Migration and deprecations¤

Renames, deprecated aliases and behavior changes. Deprecated names still work and emit a DeprecationWarning; they are removed only in a future major release. To find every deprecated name your code still uses, run your tests with -W error::DeprecationWarning. The dated record is the Changelog.

October 2026, round 11: typed API, docs that run¤

No result changes except the fixes listed under Fixes that change output.

New and wider imports¤

from ts_shape import X is fully typed for mypy, pyright and IDEs (it was object). Deprecated classes, methods and attributes are reported by type checkers (mypy: --enable-error-code deprecated); runtime warnings are unchanged.
New top-level exports: every transform and feature class, ValueMapper, PipelineResult, and TsShapeWarning, DataQualityWarning, PerformanceWarning, LoaderConfigWarning, ColumnNotFoundError, LoaderError. ts_shape.events re-exports every detector; ts_shape.context has __all__; ts_shape.loader also exports the Azure flexible-file, Databricks and energy API loaders and ContextEnricher.
list_detectors() lists every top-level public class, now including the transform and feature classes, and no longer imports the modules.

Import time¤

from ts_shape import OutlierDetectionEvents takes about 0.35 s (was 1.2–1.9 s): scipy and scikit-learn load only in the methods that use them. outlier_detection.IsolationForest is no longer a module attribute (SKLEARN_AVAILABLE is kept). The z-score is computed in numpy, identical to current scipy; on scipy older than 1.18 the last bits can differ, nearly constant data gives NaN scores, and scipy's "Precision loss" warning is no longer emitted.

PostgreSQL loaders¤

TimescaleDBDataAccess and DatapointDB build their URL with the psycopg2 driver named explicitly (works on SQLAlchemy 2.1) and escape the password. New settings: db_port (default 5432) and db_driver (default "psycopg2"), as keys of TimescaleDBDataAccess's db_config and as keyword arguments of DatapointDB, which also takes db_name (default "config_repository"). Both have a .url; without psycopg2 an ImportError names ts-shape[postgres]. Loader repr() shows the settings with passwords, keys, tokens and connection strings hidden.

Fixes that change output¤

PatternRecognition.discover_motifs(top_k > 1) no longer returns trivial matches: a later motif never starts within exclusion_zone (default window_size // 2) of an earlier one, so ranks 2+ are different shapes. top_k=1 is unchanged.
StartupDetectionEvents.detect_startup_by_slope works with repeated index labels (it raised IndexError).
detect_startup_multi_signal returns its full column set when nothing is found, for logic="any" and "all" (it returned a frame with no columns).
DatapointDB gives an empty frame for a device without data points (it raised KeyError: 'uuid').

Names follow docs/naming.md¤

Old names keep working until the next major release: a renamed keyword or attribute warns with a DeprecationWarning naming the replacement; a renamed output column stays as a twin with the same values. A canonical column sits right after its legacy twin; window bounds (start, end, duration_seconds) lead every summary.

Renamed parameters and attributes¤

Old	New	Where
`state_uuid`	`run_state_uuid`	`DowntimeEvents` (`downtime_by_shift`, `downtime_by_reason`, `top_downtime_reasons`, `availability_trend`), `SetupTimeEvents` (`setup_durations`, `setup_by_product`, `setup_statistics`, `setup_trend`), `ShiftHandoverEvents` (`generate_report`, `highlight_issues`), `IdleEnergyDetectionEvents` (`idle_energy_by_window`, `idle_energy_by_shift`, `idle_energy_trend`), `EnergyEfficiencyEvents.idle_energy_waste`, `LongDowntimeEvents`, `ReliabilityEvents`
`.state_uuid` (attribute)	`.run_state_uuid`	`LongDowntimeEvents`, `ReliabilityEvents`
`run_value_column`	`value_column_state`	`OEEEvents.performance`
`cut_value_column`	`value_column_cut`	`ContinuousProcessAlignmentEvents.segment_by_cut`
`id_value_column` (argument and attribute)	`value_column_id`	`MultiProcessTraceabilityEvents`, `RoutingTraceabilityEvents`
`handover_value_column` (argument and attribute)	`value_column_handover`	`MultiProcessTraceabilityEvents`
`routing_value_column` (argument and attribute)	`value_column_routing`	`RoutingTraceabilityEvents`
`df`	`dataframe`	`to_event_log`, `object_intervals`, `detect_objects`, `attach_objects`, `run_backtest`, `evaluate` and `LambdaDetector.evaluate` / `.to_event_log` (lambda rules), `EnergyConsumptionEvents.normalize`, `EnergyEfficiencyEvents.normalize`, `StatisticalProcessControlEvents.rule_1` .. `rule_8`
`data`	`dataframe`	`CycleDataProcessor.group_by_cycle_uuid`
`column`	`group_column`	`CycleDataProcessor.split_dataframes_by_group`
`id_column`	`uuid_column`	`EnergyConsumptionEvents.normalize`, `EnergyEfficiencyEvents.normalize`
`expected_freq`	`expected_interval`	`SignalQualityEvents.detect_missing_data`, `data_completeness`
`expected_frequency`	`expected_frequency_hz`	`SamplingResolutionEvents.detect_aliasing` (a frequency in Hz, not an interval)
`uuid_a` / `uuid_b`	`signal_uuid_a` / `signal_uuid_b`	`SignalCorrelationEvents.rolling_correlation`, `correlation_breakdown`, `lag_correlation`
`period1` / `period2`	`period_a` / `period_b`	`PeriodSummaryEvents.compare_periods`
`parquet_only`	`only_parquet`	`AzureBlobParquetLoader.list_structure`, `DatabricksUnityParquetLoader.list_structure`

Renamed output columns (old column kept as a twin)¤

Old	New	Where
`window_start` only	`start`, `end`, `duration_seconds` lead; `window_start` kept	`SamplingResolutionEvents` (all four methods), `EnvironmentalCompensationEvents.detect_correlation` / `sensitivity` / `detect_env_exceedance` (also keeps `window_end`), `MeasurementUncertaintyEvents.type_a_uncertainty` / `combined_uncertainty`, `CalibrationValidationEvents.detect_recalibration_need` / `calibration_stability`, `TimeToThresholdEvents.time_to_threshold_windows`. The event-log taxonomy now treats them as summaries (they were static), so their events carry the window's time instead of the default timestamp.
`time`	`start`, `end` (= `time`), `duration_seconds` (of `duration`)	`RateOfChangeEvents.detect_step_changes`
`mean_interval_ms`, `std_interval_ms`, `max_gap_ms`	`mean_interval_seconds`, `std_interval_seconds`, `max_gap_seconds`	`SamplingResolutionEvents.sampling_jitter`
`n_samples`	`sample_count`	`CapabilityTrendingEvents.capability_over_time`, `MeasurementUncertaintyEvents.type_a_uncertainty`, `ValueDistributionEvents.detect_bimodal` / `normality_windows` / `percentile_tracking`
`pct_of_total`	`share_of_total_pct`	`DowntimeEvents.downtime_by_reason` / `top_downtime_reasons`, `QualityEvents.nok_by_reason`, `ReworkEvents.rework_by_reason`, `ScrapEvents.scrap_by_reason`, `MaterialBalanceEvents.contribution_breakdown`
`pct_of_best`	`share_of_best_pct`	`OperatorPerformanceEvents.operator_comparison`, `ProcessStabilityIndexEvents.stability_comparison`
`pct_of_window`	`share_of_window_pct`	`MicroStopEvents.micro_stop_frequency`
`pct_of_available_time`	`share_of_available_time_pct`	`SetupTimeEvents.setup_statistics`
`pct_above`	`time_above_pct`	`ThresholdMonitoringEvents.time_above_threshold`
`pct_time_at_high` / `pct_time_at_low` / `pct_time_saturated`	`time_at_high_pct` / `time_at_low_pct` / `time_saturated_pct`	`ControlLoopHealthEvents.output_saturation`
`pct_saturated`	`saturated_pct`	`ControlLoopHealthEvents.loop_health_summary`
`pct_time_disturbed`	`time_disturbed_pct`	`DisturbanceRecoveryEvents.disturbance_frequency`
`pct_GRR` / `pct_PV` / `pct_GRR_tolerance`	`grr_pct` / `pv_pct` / `grr_tolerance_pct`	`GaugeRepeatabilityEvents.gauge_rr_summary`
`pct_contribution`	`contribution_pct`	`MeasurementUncertaintyEvents.uncertainty_budget`
`availability`, `performance`, `oee_score`	`availability_ratio`, `performance_ratio`, `oee_ratio`	`LineThroughputEvents.throughput_oee`
`period1_value`, `period2_value`	`period_a_value`, `period_b_value`	`PeriodSummaryEvents.compare_periods`

October 2026, round 10: fast at 10 and 100 million rows¤

Results are unchanged except where listed below; event ids from to_event_log are byte-identical.

ChangePointEvents.detect_change_points and segment_statistics default to method="auto": exact PELT up to 20,000 samples (results unchanged), binary segmentation above that (approximate; results on long series can differ and the slowness warning no longer fires). Pass method="pelt" for the exact search on long series.
Rolling window moments (degradation and health scores) and matrix-profile distances (motif and discord discovery) differ from before by at most 1e-9 relative; motif and discord positions are unchanged.
Setpoint KPIs no longer fail on an actual signal with duplicate index labels.
A missing uuid no longer matches the text "nan" or "None" when selecting signals.

October 2026, round 9: KPIs match their textbook definitions¤

Several KPIs gave wrong numbers. They now follow the published definitions (ISO 22400-2, Montgomery Introduction to Statistical Quality Control, Ogata Modern Control Engineering, ISO 50006, Nakajima), and each one is pinned by a hand-computed reference test in tests/reference/. Results change; no code change is needed unless you depended on the old values. Where an option keeps the old behavior, it is named below.

Time and parts are booked to the period they belong to¤

Durations used to be booked to the shift/hour/day of the reading that started them, and counter increments to the reading that reported them. Both are now split at period boundaries (wall-clock, DST-safe), via the new ts_shape.utils.time_accounting helpers.

Affected: DowntimeEvents.downtime_by_shift / availability_trend, ShiftHandoverEvents, RuntimeAccountingEvents.runtime_by_window / operating_hours_meter, MicroStopEvents.micro_stop_impact, OEEEvents (availability, performance, daily OEE) and QualityTrackingEvents.nok_by_shift.

Probe	Old	New
Stop 07:00–15:00, shift_1 (06–14) availability	0 % (480 min down)	12.5 % (60 up / 420 down)
Same stop, `availability_trend(window="1h")`, 07:00 hour	480 min down	60 min down (one row per hour)
Night shift 22–06 running at 100 %, OEE day 1 / day 2	12.5 % / 0 %	100 % / 100 %
30 parts counted 13:50–14:10, shift_1 / shift_2	0 / 30	15 / 15

A counter increment is spread over (previous reading, reading] in proportion to time; a reset counts in the period of the reading that shows it. handle_resets=False keeps the per-period max(0, last - first).
Part counts in OEE (actual_parts, total_parts, reject_parts, good_parts) and shift-handover quantities are floats now.
Shift and trend outputs list every period that holds observed time, not only periods that contain a reading.
A day without state rows has NaN availability, performance and OEE (was: the counter span as run time, availability 0).
Performance or availability above 100 % is no longer clipped; the value is kept and a DataQualityWarning names the period.
micro_stop_frequency still counts a stop in the window it starts in.

Production¤

PerformanceLossEvents counts parts (counter increments), not counter rows; cycle time = run time / parts. Counter +2 per minute for 60 min at a 30 s ideal cycle: performance 50 % → 100 %, loss 1770 s → 0 s. A new run_seconds column carries the basis.
PeriodSummaryEvents.from_daily_data recomputes KPIs from summed bases (quality = Σgood/Σtotal, availability = Σrun/Σplanned, OEE = A×P×Q, MTTR = Σrepair/Σfailures, EnPI = Σenergy/Σunits, ...). Ratios without a basis column are weighted means (plain mean with a DataQualityWarning); numeric columns with no rule are dropped with a warning instead of summed. A new aggregations= argument overrides the rule per column. Quality of 900/1000 and 0/10: 45.0 % → 89.1 %; MTTR 1 h and 3 h: 4.0 → 2.0.
MicroStopEvents.micro_stop_impact gains performance_loss_pct (minor stops are a performance loss); availability_loss_pct stays with the same value.

Setpoint response¤

time_to_settle is the time after which the response stays in the band until the end of the window (last entry, not first). The window now also ends at the next setpoint change. New settled_within_window column; t_settle_seconds is NaN when the response does not settle. settled means settled within the window with a final in-band stretch of at least hold. Underdamped step ζ=0.2, ωn=0.1, 2 % band: 18 s → 197 s.
overshoot_pct / undershoot_pct are percent (0–100) like every other *_pct column: 0.5266 → 52.66. The fraction is in the new overshoot_ratio / undershoot_ratio columns.

Statistical process control and capability¤

StatisticalProcessControlEvents(..., sigma_method="moving_range") (default) estimates σ as MR̄/1.128 (individuals chart); sigma_method="std" brings back the sample standard deviation.

Probe	Old	New
Reference 1..20: control limits	−7.25 / 28.25	7.84 / 13.16

The rule 1–8 zones and detect_cusum_shifts use the same σ̂.
dynamic_control_limits(method="moving_range") tests each point against the window_size points before it (center = their mean, σ̂ = their MR̄/1.128); the first window_size points have NaN limits.
dynamic_control_limits(method="ewma") is an EWMA chart: z_i = λx_i + (1−λ)z_{i−1}, limits μ0 ± L·σ̂·√(λ/(2−λ)·[1−(1−λ)^{2i}]). ewma is z_i, mean is μ0; new ewma_lambda= (default 2/(window_size+1)) and n_sigma= (L) set the new lcl / ucl columns. Both methods add a value column.
Cp/Cpk use the within σ (MR̄/1.128) and Pp/Ppk the overall σ (AIAG), in CapabilityTrendingEvents (sigma_method= on the constructor) and ToleranceDeviationEvents.capability_indices (sigma_method= on the method). sigma_method="std" makes Cp/Cpk equal Pp/Ppk again. New std_within column. Per-window Pp/Ppk use the window's own σ (they used the whole signal's). Example x = 10.0, 10.4, 10.2, 10.6, 9.8, specs 9–11: Cp 1.054 → 0.836, Pp unchanged 1.054.

Energy and supply chain¤

The EnPI rolling baseline, EnergyEfficiencyEvents.sec_trend and efficiency_trend rolling_avg_efficiency are Σnumerator/Σdenominator over the window (ISO 50006), not the mean of ratios: EnPI 1 and 3 over 130 kWh / 110 units: 2.0 → 1.182.
InventoryMonitoringEvents.stockout_prediction uses the refill-aware consumption rate (handle_refills=True, new); a refill no longer cancels consumption. handle_refills=False keeps first-minus-last.

Documented, unchanged¤

SignalCorrelationEvents.lag_correlation: positive lag_periods means uuid_b leads uuid_a (b(t) = a(t−5) gives −5). The docstring now says so.

October 2026, round 8: legacy code removed¤

Requirements¤

ts-shape now needs pandas 2.2.2 or later and numpy 2.0 or later (was pandas 2.1, numpy 1.26). A CI job tests the oldest supported versions, and results are the same on pandas 2.2 and pandas 3: daily windows are calendar days (23 h / 25 h on a DST change) on both.

setup.py, MANIFEST.in and the requirements*.txt files are gone; pyproject.toml is the only place dependencies are declared.

Deprecation warnings you can see¤

Deprecation warnings point at your own line now. Before, warnings from detector methods pointed inside ts-shape, and Python's default filters hid them, so a deprecated name could be used without a hint. Run your tests with -W error::DeprecationWarning once to find them.

Newly deprecated (old name still works and warns):

Loader time ranges are start / end everywhere; positional calls are unchanged.

Old	New
recipe spec key `duration_s`	`duration_seconds`
`ShiftHandoverEvents.from_shift_data(downtime_df=...)` with only `downtime_minutes`	a `downtime_seconds` column
`CycleDataProcessor.cycle_uuid_col` / `.systime_col` (attributes)	`.cycle_uuid_column` / `.time_column`
`RuntimeAccountingEvents.run_uuid` (attribute)	`.run_state_uuid`
`MultiSensorValidationEvents.sensor_uuids` (attribute)	`.signal_uuids`
`start_timestamp=` / `end_timestamp=` (Azure Blob, Databricks Unity and Spark parquet loaders, `TimescaleDBDataAccess`, `S3ProxyDataAccess`)	`start=` / `end=`
`start_time=` / `end_time=` (`ParquetLoader.load_by_time_range`, `load_files_by_time_range_and_uuids`)	`start=` / `end=`
`TimescaleDBDataAccess.start_timestamp` / `.end_timestamp`, same on `S3ProxyDataAccess` (attributes)	`.start` / `.end`

Behavior changes¤

Time-named numeric columns stay numbers. A class without its time column used to convert every column whose name contained "time" or "date" (cycle_time_mean, update_rate) to datetimes. Only datetime or text columns are used now, with a DataQualityWarning.
Mixed time units. SegmentProcessor.apply_ranges and ContinuousProcessAlignmentEvents.segment_by_cut compare times in one unit; ranges in seconds against data in nanoseconds matched no rows.
Hour folders are UTC in every loader (Azure, Databricks Unity, Parquet): a tz-aware bound is converted to UTC, a naive bound is UTC. Before, some loaders used local wall-clock hours.
CycleDataProcessor.split_dataframes_by_group raises ColumnNotFoundError for a frame without the group column (it was logged and skipped). CriticalParameterRankingEvents.rank raises ColumnNotFoundError for a missing outcome column.
Problems that were only logged are DataQualityWarnings: degenerate SPC limits, a singular MSPC covariance, ignored or overlapping segment ranges, invalid or overlapping cycles, empty feature tables. Text in numeric columns is reported in pattern recognition, feature export, segment statistics, cycles and continuous-process alignment.
StatisticalProcessControlEvents.detect_cusum_shifts counts a repeated sample (same time, same value) once, with a DataQualityWarning. A duplicated row was added to the running sum twice and could report a shift on its own.
ProfileComparison.cluster(n_clusters=...) always returns the requested number of clusters (SciPy's maxclust cut could return fewer).
New exports: ts_shape.events.production lists FlowMetricsEvents, LineBalancingEvents, RuntimeAccountingEvents; ts_shape.features lists PatternRecognition, DescriptiveFeatures, TimeWindowedFeatureTable.
The API reference no longer has pages for private modules, and package pages have clean URLs (.../events/production/ instead of .../events/production/__init__/).

October 2026, round 7: every function robust and consistent¤

Every public function, not only the detectors, now handles empty frames, timezones, nullable and pyarrow dtypes, text time columns and missing columns the same way. Contract tests check this for every public callable.

Renamed parameters¤

The old names work and warn.

Old	New	Where
`column_name` (a value column)	`value_column`	filters, `IntegerCalc`, `LambdaProcessor.apply_function`, `UnitConverter.convert_column`, `Numeric/Boolean/StringStatistics`
`column_name` (the time column)	`time_column`	`DateTimeFilter`, `TimestampStatistics`, `ValueMapper`
`freq`	`window`	`PowerIntegrator`, `DataHarmonizer.resample_to_uniform` / `merge_multi_signals`, `TimeGroupedStatistics`
`df`, `uuid_col`, `value_cols`, `group_col`	`dataframe`, `uuid_column`, `value_columns`, `group_column`	`FeatureMatrixExporter.to_feature_matrix`
`cycle_uuid_col`, `systime_col`	`cycle_uuid_column`, `time_column`	`CycleDataProcessor`
`metric`	`value_column`	`CycleDataProcessor.compare_cycles`, `identify_golden_cycles`
`metadata_uuid_col`, `tolerance_uuid_col`, `mapping_uuid_col`	`*_uuid_column`	`ContextEnricher`

DateTimeFilter takes the bound right after the frame: filter_after_date(df, "2024-01-02", time_column="systime"). The old order still works and warns.

Results that were wrong¤

CrossSignalAnalytics.lead_lag reported the leader backwards. A positive optimal_lag now means signal_a leads. All its methods sort by time and use only rows where both signals have a value.
PatternRecognition sorts by time, keeps the timezone, and needs uuid= when the frame holds several signals. normalize=False now really uses the raw distance.
BooleanFilter edges are found per uuid in time order; a missing value holds the previous state. StringFilter.detect_changes_in_string always reports the first row of each uuid, also for string dtype.
DataHarmonizer.fill_gaps(max_gap=) measures the gap between the valid samples around it; signals without values are kept as NaN columns.
LambdaProcessor.apply_function and TimezoneShift.revert_to_original_timezone no longer change your frame.
FeatureMatrixExporter uses the sample standard deviation (ddof=1) and aggregates only value_* columns by default.
ReliabilityEvents, InventoryMonitoringEvents.stockout_prediction: NaN instead of inf. LeadTimeAnalysisEvents.lead_times pairs each order with the first later delivery, never a negative lead time.
A zero-length step (duplicate timestamp) is no data gap and starts no flatline. Duplicated rows no longer duplicate outlier and tolerance events.
object_intervals(min_duration=) no longer loses an id that comes back after a short blip.

Event ids¤

to_event_log builds ocel:eid from the detector, source_uuid, the timestamps, the activity and the row's values, no longer from the row position. The same detector on two machines no longer collides, and concat no longer drops events. Event ids from earlier releases change.

Also in to_event_log: a string in objects= that is not a column is used as the object id for every row; a NaN in an activity template gives unknown with a warning; interval rows with end < start are dropped with a warning; the static shape uses timestamp= (default 1970-01-01 UTC), not the current time.

No data, unknown codes¤

Percentages of no data (true_percentage, uppercase_percentage, ...) are NaN, not 0. A mode of no data is None.
CrossSignalAnalytics with too little data gives NaN, not 0.0 / 1.0.
ValueMapper keeps codes the mapping table does not know, with a DataQualityWarning (they were NaN). Duplicate keys raise ValueError.
ContextEnricher calls build on each other; get_enriched_dataframe() returns the cumulative result. Duplicate keys and colliding column names raise ValueError.
ToleranceDeviationEvents.capability_indices and MultivariateSPCEvents with too little data give NaN / an empty result with a warning instead of raising.

Errors¤

A missing column raises ColumnNotFoundError (a ValueError) in every public function; it was often a KeyError.

Loaders, signals, rules¤

DatabricksSparkParquetLoader.build_paths includes the hour of a mid-hour start and builds folder hours in UTC. ParquetLoader matches uuids to whole file names. EnergyAPILoader returns UTC times.
signals.align handles mixed time units and names the time column after time_column; per_window handles booleans with missing values. SignalSpec validates its fields.
Lambda rule expressions reject and, or, not, in and chained comparisons at compile time (use &, |, ~, .isin). Interval rules end at the next sample after the last true one, like signals.intervals.
Pipeline passes arrays and frames as step arguments and never changes your input frame.

New¤

max_gap on detect_flatline, threshold_with_hysteresis and on_off_intervals.
DataHarmonizer.detect_gaps adds start, end, duration_seconds.
TimeGroupedStatistics.calculate_custom_func(group_by_uuid=True).

October 2026, round 6: consistent outputs¤

Summary methods return a DataFrame¤

Like best_and_worst_shifts (round 5), these methods take as_frame: bool = True. The default returns a DataFrame whose columns are the old keys, plus start / end / duration_seconds for the data the summary covers. as_frame=False returns the old dict or list and emits a DeprecationWarning; that form will be removed in the next major release. To keep old code working unchanged, pass as_frame=False. Alternatively, read a one-row result with .iloc[0], which supports the same ["key"] access.

Method	Default DataFrame
`SteadyStateDetectionEvents.steady_state_statistics`	one row; the dict keys plus `uuid`, `source_uuid`
`MachineStateEvents.state_quality_metrics`	one row; the dict keys plus `uuid`, `source_uuid`
`TargetEvents.target_hit_rate`	one row; `start` / `end` span the analysed days
`ToleranceDeviationEvents.capability_indices`	one row; `Cpm` is always a column (NaN without `target_value`), plus `sample_count`
`BottleneckDetectionEvents.throughput_constraint_summary`	one row per station: `bottleneck_count`, `bottleneck_pct`, `avg_utilization_pct`, `most_frequent_bottleneck`
`FlowConstraintEvents.flow_constraint_analytics`	one row per `constraint_type` (`blocked`, `starved`): `event_count`, `total_duration_seconds`, `avg_duration_seconds`, `minor_count`, `moderate_count`, `severe_count`, `time_alignment_quality`. The event lists come from `blocked_events` / `starved_events`.
`ShiftHandoverEvents.highlight_issues`	one row per issue, with the report row's shift window and `date`

KPI percentages carry their basis¤

New columns; none were removed:

Method	New columns
`OEEEvents.oee`	`run_seconds`, `planned_seconds`, `excluded_seconds`, `actual_parts`, `ideal_parts`, `total_parts`, `reject_parts`, `good_parts` (NaN without quality counters), `resets`
`OEEEvents.availability`	`excluded_seconds` (time `max_gap` removed as no data)
`LineThroughputEvents.throughput_oee`	`run_seconds`, `planned_seconds`, `resets` (with `handle_resets=True`)
`DowntimeEvents.downtime_by_shift` / `availability_trend`	`excluded_seconds`
`PerformanceLossEvents` (all three methods)	`run_seconds`

Durations in seconds¤

Each *_minutes or *_s duration column now has a *_seconds twin right after it: total_seconds, downtime_seconds, uptime_seconds, avg_seconds (downtime), avg_cycle_time_seconds, target_cycle_time_seconds, loss_seconds (performance loss), avg_seconds, min_seconds, max_seconds, median_seconds, std_seconds, avg_setup_seconds, total_setup_seconds (setup time), downtime_seconds (shift handover, also in OUTPUT_COLUMNS). The old columns stay as a convenience. Their removal would be a breaking change and is left for a future major release; new code should read the *_seconds columns. In recipe-phase specs duration_seconds is the new spelling of the duration_s criterion; duration_s still works and, since round 8, warns.

Run state as string or boolean¤

DowntimeEvents and ShiftHandoverEvents read either a string state signal (running_value="Running" in value_string) or a boolean run signal (value_bool, True = running). The boolean signal is used when the state rows carry no value_string, or when running_value / stopped_value is a bool. Both forms give the same downtime. Before this change, a boolean run signal was compared with "Running" and therefore reported 100 % downtime.

Fixes¤

FlowConstraintEvents.flow_constraint_analytics applies its warning_threshold / critical_threshold to the severity counts and to the event lists. Before, it accepted them but classified with the 5 s / 30 s defaults.

Same numbers for any data source, totals that add up¤

Dtypes no longer change results. Value columns with pandas nullable, pyarrow or text-number dtypes are converted once when a detector is built. Text that is not a number is reported with a DataQualityWarning at construction, not when a method runs. Counter totals on nullable or pyarrow data were too high (the first reading counted as production): they now match NumPy data.
Group totals add up to the counter total. Per shift, day, operator, part, reason, batch and schedule period, the part made between the last reading of one group and the first reading of the next now counts in the later group (utils.counters.counter_by_group). Quantities can grow by one increment per group boundary; resets can grow by one when a reset falls on a group boundary. Per-operator/part/reason totals no longer double count when those groups alternate. shift_production(date=...) includes the increment into the day's first reading. batch_yield books production between two batches to the earlier batch. handle_resets=False is unchanged.
FlowConstraintEvents intervals end at the next change (interval_end="next_change"); pass interval_end="last_sample" for the previous, shorter durations.
MultiSensorValidationEvents.detect_outlier_sensor needs three sensors to name an outlier; with two it skips the window and warns.
ToleranceDeviationEvents.capability_indices on empty data returns an empty frame with the full schema; CapabilityTrendingEvents uses the spec limits in force in each window.
Energy methods with window="W" / "ME" report the real week or month (end was start + 1 day).
Shift labels: a row with a missing timestamp gets the shift "unknown".
Degenerate inputs: duplicate timestamps give NaN rates (with a warning) instead of inf; detect_oscillation(window<=0) raises.

Cycles, segments and objects¤

These were verified against hand-built scenarios; the fixes change results:

Cycles (CycleExtractor):
back-to-back cycles sharing a timestamp come out correctly;
a second start while a cycle is open aborts it (is_complete=False) instead of producing overlapping cycles;
a held start pulse is one start;
value/state-change cycles read only start_uuid;
a cycle already running at the first sample has is_complete=False;
cycle_uuid is deterministic (same data, same id).
Cycle data (CycleDataProcessor): every method uses start <= t < end. A sample exactly at a cycle end belongs to the next cycle, not to both; open cycles get no samples. identify_golden_cycles(method="target_value") needs target=.
Segments (SegmentExtractor / SegmentProcessor):
the last segment runs to the end of the data (it had zero length);
apply_ranges cuts half-open, so boundary samples are no longer duplicated;
the first sample of an empty-id gap is no longer given to the previous order.
segment_by_cut: material is assigned to the piece the next cut finishes. Before, it was one piece off and piece 1 got two pieces' worth. Material after the last cut has piece_id <NA>.
Objects (object_intervals / detect_objects):
an empty or NaN reading ends a presence;
a change-encoded id stays present to the end of the data;
at a hand-over instant only the incoming object gets the event;
a child is part_of only the parent it overlaps longest;
min_duration measures the presence, not first-to-last sample;
use to_event_log_ocel(log, pm4py=True) for pm4py export.
production_dates: correct after a DST change (it was one day off).

Severity levels and small fixes¤

severity_level. Every result with a categorical severity also has severity_level (info / warn / critical), right after severity. The severity labels are unchanged. Code that selects columns by position sees one more column.
TimeWindowedFeatureTable: freq is renamed window. freq still works and warns.
CriticalParameterRankingEvents(): the long-format frame is optional and rank() never used it, so a time column is no longer needed.
ContinuousProcessAlignmentEvents.lag_profile: a window without speed readings has NaN speed and lag. Before, it reported the lag at min_speed: thousands of seconds.

October 2026, round 5: correct by default¤

The correct behavior is now the default, and results are returned at full precision. No name was removed, so your code keeps running, but results change wherever your data has counter resets, change-encoded signals, uneven sampling or several signals in one frame. Each old behavior is one argument away.

Defaults that changed¤

Parameter	Old default	New default	What it fixes	Where
`handle_resets`	`False`	`True`	A counter that resets (shift start, rollover) no longer makes a window report too little, often 0. Results carry a `resets` column.	production (OEE, parts, quality, scrap, rework, operator, target, period summary, shift report and handover, batch yield, line throughput, process alignment), energy (EnPI, efficiency, consumption, carbon), `utils.counters`, `TimeGroupedStatistics` (`'diff'`)
`interval_end`	`"last_sample"`	`"next_change"`	An interval lasts until the signal changes. A change-encoded (deadband) signal's interval is no longer cut to zero length.	machine state, duty cycle, micro stops, long downtime, steady state, threshold monitoring, degradation, vibration, recipe phases, inventory, signal quality, flatline, correlation breakdown, `SegmentExtractor`, `utils.intervals`
`time_weighted`	`False`	`True`	Shares and averages are weighted by time, not by sample count, so irregular sampling no longer biases them.	bottleneck detection, duty cycle by window, operating range, material balance, idle energy, efficiency idle waste
`group_by_uuid`	`False`	`True`	Statistics are computed per signal instead of pooling unrelated signals.	`OutlierDetectionEvents`, `TimeGroupedStatistics`, `BooleanFilter` edges, `StringFilter.detect_changes_in_string`
`slope_per`	`"sample"`	`"second"`	Drift slopes are per second, so irregular sampling no longer distorts them; `min_slope` is per second too.	`AnomalyClassificationEvents.detect_drift`, `classify_anomalies`
`missing_as_zero`	`True`	`False`	A missing flow signal makes the balance NaN instead of a false imbalance.	`MaterialBalanceEvents`
`handle_refills`	`False`	`True`	A refill no longer turns into negative consumption.	`InventoryMonitoringEvents.consumption_rate`
`include_severity`	`False`	`True`	Violations carry `triggered_rule` and `severity`.	`StatisticalProcessControlEvents.detect_violations`
`regex`	`True`	`False`	Substring filters match literally (`"A.1"` no longer matches `"AB1"`).	`StringFilter.filter_string_contains`, `StringStatistics.contains_substring_count`
`trim_to_range`	`False`	`True`	Loaders return only rows inside the requested time range.	Azure Blob and Databricks Unity parquet loaders
`as_frame`	`False`	`True`	`best_and_worst_shifts` returns one DataFrame with a `rank_type` column; `as_frame=False` (a dict) is deprecated.	`ShiftReportEvents`
`layout` (new)	—	`"long"`	`batch_transition_matrix` returns `from_state`, `to_state`, `transition_count`, and an empty result keeps that schema. `layout="wide"` gives the old matrix.	`BatchTrackingEvents`

To compare old and new numbers, pass the old default explicitly, for example oee(..., handle_resets=False).

Other result changes¤

Full precision. Detectors no longer round results (they used 0 to 8 decimals). Round when you display: df.round(2).
machine_running_pct is a percentage (0–100). In IdleEnergyDetectionEvents.idle_energy_by_window and EnergyEfficiencyEvents.idle_energy_waste it held a 0–1 ratio. The ratio is machine_running_ratio.
Flatlines: with "next_change", a single sample counts as a flatline only when the gap after it is longer than the signal's usual sampling interval.
ValueDistributionEvents.detect_bimodal returns a unimodal row for a constant signal (it was empty or had a None mode).
CriticalParameterRankingEvents.rank raises ValueError for an unknown candidate column (it was skipped).
TimezoneShift.list_available_timezones returns a sorted list (it was a set).
make_timeseries no longer rounds value_double.

Nothing is skipped silently¤

Loaders: an unreadable blob or file used to disappear from the result. Multi-object Azure Blob and Databricks Unity methods take errors="warn" (default: skip it and name it in one DataQualityWarning) or errors="raise" (LoaderError). Series metadata that exists but cannot be read raises LoaderError.
Non-numeric values in numeric inputs are still treated as missing, but a DataQualityWarning now says how many there were.
A missing value column (value_column, value_column_<role>, ...) raises ColumnNotFoundError naming it and listing the columns, instead of a bare KeyError.
Malformed recipe-phase specs, untestable normality windows, failed fits and event-log object bindings to a missing column now warn instead of being dropped.
lambda_rules.load_yaml without PyYAML raises ImportError pointing to ts-shape[yaml].

Newly deprecated¤

Old	New
`OrderTraceabilityEvents`	`ValueTraceabilityEvents`
`ToleranceDeviationEvents(tolerance_uuid=...)` alone	`upper_tolerance_uuid` and/or `lower_tolerance_uuid` (either one alone is a one-sided spec)
`RoutingTraceabilityEvents(station_map=...)`	`state_map`
`eventlog.adapters.adapt(legacy=...)`	`frame=`
integer `window` in `SignalCorrelationEvents`	`window_size`
`best_and_worst_shifts(as_frame=False)`	the default DataFrame

New¤

OutlierDetectionEvents(signal_uuids=[...]) restricts the analysis to some signals.
pip install ts-shape[parquet] installs pyarrow (also pulled in by [s3] and [azure]); [all] now installs every optional feature.

Renamed parameters (old keywords deprecated)¤

Parameter names follow Naming & Output Conventions. The old keyword still works, but emits a DeprecationWarning.

Old	New	Where
`column_name` (the time column)	`time_column`	every detector constructor (`Base`)
`freq`, `period`, `resample`	`window`	data-gap coverage, value distribution, demand patterns, period summary, correlation and root-cause methods
integer `window` (a row count)	`window_size`	`SignalCorrelationEvents.rolling_correlation` / `correlation_breakdown`, `StatisticalProcessControlEvents.dynamic_control_limits`
`z_threshold`, `threshold_sigma`	`n_sigma`	anomaly classification, anomaly correlation, disturbance recovery
`low`/`high`, `lower`/`upper`, `low_limit`/`high_limit`, `lower_bound`/`upper_bound`	`lower_limit` / `upper_limit`	threshold hysteresis, operating range, control-loop saturation, time-to-threshold
`minor_threshold` / `moderate_threshold`	`warning_threshold` / `critical_threshold`	`FlowConstraintEvents.flow_constraint_analytics`
`include_first`, `count_initial_active`	`include_initial`	reliability (Weibull), alarm frequency
`significant_only`, `on_crossing_only`	`only_significant`, `only_crossings`	root-cause ranking, asset-health events
`sensor_uuids`	`signal_uuids`	`MultiSensorValidationEvents`
`run_uuid`	`run_state_uuid`	`RuntimeAccountingEvents`
`counter_column`, `energy_column`, `state_column`	`value_column_counter`, `value_column_energy`, `value_column_state`	energy pack

Optional arguments are keyword-only. Passing them positionally still works, but warns, for example detect_outliers_zscore(3.0) instead of detect_outliers_zscore(threshold=3.0).

The default event_uuid values now all follow "<pack>:<name>": - outlier_event → quality:outlier - startup_event → eng:startup - setpoint_change_event → eng:setpoint_change - FlowMetricsEvents: prod:flow → prod:flow_metrics (it used to collide with FlowConstraintEvents)

If you filter results by these values, update the filters.

Renamed classes and methods (old names deprecated)¤

Detector classes all end in Events. Event classes no longer use calculate_ / compute_ / process_: methods that find events start with detect_, and metrics are nouns (see Naming & Output Conventions).

The old names still work. They emit a DeprecationWarning, and the event log accepts them too: for example, to_event_log(df, "OEECalculator.calculate_oee") is resolved to OEEEvents.oee.

Old	New
`CycleTimeTracking`	`CycleTimeEvents`
`DowntimeTracking`	`DowntimeEvents`
`OperatorPerformanceTracking`	`OperatorPerformanceEvents`
`PartProductionTracking`	`PartProductionEvents`
`PerformanceLossTracking`	`PerformanceLossEvents`
`QualityTracking`	`QualityEvents`
`ReworkTracking`	`ReworkEvents`
`ScrapTracking`	`ScrapEvents`
`SetupTimeTracking`	`SetupTimeEvents`
`TargetTracking`	`TargetEvents`
`OEECalculator` (`calculate_availability`, `calculate_performance`, `calculate_quality`, `calculate_oee`)	`OEEEvents` (`availability`, `performance`, `quality`, `oee`)
`PeriodSummary`	`PeriodSummaryEvents`
`ShiftReporting`	`ShiftReportEvents`
`ShiftHandoverReport`	`ShiftHandoverEvents`
`ProcessStabilityIndex`	`ProcessStabilityIndexEvents`
`StatisticalProcessControlRuleBased` (`process`, `apply_rules_vectorized`, `calculate_control_limits`, `calculate_dynamic_control_limits`)	`StatisticalProcessControlEvents` (`detect_violations`, `detect_rule_violations`, `control_limits`, `dynamic_control_limits`)
`ToleranceDeviationEvents.process_and_group_data_with_events` / `.compute_capability_indices`	`.detect_deviations` / `.capability_indices`
`DataGapAnalysisEvents.find_gaps`	`.detect_gaps`
`MultiSensorValidationEvents.identify_outlier_sensor`	`.detect_outlier_sensor`
`LeadTimeAnalysisEvents.calculate_lead_times`	`.lead_times`
`EnvironmentalCompensationEvents.estimate_sensitivity`	`.sensitivity`
`DesignOfExperimentsEvents.compute_effects`	`.effects`
`.alarm_duration_stats`, `.batch_duration_stats`, `*.state_duration_stats`	`*_statistics`
`DutyCycleEvents.duty_cycle_per_window`, `RuntimeAccountingEvents.runtime_per_window`	`*_by_window`

Behavior fixes that change results¤

These change output because the old output was wrong or contradicted the documentation. Each one has a regression test.

SPC rules now follow Nelson (1984) / Western Electric (StatisticalProcessControlEvents):
rule 8 flags 8 points in a row with none within 1σ (it used to flag 8 points within 1σ);
rule 3 is 6 points steadily rising or falling (was 7), and rule 4 is 14 points alternating up and down (was 15);
rules 5 and 6 count points beyond the next zone too, per side;
rule 2 measures sides against the chart center line;
the rules no longer add columns to your frame, and the vectorized and per-rule paths flag the same points.
ToleranceDeviationEvents.detect_deviations keeps the measured value and reports single-sample deviations. A missing actual_uuid raises instead of returning nothing. capability_indices adds Cpm when target_value is given.
LineThroughputEvents.takt_adherence:
a cycle exactly equal to takt is no longer a violation;
a decrease in a numeric counter (a reset) is no longer a cycle.
MachineStateEvents.state_quality_metrics:
follows the signal's spec like detect_run_idle;
run_idle_ratio is NaN, not 0.0, when there is no idle time.
ShiftHandoverEvents.highlight_issues without report_df or the four uuids raises instead of returning [].
PeriodSummaryEvents.from_daily_data honors anchored weeks ('W-FRI' …).
GoldenBatchDeviationEvents.phase_breakdown compares each phase with the same phase of the golden batch.
EnvironmentalCompensationEvents:
constructing without env_uuid raises;
compensated_signal's uuid is the event uuid.
SetpointChangeEvents: two setpoint changes at the same instant no longer multiply the KPI rows.
ProcessWindowEvents uses the strict > its docstrings describe.
Hour-folder loaders no longer drop the hour that contains the range end. Databricks loaders apply a single start or end bound.

Parameters that had no effect now warn when passed: RateOfChangeEvents.detect_rapid_change(window=...), MeasurementUncertaintyEvents(part_uuid=...), ContinuousProcessAlignmentEvents(ref_uuid=...). When both settle_pct and tol are passed to a setpoint KPI, a warning says settle_pct wins. When both state_map and station_map are passed, a warning says state_map wins.

Every detector output carries its time bounds¤

Each detector returns the canonical columns of its shape (see Naming & Output Conventions): - intervals: start, end, duration_seconds, uuid, source_uuid; - summaries: start, end, duration_seconds, also when they are keyed by date, (date, shift) or period; - points: systime, uuid, source_uuid.

The new columns were added next to the old ones; none was removed or renamed. Two consequences:

Merging two summaries on date or (date, shift) now finds start, end and duration_seconds on both sides, and pandas suffixes them _x / _y. Merge on the window instead, or select your columns first:

import numpy as np
import pandas as pd
from ts_shape import PartProductionEvents, QualityEvents

t = pd.date_range("2024-01-01", periods=96, freq="h")
df = pd.concat([
    pd.DataFrame({"systime": t, "uuid": "line1:ok", "value_integer": np.arange(96) * 9}),
    pd.DataFrame({"systime": t, "uuid": "line1:nok", "value_integer": np.arange(96)}),
    pd.DataFrame({"systime": t, "uuid": "line1:part", "value_string": "P-100"}),
    pd.DataFrame({"systime": t, "uuid": "line1:count", "value_integer": np.arange(96) * 10}),
], ignore_index=True)

quality = QualityEvents(df).daily_quality_summary("line1:ok", "line1:nok")
production = PartProductionEvents(df).daily_production_summary("line1:part", "line1:count")
daily = quality.merge(production, on=["start", "end", "duration_seconds"])

Night shifts are one row. An overnight shift (e.g. 22:00 to 06:00) is attributed to the day it starts (production_day=True is now the default). Before, grouping by calendar date split one night across two rows, and each row mixed two different nights. Pass production_day=False to get the old grouping.

Removed exporter aliases¤

The event-log exporters were renamed into a consistent to_event_log_* family. The old aliases (to_flat_df, to_ocel_tables) have now been removed — update any remaining imports:

Removed	Use instead
`to_flat_df(...)`	`to_event_log_xes(...)`
`to_ocel_tables(...)`	`to_event_log_ocel(...)`

# Before
from ts_shape.eventlog import to_flat_df, to_ocel_tables

# After
from ts_shape.eventlog import to_event_log_xes, to_event_log_ocel

`to_event_log_ocel` now returns five tables¤

To fully model OCEL 2.0, to_event_log_ocel(log) returns an OCEL2Tables bundle (object-to-object relations and time-varying object attributes were added) instead of a 3-tuple. Access the frames by attribute:

# Before
events_df, objects_df, relations_df = to_event_log_ocel(log)

# After
tables = to_event_log_ocel(log)
tables.events, tables.objects, tables.relations          # as before
tables.o2o, tables.object_changes                        # new OCEL 2.0 tables

`org:resource` in XES export¤

to_event_log_xes previously copied the case object id into org:resource. Per the XES standard org:resource is the performing resource, so it is now populated only from an operator-type object relation, and omitted when the log has no operators. Supply one via objects={"operator": ...} if you need it.

Standardized detector output (May 2026)¤

Every public DataFrame-returning detector method now emits one of three canonical shapes — point, interval, or summary — defined in src/ts_shape/events/_output.py. If you previously read shape-specific column names, switch to the canonical ones:

Old column(s)	Canonical column
`start_time`, `gap_start`, `window_start`, `period_start`, `disturbance_start`	`start`
`end_time`, `gap_end`, `window_end`, `period_end`, `disturbance_end`	`end`
`severity_score`, `severity_level`	`severity`
`OEECalculator`: `date`	`start` + `end` + `duration_seconds` (86400)

See the Changelog for the full dated history of breaking changes.

Migration and deprecations¤

October 2026, round 11: typed API, docs that run¤

New and wider imports¤

Import time¤

PostgreSQL loaders¤

Fixes that change output¤

Names follow docs/naming.md¤

Renamed parameters and attributes¤

Renamed output columns (old column kept as a twin)¤

October 2026, round 10: fast at 10 and 100 million rows¤

October 2026, round 9: KPIs match their textbook definitions¤

Time and parts are booked to the period they belong to¤

Production¤

Setpoint response¤

Statistical process control and capability¤

Energy and supply chain¤

Documented, unchanged¤

October 2026, round 8: legacy code removed¤

Requirements¤

Deprecation warnings you can see¤

Behavior changes¤

October 2026, round 7: every function robust and consistent¤

Renamed parameters¤

Results that were wrong¤

Event ids¤

No data, unknown codes¤

Errors¤

Loaders, signals, rules¤

New¤

October 2026, round 6: consistent outputs¤

Summary methods return a DataFrame¤

KPI percentages carry their basis¤

Durations in seconds¤

Run state as string or boolean¤

Fixes¤

Same numbers for any data source, totals that add up¤

Cycles, segments and objects¤

Severity levels and small fixes¤

October 2026, round 5: correct by default¤

Defaults that changed¤

Other result changes¤

Nothing is skipped silently¤

Newly deprecated¤

New¤

Renamed parameters (old keywords deprecated)¤

Renamed classes and methods (old names deprecated)¤

Behavior fixes that change results¤

Every detector output carries its time bounds¤

Removed exporter aliases¤

to_event_log_ocel now returns five tables¤

org:resource in XES export¤

Standardized detector output (May 2026)¤

`to_event_log_ocel` now returns five tables¤

`org:resource` in XES export¤