Elevator fault prediction method and system based on multi-source data fusion

Abstract

This invention discloses a method and system for predicting elevator faults based on multi-source data fusion, belonging to the field of elevator fault prediction technology. The method includes: collecting multi-source operating data during elevator operation; performing unified time alignment processing and segmented feature extraction processing on the multi-source operating data to form a multi-source fusion feature vector; identifying operating condition-related features caused by operating conditions based on the multi-source fusion feature vector and performing weakening processing to construct a state representation without operating condition bias; extracting a set of fault-sensitive features characterizing abnormal operating behavior based on the state representation without operating condition bias; converting window-level anomaly judgment results in the fault-sensitive feature set into abnormal events, and constructing an anomaly propagation sequence according to the chronological order and time interval between abnormal events; matching the anomaly propagation sequence with a preset fault propagation pattern library, and outputting corresponding fault level information based on the matching results. The method described in this invention improves early fault identification and fault type differentiation.

Description

Technical Field

[0001] This invention relates to the field of elevator fault prediction technology, specifically to an elevator fault prediction method and system based on multi-source data fusion. Background Technology

[0002] With the development of intelligent manufacturing, elevators are gradually incorporating multi-sensor monitoring to collect operational data such as current, vibration, and displacement to assist in judging the elevator's status. Most existing methods are based on single signals or simple fusion results, identifying anomalies through thresholds or statistical features. Some solutions combine operational phase division and model analysis for fault identification, and have already established a certain application foundation in engineering.

[0003] However, in practice, the utilization of multi-source data remains insufficient. Different signals are often processed separately, lacking unified modeling and correlation analysis, resulting in fragmented state representation and difficulty in forming a holistic characterization of equipment operation. Furthermore, elevators are significantly affected by load changes and start-stop switching during operation. These normal operating condition fluctuations are directly reflected in current, vibration, and speed signals, and without effective differentiation, they are easily misjudged as anomalies. Existing methods mostly rely on single time windows or simple continuity judgments, with little consideration for the temporal relationships between multi-source features, making it difficult to reflect the process of anomaly generation and development, and also hindering the differentiation between transient disturbances and genuine faults. Judgment methods based on fixed rules or single models have limited adaptability under complex operating conditions, and there is still room for improvement in the ability to finely distinguish fault types. Summary of the Invention

[0004] In view of the above-mentioned problems, the present invention is proposed.

[0005] Therefore, the technical problem solved by this invention is that existing elevator fault prediction methods suffer from the following problems: the lack of unified fusion modeling of multi-source operating data leads to scattered state representations; normal operating conditions such as load changes, operating height, and start-stop conditions interfere with fault characteristics, resulting in a high misjudgment rate; the lack of modeling of the temporal correlation and propagation process between various types of fault-sensitive features makes it difficult to distinguish between instantaneous disturbances and actual fault evolution; and how to achieve fault type identification, early warning output, and fault level assessment based on abnormal propagation relationships under complex operating conditions.

[0006] To address the aforementioned technical problems, this invention provides the following technical solution: a method for predicting elevator faults based on multi-source data fusion, comprising: collecting multi-source operating data during elevator operation; performing unified time alignment processing and condition segmentation feature extraction processing on the multi-source operating data to form a multi-source fusion feature vector; identifying condition-related features caused by operating conditions based on the multi-source fusion feature vector and performing weakening processing to construct a condition-bias-free state representation; extracting a set of fault-sensitive features characterizing abnormal operating behavior based on the condition-bias-free state representation; converting window-level anomaly judgment results in the fault-sensitive feature set into abnormal events, constructing an anomaly propagation sequence according to the chronological order and time interval between abnormal events; matching the anomaly propagation sequence with a preset fault propagation pattern library, and outputting corresponding fault level information based on the matching results.

[0007] As a preferred embodiment of the elevator fault prediction method based on multi-source data fusion described in this invention, the collection of multi-source operating data during elevator operation includes: collecting motor load change data through current and voltage sensors installed at the input of the drive motor; collecting mechanical vibration data through vibration sensors installed on the reducer or guide rail structure; collecting elevator operating speed data, displacement data, and acceleration data through encoders or position detection devices; determining the operating stage marker corresponding to each sampling point based on the operating speed data and the acceleration data; the operating stage marker includes the start-up stage, stable operation stage, braking stage, and stop stage; when any data source experiences a sudden change, interruption, fixed value, over-range, or continuous missing data, the data for the corresponding time period is marked as valid, and trend compensation is performed based on adjacent valid time period data.

[0008] As a preferred embodiment of the elevator fault prediction method based on multi-source data fusion described in this invention, the formation of the multi-source fusion feature vector includes: aligning motor load change data, mechanical vibration data, operating speed data, displacement data, and acceleration data according to a unified time axis; using the stopping stage as the boundary condition for adjacent operating stages based on the operating stage markers, and extracting stage features in the starting stage, stable operating stage, and braking stage respectively; extracting current rise rate features and vibration response delay features in the starting stage; extracting current fluctuation range features and vibration amplitude stability features in the stable operating stage; extracting deceleration trend features and vibration attenuation features in the braking stage; and combining the current rise rate features, vibration response delay features, current fluctuation range features, vibration amplitude stability features, deceleration trend features, and vibration attenuation features in a preset order to form a multi-source fusion feature vector corresponding to the operating cycle.

[0009] As a preferred embodiment of the elevator fault prediction method based on multi-source data fusion described in this invention, the following steps are included: constructing a condition-bias-free state representation: constructing a condition description quantity based on multi-source fused feature vectors from multiple operating cycles; dividing the condition description quantity into corresponding condition intervals, and statistically analyzing the mean, fluctuation amplitude, and cross-cycle change trend of each feature component within the condition intervals; performing condition correlation analysis on each feature component based on the condition description quantity; determining a feature component as a condition-related feature when it remains stable under the same operating conditions, exhibits repeatable regular changes under different operating conditions, and does not show sudden abnormal changes; forming a condition-related feature index set based on the condition-related features, and constructing a trend component characterizing the influence of the operating conditions based on the degree of correlation between each condition-related feature and the corresponding condition description quantity; performing trend weakening processing on the non-fault-related change components of the condition-related features; and combining the condition-related features after trend weakening processing with feature components not determined as condition-related features to form a condition-bias-free state representation.

[0010] As a preferred embodiment of the elevator fault prediction method based on multi-source data fusion described in this invention, the extraction of fault-sensitive feature set characterizing abnormal operating behavior includes: calculating the comprehensive deviation of the operating cycle based on the state representation without operating condition bias, and performing coarse screening of the operating cycle according to the comprehensive deviation; when the comprehensive deviation is less than a preset normal deviation threshold, the corresponding operating cycle is included in the normal operating benchmark update sample; when the comprehensive deviation is greater than or equal to the preset normal deviation threshold, a sliding window is performed on the aligned data sequence within the corresponding operating cycle; within each local time window, the normal operating benchmark is called according to the corresponding operating stage; the normal operating benchmark includes a current benchmark, a vibration benchmark, and a speed benchmark; based on the mean current of the window and The current anomaly score is calculated based on the difference between the current mean benchmark and the current dispersion benchmark. The slope of the current change is calculated by combining the current difference at the beginning and end of the window with the time difference to mark the current anomaly. The vibration anomaly score is calculated based on the difference between the window vibration mean and the vibration mean benchmark, as well as the vibration dispersion benchmark. The vibration dispersion is calculated based on the mean of the squared difference between the window vibration values ​​and the vibration mean to mark the vibration anomaly. The displacement discontinuity score is calculated and marked as a motion state anomaly based on the window velocity mean, window velocity range, and window displacement increment compared with the corresponding velocity benchmark and dispersion benchmark. Based on the window marking results of current anomaly, vibration anomaly, and motion state anomaly, a window-level anomaly judgment vector is constructed, and the anomaly window ratio and the anomaly score mean are calculated respectively to form a fault-sensitive feature set.

[0011] As a preferred embodiment of the elevator fault prediction method based on multi-source data fusion described in this invention, the process of constructing the anomaly propagation sequence includes: inputting the window-level anomaly judgment vector into the fault discrimination model according to the time window sequence, and taking the average of the first and last sampling times of each time window as the window center time of the abnormal state; establishing current abnormal state sequences, vibration abnormal state sequences, and motion abnormal state sequences respectively; when the same type of abnormal state is abnormal within a consecutive time window of no less than a preset minimum duration window number, the corresponding continuous window segment is determined as a valid abnormal segment; for each valid abnormal segment, the center time of the first abnormal window is taken as the first abnormal time point. The duration is calculated by adding the center time difference between the last and first abnormal windows to the single window duration. Each valid abnormal segment is encoded as an abnormal event containing the abnormal type, the first abnormal time point, and the duration, and sorted according to the first abnormal time point to obtain a candidate abnormal propagation sequence. The difference between the first abnormal time points of adjacent abnormal events in the candidate abnormal propagation sequence is calculated as the propagation time interval, and the valid propagation connection is determined based on the positive time lag, the upper limit of the propagation time, and the minimum duration. When there are at least two consecutive valid propagation connections within the same running cycle, the corresponding abnormal event chain is encoded as a valid abnormal propagation sequence according to the abnormal type, the first abnormal time point, and the duration.

[0012] As a preferred embodiment of the elevator fault prediction method based on multi-source data fusion described in this invention, the step of outputting corresponding fault level information based on the matching results includes: constructing a preset fault propagation pattern library; the pattern library includes multiple preset fault propagation patterns, each pattern corresponding to a fault type; the preset fault propagation pattern includes anomaly type combinations, anomaly occurrence order, minimum and maximum allowable propagation time intervals between adjacent anomaly events, and minimum duration requirements for each anomaly event; converting each anomaly event in the effective anomaly propagation sequence into a standardized actual propagation sequence by performing field mapping and order reconstruction; and aligning the standardized actual propagation sequence with the preset fault propagation pattern in order. The system calculates the type order matching rate by the ratio of the number of events of the same type to the number of pattern events; it determines whether the actual propagation time interval of each abnormal event falls within the corresponding allowed propagation time interval range and whether the duration is not less than the minimum duration requirement, and calculates the time matching rate by the ratio of the number of satisfied items to the total number of constraint items; it calculates the comprehensive matching score by weighted summation of the type order matching rate and the time matching rate, and outputs the fault type, warning information or normal status according to the full matching threshold and the warning matching threshold; it calculates the fault severity index by weighted summation of the number of abnormal events in the effective abnormal propagation sequence, the average score of each type of abnormality, and the average duration of abnormal events, and outputs the corresponding fault level information.

[0013] As a preferred embodiment of the elevator fault prediction system based on multi-source data fusion described in this invention, the system comprises: a multi-source data acquisition module, a fusion feature construction module, a working condition bias reduction module, a fault-sensitive feature extraction module, an anomaly propagation sequence construction module, and a fault matching output module. The multi-source data acquisition module collects multi-source operating data during elevator operation. The fusion feature construction module performs unified time alignment processing and working condition segmentation feature extraction processing on the multi-source operating data to form a multi-source fusion feature vector. The working condition bias reduction module identifies working condition-related features caused by operating conditions based on the multi-source fusion feature vector and performs reduction processing to construct a state representation without working condition bias. The fault-sensitive feature extraction module extracts a set of fault-sensitive features characterizing abnormal operating behavior based on the state representation without working condition bias. The anomaly propagation sequence construction module converts window-level anomaly judgment results in the fault-sensitive feature set into anomaly events and constructs an anomaly propagation sequence according to the chronological order and time interval between the anomaly events. The fault matching output module matches the anomaly propagation sequence with a preset fault propagation pattern library and outputs corresponding fault level information based on the matching result.

[0014] A computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement a method for predicting elevator faults based on multi-source data fusion.

[0015] A computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of a method for predicting elevator faults based on multi-source data fusion.

[0016] The beneficial effects of this invention are as follows: By constructing multi-source fusion features according to the operating stages, motor load, mechanical vibration, and motion state are transformed from raw sampled data into periodic features with stage-specific physical meaning. This reduces errors caused by mixed judgments of starting shock, stable operation fluctuations, and braking attenuation. By constructing operating condition descriptors and identifying operating condition-related features, repeatable changes caused by speed, load, stroke, and cycle duration are weakened from the state representation. This makes fault judgment rely more on residual anomalies that cannot be explained by normal operating conditions, reducing false alarms caused by heavy loads, short strokes, or speed changes. By further converting window-level anomalies into anomalous events and anomaly propagation sequences and matching them with a fault propagation pattern library containing type order and time constraints, fault prediction can utilize the sequential relationship between anomalies on the motor, mechanical, and motion sides, rather than relying solely on single-point thresholds or isolated scores. This improves the consistency of early fault identification, fault type differentiation, and fault level output. Attached Figure Description

[0017] To more clearly describe the technical solutions of the embodiments of the present invention, the accompanying drawings required in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention, and for those of ordinary skill in the art, other drawings can also be obtained based on these drawings without any creative effort.

[0018] Figure 1 is the overall flowchart of the elevator fault prediction method based on multi-source data fusion provided by the present invention.

[0019] Figure 2 is a schematic diagram of the computer device of the present invention. DETAILED DESCRIPTION

[0020] To make the above objects, features and advantages of the present invention more obvious and easy to understand, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings of the specification. Obviously, the described embodiments are part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0021] referring to Figure 1 , an embodiment of the present invention provides an elevator fault prediction method based on multi-source data fusion, including:

[0022] S1: Collect multi-source operation data during the operation of the elevator.

[0023] Further, collecting the multi-source operation data during the operation of the elevator includes: collecting motor load change data through the current sensor and voltage sensor installed at the input end of the drive motor; collecting mechanical vibration data through the vibration sensor installed on the reducer or guide rail structure; collecting operation speed data, displacement data and acceleration data of the elevator through the encoder or position detection device; determining the operation phase mark corresponding to each sampling point according to the operation speed data and acceleration data; the operation phase mark includes a start phase, a stable operation phase, a braking phase and a stop phase; when any data source has a mutation, interruption, fixed value, out-of-range or continuous missing, the data of the corresponding time period is marked for validity, and trend compensation is performed based on the data of adjacent valid time periods.

[0024] It should be noted that a specific solution for collecting multi-source operation data during the operation of the elevator specifically includes that, in this embodiment, one operation cycle refers to the complete process in which the elevator enters the start phase from the stop phase, and then enters the stop phase again after passing through the stable operation phase and the braking phase. Denote the th operation cycle as , , The number of runtime cycles involved in the analysis. This represents the runtime index. The uniform sampling interval is denoted as... The preferred value is 0.05s. If If it is too large, short-term anomalies during the start-up and braking phases will easily be averaged out; if If the value is too small, vibration noise and communication jitter will significantly increase the computational burden.

[0025] The elevator collects motor input current via a current sensor installed at the drive motor input terminal, motor input voltage via a voltage sensor installed at the drive motor input terminal, and mechanical vibration data via vibration sensors installed on the reducer housing, traction mechanism, transmission bearing housing, or guide rail structure. The elevator's operating speed and displacement data are collected via an encoder or position detection device. Motor input voltage includes motor input current, motor input voltage, and motor input power calculated from the motor input current and motor input voltage. Mechanical vibration data includes raw sampled values ​​of vibration acceleration, root mean square values ​​of vibration acceleration, or vibration amplitude envelope values. Acceleration data is obtained either directly from an acceleration sensor or through velocity data differential analysis. When both acceleration sensor and velocity differential results are available, acceleration sensor data is used preferentially, and velocity differential results are used as consistency verification data.

[0026] To ensure that all data has a clear source and a consistent structure, this embodiment will... In the first running cycle The data record of a uniform sampling point is represented as follows:

[0027]

[0028] in, Indicates the first In the _th operating cycle, the _th Multi-source operational data records from a unified sampling point Indicates the first A unified sampling point index within each running cycle. , Indicates the first The number of uniform sampling points within each operating cycle; Indicates the first In the _th operating cycle, the _th Motor input current at a unified sampling point Indicates the first In the _th operating cycle, the _th Motor input voltage at a unified sampling point Indicates the first In the _th operating cycle, the _th Mechanical vibration values ​​at a uniform sampling point; Indicates the first In the _th operating cycle, the _th The running speed of a unified sampling point Indicates the first In the _th operating cycle, the _th Displacement of a uniform sampling point Indicates the first In the _th operating cycle, the _th Acceleration at a uniform sampling point Indicates the first In the _th operating cycle, the _th A unified sampling point's operational phase marker. Indicates the first In the _th operating cycle, the _th A unified sampling point data validity marker. When acceleration is obtained from velocity difference, it is represented as:

[0029]

[0030] in, Indicates the first In the _th operating cycle, the _th The acceleration corresponding to each sampling point and They represent the first In the _th operating cycle, the _th The velocity of the next sampling point and the previous sampling point after each sampling point This indicates a uniform sampling interval. For the first and last sampling points of the running cycle, the acceleration is calculated using an adjacent one-sided difference method.

[0031] In this embodiment, the first In the _th operating cycle, the _th The running stage marker corresponding to each sampling point It includes the stopping phase, the starting phase, the stable operation phase, and the braking phase. Specifically, when the absolute value of the speed does not exceed the rated speed... When the displacement change of three or more consecutive sampling points is no greater than 2mm, the corresponding sampling point is marked as a stop phase. Since the encoder may experience slight jitter when stationary, setting the stop speed threshold too low may cause the stop phase to be cut off; setting it too high may mistakenly identify low-speed crawling as a stop. Therefore, in this embodiment, the stop speed threshold is preferably set to 2%. This is confirmed by combining the continuous displacement change.

[0032] After the stationary phase, the velocity increases continuously and the acceleration exceeds the rated acceleration. When the speed reaches 10% of the target operating speed and remains there for at least 0.3 seconds, the corresponding segment is marked as the start-up phase. When the speed enters the vicinity of the target operating speed, and the speed fluctuation does not exceed 10% of the target operating speed and the absolute value of the acceleration does not exceed [a certain value], the corresponding segment is marked as the start-up phase. When the speed decreases and the acceleration direction is opposite to the speed direction, and this lasts for at least 0.5 seconds, the corresponding segment is marked as the stable operation phase. When the speed decreases and the acceleration direction is opposite to the speed direction, and this lasts for at least 0.3 seconds, the corresponding segment is marked as the braking phase. For isolated phase markings of less than 0.3 seconds, if the preceding and following phases are the same, they are merged into the phase with the same preceding and following phases; if the preceding and following phases are different, they are merged into the starting phase or braking phase according to the direction of speed change.

[0033] This embodiment sets a data validity flag. , Indicates the first In the _th operating cycle, the _th The validity markers for each sampling point contain the availability status of each data source. It includes six sub-markers: current validity, voltage validity, vibration validity, velocity validity, displacement validity, and acceleration validity. Each sub-marker has a value of 1, 0, or 2; 1 indicates original valid data, 0 indicates invalid data, and 2 indicates data after trend compensation.

[0034] For any data source , It can be If any of the following conditions are met, the corresponding sampling point or time period will be marked as invalid: First, the sampled value exceeds the sensor's range; second, consecutive sampling points are missing or timestamps are interrupted; third, the data change for more than 0.5 seconds is less than the sensor's resolution, and the data source should change according to the operating conditions; fourth, the change in adjacent sampling points exceeds five times the historical normal change range for the same operating phase; fifth, communication status, sensor status, or controller diagnostic information indicates abnormal data. For sudden change judgment, if there is no historical normal change range during the initial deployment phase of the equipment, the sensor's physical allowable change rate will be used as the judgment basis.

[0035] When the duration of the invalid segment does not exceed the preset compensation duration At that time, trend compensation is performed based on data from adjacent valid time periods. Let the nearest valid sampling points before and after the invalid segment be respectively... and Then at any time within the invalid segment The compensation value is expressed as:

[0036]

[0037] in, Indicates the time within the invalid segment The compensation data value, and These represent the times of the nearest valid sampling point before and after the invalid segment, respectively. and These represent the data values ​​of the most recent valid sampling points. Since short-term communication jitter or momentary acquisition failures typically do not change the elevator's operating trend, while prolonged periods of missing data may mask true anomalies, this embodiment preferably uses... When the duration of the invalid segment is greater than In such cases, trend compensation is not performed, but invalid markers are retained, and the data source is excluded or the confidence of the corresponding window is reduced in subsequent window scoring.

[0038] S2: Perform unified time alignment processing and working condition segmentation feature extraction processing on multi-source operating data to form a multi-source fusion feature vector.

[0039] Furthermore, forming a multi-source fusion feature vector includes aligning motor load change data, mechanical vibration data, operating speed data, displacement data, and acceleration data along a unified time axis; using the stopping stage as the boundary condition for adjacent operating stages based on the operating stage markers, and extracting stage features in the starting stage, stable operating stage, and braking stage respectively; extracting current rise rate features and vibration response delay features in the starting stage; extracting current fluctuation range features and vibration amplitude stability features in the stable operating stage; extracting deceleration trend features and vibration attenuation features in the braking stage; and combining the current rise rate features, vibration response delay features, current fluctuation range features, vibration amplitude stability features, deceleration trend features, and vibration attenuation features in a preset order to form a multi-source fusion feature vector corresponding to the operating cycle.

[0040] It should be noted that one specific scheme for forming a multi-source fusion feature vector includes performing unified time alignment processing and condition-segmented feature extraction processing on the multi-source operational data to form a multi-source fusion feature vector corresponding to the operational cycle. Since the sampling frequencies and timestamps of the current sensor, vibration sensor, encoder, and position detection device are different, this embodiment establishes a unified time axis, represented as follows:

[0041]

[0042] in, Indicates the first A unified timeline for each operating cycle. Indicates the first Within the first operating cycle A unified sampling time, Indicates the first The last unified sampling moment within each operating cycle Indicates the first The start time of each running cycle.

[0043] After completing time alignment, based on the runtime phase markers The operating cycle is divided into a start-up phase, a stable operating phase, a braking phase, and a stop phase. The stop phase is used to determine the boundaries between adjacent operating cycles and does not directly participate in phase feature extraction; the start-up phase, stable operating phase, and braking phase are used to extract multi-source fusion features. Let the first... The start-up phase, stable operation phase, and braking phase of each operating cycle are respectively: , and .

[0044] In the startup phase, the current rise rate characteristics and vibration response delay characteristics are extracted and represented as follows:

[0045]

[0046] in, Indicates the first Current rise rate characteristics during the start-up phase of each operating cycle and These represent the 95th and 5th percentiles of the starting current value, respectively. and These represent the start and end times of the startup segment, respectively. The vibration response delay characteristic is expressed as:

[0047]

[0048] in, Indicates the first Vibration response delay characteristics over one operating cycle This indicates the moment when the current in the starting segment first reaches the current-starting response threshold. This indicates the moment when the vibration value in the starting section first reaches the vibration response threshold. The current starting response threshold is the value at a distance of 30% from the 5th to 95th percentile of the starting section current; the vibration response threshold is the value at a distance of 30% from the 5th to 95th percentile of the starting section vibration value. Using relative quantile thresholds can accommodate elevators of different power levels and installation locations, avoiding the inability to migrate fixed absolute thresholds between different devices.

[0049] During the stable operation phase, the characteristics of current fluctuation range and vibration amplitude stability are extracted. The current fluctuation range characteristic is expressed as:

[0050]

[0051] in, This indicates the characteristics of current fluctuation range during stable operation. The stability characteristics of vibration amplitude are expressed as follows:

[0052]

[0053] in, Indicates the first Stability characteristics of vibration amplitude during stable operating segments of each operating cycle. Indicates the first Number of effective vibration sampling points within the stable operating segment of each operating cycle Indicates the first The average vibration during the stable operating period of each operating cycle To prevent extremely small positive numbers with zero denominators in vibration data calculations, it is preferable to use numbers that are twice the resolution of the vibration sensor.

[0054] During the braking phase, deceleration trend characteristics and vibration attenuation characteristics are extracted. The deceleration trend characteristics are represented as follows:

[0055]

[0056] in, Indicates the first Characteristics of deceleration trend during braking phase in each operating cycle and They represent the first The speed at the start and end of the braking segment of each operating cycle. and They represent the first The start and end times of the braking segment in each operating cycle. Vibration attenuation characteristics are expressed as:

[0057]

[0058] in, Indicates the first Vibration attenuation characteristics of the braking section in each operating cycle Indicates the first The average vibration of the first half of the braking section in each operating cycle Indicates the first The average vibration of the second half of the braking section in each operating cycle.

[0059] The first Each feature of the first operating cycle is combined in a preset order to form the second cycle. The multi-source fusion feature vector for each running cycle is represented as:

[0060]

[0061] in, Indicates the first The multi-source fusion feature vector for each operating cycle. If the proportion of valid sampling points is less than 70% in a certain operating phase, the corresponding feature for that phase will not be used as a high-confidence feature in subsequent judgments. When there are historical normal samples in the same operating condition interval, the median of the corresponding feature in that operating condition interval can be used as a substitute, and the confidence of that feature is marked as 0.5. When there are no substitute samples, the confidence of that feature is marked as 0 and will not participate in subsequent weighted calculations.

[0062] S3: Based on multi-source fusion feature vectors, identify operating condition-related features caused by operating conditions and perform weakening processing to construct a state representation without operating condition bias.

[0063] Furthermore, constructing a state representation free from operating condition bias includes: constructing an operating condition description quantity based on multi-source fused feature vectors from multiple operating cycles; dividing the operating condition description quantity into corresponding operating condition intervals, and statistically analyzing the mean, fluctuation amplitude, and cross-cycle change trend of each feature component within the operating condition intervals; performing operating condition correlation analysis on each feature component based on the operating condition description quantity; when a feature component remains stable under the same operating conditions, exhibits repeatable regular changes under different operating conditions, and does not show sudden abnormal changes, the feature component is determined to be an operating condition-related feature; forming an operating condition-related feature index set based on the operating condition-related features, and constructing a trend component representing the impact of operating conditions based on the degree of correlation between each operating condition-related feature and the corresponding operating condition description quantity; performing trend weakening processing on the non-fault change components of the operating condition-related features; and combining the trend-weakened operating condition-related features with feature components not determined to be operating condition-related features to form a state representation free from operating condition bias.

[0064] It should also be noted that one specific scheme for constructing a state representation without operating condition bias includes multi-source fusion feature vectors based on multiple operating cycles. It identifies condition-related features caused by normal operating conditions, performs weakening processing on the non-fault-related change components in these features, and constructs a state representation without operating condition bias.

[0065] In this embodiment, operating condition-related features refer to feature components that exhibit small fluctuations within the same operating condition range, show repeatable trends between different operating condition ranges, and do not possess sudden abnormal characteristics. For example, heavy-load operation may lead to an increased rate of increase in starting current, while high-speed operation may lead to an increase in vibration amplitude during stable operation. If these changes caused by normal operating conditions are not mitigated, normal heavy-load or high-speed operation may be easily misjudged as a fault. Therefore, this embodiment identifies such changes through operating condition descriptors and only mitigates trend components that can be explained by operating conditions, while retaining abnormal residuals that cannot be explained by operating conditions.

[0066] This embodiment uses a historical normal sample set. Establish rules for determining the correlation between operating conditions. The historical normal sample set comes from alarm-free operation cycles during the elevator installation and commissioning phase, manually confirmed normal operation cycles, or operation cycles in subsequent steps where the overall deviation is below the normal deviation threshold and no abnormal propagation sequence has formed. When the number of historical normal samples is less than 30 operation cycles, trend weakening is not performed; only standardization based on the historical normal mean and standard deviation is applied. When the number of historical normal samples reaches 30 or more, the following operating condition correlation feature identification process is performed. For the first... Each operating cycle constructs a working condition description quantity, represented as:

[0067]

[0068] in, Indicates the first The operating condition description quantity for each operating cycle, Indicates the first Average speed during the stable operating segment of each operating cycle Indicates the first Displacement increment within each operating cycle Indicates the first Duration of each running cycle Indicates the first Average input power of the motor per operating cycle Indicates the first The proportion of startup segment time to total runtime of each runtime cycle. Indicates the first The proportion of braking time to total operating cycle time. Calculated from current and voltage data, and expressed as:

[0069]

[0070] Subsequently, the operating condition description quantities are divided into operating condition intervals. Specifically, for , and The system is divided into three intervals—low, medium, and high—based on the ternary quartile of historical normal samples, and these three intervals are combined to form operating condition intervals. If the number of normal samples in a given operating condition interval is less than 10, it is merged with an adjacent operating condition interval. If the number of normal samples is still less than 10 after merging, that operating condition interval is not used for operating condition correlation determination. Let the number of normal samples in the interval be... The set of historical normal operating cycles corresponding to each operating condition interval is as follows: The first feature vector in multi-source fusion The mean and fluctuation range of each characteristic component within this operating condition range are respectively expressed as follows:

[0071]

[0072]

[0073] in, Indicates the first The eigencomponent in the th th eigenvalue 1 The average value within each operating condition interval. Indicates the first The eigencomponent in the th th eigenvalue 1 The fluctuation range within each operating condition range express Number of historical normal operating cycles within each operating condition range Indicates the first The first running cycle Each feature component Indicates feature component index, This represents the index of the operating condition range.

[0074] To determine whether a certain characteristic component remains stable under the same operating conditions, the stability index under the same operating conditions is calculated and expressed as:

[0075]

[0076] in, Indicates the first Stability index of each characteristic component under the same working condition. Indicates the number of effective operating condition intervals. This represents the smallest positive number that prevents the denominator from being zero in characteristic calculations. If If the threshold is too low, most operating condition-related features will not be identified, resulting in insufficient attenuation; if the threshold is too high, features with large fault disturbances may be mistakenly identified as stable. Therefore, this embodiment preferably uses 0.25. Further, in order to determine the... Whether each feature component exhibits repeatable variation with operating conditions, based on the fitting of operating condition trend components using historical normal samples, is expressed as follows:

[0077]

[0078] in, Indicates the first In the _th operating cycle, the _th The trend values ​​of each characteristic component are derived from the interpretation of the operating conditions. Indicates the first Trend intercept of each feature component, Represents the vector of operating condition trend coefficients. This represents the standardized operating condition description quantity. This indicates the transpose sign. The trend coefficient is determined using the least squares method based on historical normal samples. If the number of historical normal samples is insufficient to stably fit the data, the mean of the corresponding operating condition interval is used. As a trend value, it will be updated to a fitted trend value once the sample size meets the requirements. The explanatory power of the operating conditions is expressed as:

[0079]

[0080] in, Indicates the first The interpretability of each feature component under operating conditions This represents the historical normal sample set. Indicates the first The mean of each feature component in historical normal samples. If This indicates that the main changes in this characteristic can be explained by the operating condition descriptor, and it exhibits repeatable operating condition patterns. If Less than This indicates that the first The fluctuation range of each feature component in historical normal samples is lower than the stably calculable explanatory power of the operating conditions, and it does not have sufficient cross-sample variation to determine the explanatory relationship of the operating conditions. At this point, Set to 0 to avoid misjudgment due to unstable values.

[0081] To avoid misinterpreting sudden changes in fault conditions as regular operating conditions, the percentage of sudden deviations is further calculated:

[0082]

[0083] in, Indicates the first The proportion of sudden deviations of each characteristic component Indicates the first The operating condition range to which each operating cycle belongs This is an indicator function that takes the value 1 when the condition is true and 0 when the condition is false. Indicates the first The operating cycle belongs to the operating condition range of the [number]th operating cycle. The mean of each characteristic component. If This indicates that the feature has few sudden deviations in historical normal samples, making it a candidate for operating condition-related features. If the proportion of sudden deviations is high, it indicates that the feature is affected by unexplained abnormal disturbances, and trend weakening is not performed.

[0084] When the Each feature component simultaneously satisfies , and When this condition is determined, the feature component is identified as a condition-related feature, and its index is added to the condition-related feature index set. .

[0085] Determine the set of operating condition-related feature indexes Then, a trend component characterizing the impact of each operating condition is constructed based on the correlation between the relevant characteristics of each operating condition and the descriptive quantities of the operating condition. For those belonging to The characteristic components are fitted with trend values. As a component of the trend of operating conditions, for those that do not belong to The characteristic components are not subjected to trend weakening.

[0086] To avoid negating the true extent of fault changes, a mitigation factor is set based on the interpretability of the operating conditions. The trend mitigation factor is constructed and expressed as:

[0087]

[0088] in, Indicates the first The trend attenuation coefficient for each feature component. Even if a feature is highly correlated with the operating condition, it may still contain early fault residual information; if the operating condition trend is completely removed, the actual fault may also be attenuated. Therefore, the attenuation cap is set to 0.8. By setting the attenuation cap, it is possible to reduce false alarms of the operating condition while retaining fault-sensitive changes.

[0089] For the The first in the running cycle From the feature components, construct the de-biased state components:

[0090]

[0091] in, This represents the state components after removing the operating condition bias. Indicates the first The standard deviations of each feature component in historical normal samples. For operating condition-related features, only the operating condition trend bias relative to the normal mean is weakened; for non-operating condition-related features, their standardized deviations are retained. This forms the first... The debiasing state of each operating cycle is represented as follows , is represented as:

[0092]

[0093] It should also be noted that the identification of operating condition-related features based on multi-source fusion feature vectors from multiple operating cycles is to distinguish between "repeatable fluctuations caused by normal operating condition changes" and "abnormal deviations caused by faults." Under different loads, strokes, speeds, and operating cycle durations, the starting current, vibration amplitude, or braking trend of the elevator may naturally change. If these changes are directly input into fault judgment, heavy-load operation, short-stroke operation, or speed changes will be misidentified as faults. Therefore, an operating condition description quantity including average speed, displacement increment, cycle duration, motor power, starting segment ratio, and braking segment ratio is first constructed. Then, the mean value, fluctuation amplitude, and cross-cycle change trend of the features are statistically analyzed within the operating condition interval. Only features that remain stable under the same operating condition, exhibit repeatable changes under different operating conditions, and do not show sudden deviations are included in the operating condition-related feature index set. Subsequently, a trend component is constructed based on the degree of correlation with the operating condition description quantity, and non-fault trends are weakened. The resulting de-biased state retains residual changes that cannot be explained by the operating condition, making subsequent fault identification more focused on abnormal operating behavior rather than normal load or speed changes.

[0094] S4: Extract a set of fault-sensitive features that characterize abnormal operating behavior based on the state representation without operating condition bias.

[0095] Furthermore, the extraction of fault-sensitive feature sets characterizing abnormal operating behavior includes: calculating the comprehensive deviation of the operating cycle based on the state representation without operating condition bias, and performing coarse screening of the operating cycles based on the comprehensive deviation; when the comprehensive deviation is less than a preset normal deviation threshold, the corresponding operating cycle is included in the normal operating benchmark update sample; when the comprehensive deviation is greater than or equal to the preset normal deviation threshold, a sliding window partition is performed on the aligned data sequence within the corresponding operating cycle; within each local time window, the normal operating benchmark is called according to the corresponding operating stage; the normal operating benchmark includes a current benchmark, a vibration benchmark, and a speed benchmark; and the deviation is calculated based on the difference between the window current mean and the current mean benchmark, and the current dispersion benchmark. The system calculates current anomaly scores, combines the current difference at the beginning and end of the window with the time difference to calculate the slope of current change and mark current anomalies; it calculates vibration anomaly scores based on the difference between the window vibration mean and the vibration mean benchmark, as well as the vibration dispersion benchmark, and marks vibration anomalies based on the mean of the squared difference between the window vibration values ​​and the vibration mean; it calculates displacement discontinuity scores based on the window velocity mean, window velocity range, and window displacement increment, comparing them with the corresponding velocity benchmark and dispersion benchmark, and marks them as motion state anomalies; based on the window marking results of current anomalies, vibration anomalies, and motion state anomalies, it constructs a window-level anomaly judgment vector, and calculates the anomaly window ratio and the anomaly score mean respectively, forming a fault-sensitive feature set.

[0096] It should be noted that one specific scheme for extracting the set of fault-sensitive features representing abnormal operating behavior includes, based on the de-biased state representation... This involves extracting a set of fault-sensitive features characterizing abnormal operating behavior. First, a coarse screening is performed at the operating cycle level, followed by sliding window-level anomaly analysis on suspected abnormal operating cycles. Based on the de-biased state representation, the comprehensive deviation of the operating cycle is calculated, expressed as:

[0097]

[0098] in, Indicates the first The overall deviation over each operating cycle Indicates the first The basic weights of each feature component are preferably... , Indicates the first In the _th operating cycle, the _th The confidence level of the i-th feature component is used to characterize the reliability of the data source for that feature component. A high confidence level is assigned a value of 1, a replacement / compensation feature a value of 0.5, and an invalid feature a value of 0; if the i-th feature component... Each feature component is directly calculated from the original valid sampled data, then... Set it to 1; if the first If each feature component contains trend-compensated data, historical median substitute data, or low-sample substitute data, then... Set it to 0.5; if the first If a feature component cannot be reliably calculated due to insufficient effective sampling points during the corresponding operational phase, prolonged sensor inoperability, or lack of available alternative samples, then... Set to 0. By introducing It can reduce the impact of low-quality features when calculating the overall deviation and eliminate invalid features, thus avoiding the mis-amplification of the degree of abnormality in the operating cycle due to missing data or compensation data.

[0099] Set the preset normal deviation threshold. When the number of historical normal samples is no less than 30 running cycles, The optimal value is the sum of the 90th percentile and 0.5 times the interquartile range of the overall deviation from historical normal samples; when historical normal samples are insufficient, The preferred value is 2. When When a valid anomaly propagation sequence is not formed during the operation cycle, and the anomaly window proportions are all below the preset update exclusion threshold, the corresponding operation cycle will be included in the normal operation benchmark update sample. The normal operation benchmark is maintained separately according to the operation stage, and includes at least the current average benchmark, vibration average benchmark, velocity average benchmark, current dispersion benchmark, vibration dispersion benchmark, and velocity dispersion benchmark. For data sources... Corresponding operational phase The baseline mean and baseline dispersion are updated using a smoothing method, as shown below:

[0100]

[0101]

[0102] in, and They represent the first The data source in the first The updated baseline mean and baseline dispersion for each running phase Indicates the runtime index. and They represent the first The data source in the first The baseline mean and baseline dispersion before each operational phase update Indicates the first The first operating cycle The average value of the data source for each running phase. Indicates the first In the _th operating cycle, the _th The data source in the first Standard deviation of each operating phase This represents the baseline update factor, which is preferably set to 0.05.

[0103] when At that time, a sliding window is performed on the aligned data sequence within the runtime. Window length The preferred value is 1 second, with a sliding step size of 1 second. The preferred value is 0.5s. Record the... In the _th operating cycle, the _th A sliding window is The time at the center of the window is represented as:

[0104]

[0105] in, Indicates the first In the _th operating cycle, the _th The center moment of a sliding window Indicates the sliding window index. and They represent the first In the _th operating cycle, the _th The start and end times of each sliding window are defined. Each window calls the normal operating baseline for the corresponding stage based on its center time. If the proportion of valid sampling points from a data source in a window is less than 70%, that data source will not participate in the corresponding anomaly scoring for that window. If fewer than two data sources are valid in the same window, that window is marked as a low-confidence window and will not be used for anomaly event generation. For current anomalies, a current anomaly score is calculated, expressed as:

[0106]

[0107] in, Indicates the first In the _th operating cycle, the _th Current anomaly scoring for a sliding window Indicates the first In the _th operating cycle, the _th The average current of each sliding window, Indicates the first The running phase corresponding to each sliding window and They represent the first The current mean and current dispersion benchmarks for each sliding window corresponding to the operating phase. Indicates the first In the _th operating cycle, the _th The ratio of the difference between the first and last currents of a sliding window to the duration of the window. Indicates the first The 95th percentile of the slope of the historical normal current change during the operation phase corresponding to each sliding window. This represents the smallest positive number used in current anomaly calculations to prevent the denominator from being zero. This represents the weight of the slope term, preferably 0.4. If... If so, the window is marked as having an abnormal current. For vibration anomalies, a vibration anomaly score is calculated, expressed as:

[0108]

[0109] in, Indicates the first In the _th operating cycle, the _th Vibration anomaly score for each sliding window, Indicates the first In the _th operating cycle, the _th The average vibration value of each sliding window. and They represent the first The vibration mean benchmark and vibration dispersion benchmark for each sliding window corresponding to the operating phase. Indicates the first In the _th operating cycle, the _th The number of valid vibration sampling points within a sliding window This represents the smallest positive number used in vibration anomaly calculations to prevent the denominator from being zero. This represents the weight of the discrete vibration term, preferably 0.5. If... If so, the window is marked as a vibration anomaly. For motion state anomalies, which include velocity anomalies and displacement continuity anomalies, a motion state anomaly score is calculated:

[0110]

[0111] in, Indicates the first In the _th operating cycle, the _th Anomaly score for the motion state of a sliding window Indicates the first In the _th operating cycle, the _th The average speed of each sliding window and They represent the first The speed mean benchmark and speed dispersion benchmark for each sliding window corresponding to the running phase Indicates the first In the _th operating cycle, the _th The sliding window is extremely slow. Indicates the first The 95th percentile of the historical normal window speed range for each sliding window's running phase. Indicates the first In the _th operating cycle, the _th The difference between the first and last displacements of a sliding window Indicates the window length. Indicates the first The standard deviation of the historical normal displacement continuity deviation for each sliding window's operating phase and These represent the weights of the velocity range term and the displacement continuity term, respectively, preferably 0.4 and 0.6. If... If the window is marked as having an abnormal motion state, then based on the window marking results for current anomalies, vibration anomalies, and motion state anomalies, a window-level anomaly determination vector is constructed, represented as:

[0112]

[0113] in, Indicates the first In the _th operating cycle, the _th A window-level anomaly detection vector for a sliding window. , and They represent the first In the _th operating cycle, the _th The sliding window displays abnormal current, abnormal vibration, and abnormal motion status markers, with a value of 1 for abnormality and 0 otherwise. Indicates the first In the _th operating cycle, the _th The confidence level of a sliding window is determined as follows: a window is considered a high-confidence window when at least two data sources within the window are valid, and the proportion of valid sample points from each of the data sources participating in the scoring is not less than 70%. Set to 1; when fewer than two data sources are valid within the window, or when the proportion of valid sample points from any data source participating in anomaly scoring is less than 70% within the sliding window, it is determined to be a low-confidence window. Set it to 0. Form a set of fault-sensitive features, represented as:

[0114]

[0115] in, Indicates the first A set of fault-sensitive characteristics for each operating cycle. , and They represent the first The proportion of abnormal windows in current, vibration and motion status for each operating cycle; , and They represent the first The average score of current, vibration and motion state abnormality for each operating cycle is preferred. The average score of abnormality is preferably the average score of the abnormality in the high confidence window that is marked as the corresponding abnormality. If there is no corresponding abnormality window, the average score of this type of abnormality is recorded as 0 or invalid and is not included in the severity calculation. Indicates the first The number of sliding windows per runtime. If a certain type of anomaly has no valid scoring window, the proportion of windows for that type of anomaly and the average score of the anomaly are considered invalid and will not be included in subsequent severity calculations.

[0116] S5: Convert the window-level anomaly determination results in the fault-sensitive feature set into anomaly events, and construct an anomaly propagation sequence based on the order and time interval between the anomaly events.

[0117] Furthermore, the process of constructing the anomaly propagation sequence includes: inputting the window-level anomaly judgment vector into the fault discrimination model according to the time window sequence, and taking the average of the first and last sampling times of each time window as the center time of the anomaly state window; establishing current anomaly state sequences, vibration anomaly state sequences, and motion anomaly sequences respectively; when the same type of anomaly state is abnormal within a consecutive time window of no less than a preset minimum duration window number, the corresponding continuous window segment is determined as a valid anomaly segment; for each valid anomaly segment, the center time of the first anomaly window is taken as the first anomaly time point, and the center time of the last anomaly window is compared with the first anomaly window. The time difference plus the duration of a single window is used as the duration; each valid abnormal segment is encoded as an abnormal event containing the abnormal type, the first abnormal time point, and the duration, and sorted according to the first abnormal time point to obtain a candidate abnormal propagation sequence; the difference between the first abnormal time points of adjacent abnormal events in the candidate abnormal propagation sequence is calculated as the propagation time interval, and the valid propagation connection is determined according to the positive time lag, the upper limit of the propagation time, and the minimum duration; when there are at least two consecutive valid propagation connections within the same running cycle, the corresponding abnormal event chain is encoded as a valid abnormal propagation sequence according to the abnormal type, the first abnormal time point, and the duration.

[0118] It should be noted that a preferred scheme for constructing anomaly propagation sequences based on various fault sensitivity characteristics specifically includes, in this embodiment, three sequence states generated sequentially: candidate anomaly propagation sequence, effective anomaly propagation sequence, and standardized actual propagation sequence. The candidate anomaly propagation sequence is used for preliminary sorting, the effective anomaly propagation sequence is used to filter event chains with propagation relationships, and the standardized actual propagation sequence is used for matching with a preset fault propagation pattern library to ensure consistent fields.

[0119] The window-level anomaly determination results are converted into anomaly events, and an anomaly propagation sequence is constructed based on the chronological order and time intervals between these events. The fault discrimination model in this embodiment is a rule-based model, which generates anomaly events based on the window-level anomaly determination vector, the number of consecutive anomaly windows, window confidence, and time constraints, without involving a neural network model that requires training. The input to this rule-based model is the window-level anomaly determination vector. The output consists of a sequence of abnormal current states, a sequence of abnormal vibration states, a sequence of abnormal motion states, and the abnormal events generated therefrom. The abnormal sequences are established according to the time window order, as follows:

[0120]

[0121]

[0122]

[0123] in, , and They represent the first The abnormal current state sequence, abnormal vibration state sequence, and abnormal motion state sequence for each operating cycle. , and They represent the first Anomalies in current, vibration, and motion status are marked in the last sliding window of each operating cycle.

[0124] For any exception type , When the same type of abnormal state occurs continuously for no less than a preset minimum duration window number When all segments within a high-confidence window are anomalous, the corresponding continuous window segment is determined as a valid anomalous segment. Preferably, If there is only one low-confidence window in the middle of a series of anomaly windows, and the anomalies before and after the low-confidence window are of the same type, then the low-confidence window is allowed not to interrupt the series; if there are two or more consecutive low-confidence windows, then the anomaly series is interrupted.

[0125] Suppose a certain valid anomaly segment is formed by the first... The window continues until the [number]th [window]. If there are multiple windows, then they are encoded as exception events, represented as:

[0126]

[0127] in, Indicates the first The first in the running cycle An abnormal event. Indicates an index of abnormal events. Indicates the first One exception type; , Indicates the first The first in the running cycle The first abnormal time point of the abnormal event Indicates the first The first operating cycle The center moment of a sliding window; Indicates the first The first in the running cycle The average score of each abnormal event; Indicates the first The first in the running cycle The duration of each anomaly event score is represented as:

[0128]

[0129] in, Indicates the first The first operating cycle The center moment of a sliding window Indicates the length of the sliding window.

[0130] If the time interval between two valid anomaly segments of the same anomaly type does not exceed 1 second, and there are no other high-scoring anomaly events of the same type in between, then the two segments are merged into one anomaly event. All anomaly events within the same runtime are sorted in ascending order of their first anomaly time point to obtain the candidate anomaly propagation sequence, represented as:

[0131]

[0132] in, Indicates the first Candidate anomaly propagation sequence for each running cycle, Indicates the number of abnormal events. Indicates the first The last abnormal event in each runtime cycle. If the difference between the first abnormal event timestamps of two abnormal events is less than the window step size. If two events occur approximately simultaneously, they are considered to have occurred at approximately the same time. Events that occur approximately simultaneously retain their respective anomaly types and durations, but by default do not constitute a forward propagation connection. For two adjacent anomaly events in a candidate anomaly propagation sequence... and The propagation time interval is calculated and expressed as:

[0133]

[0134] in, Indicates the first The first in the running cycle The abnormal time difference between an abnormal event and its adjacent abnormal events. Indicates the first The first in the running cycle The time of the first occurrence of an abnormal event. Indicates the first The first in the running cycle The time of the first occurrence of each anomalous event. A valid propagation connection is considered to exist between the two if the following conditions are met:

[0135]

[0136] in, Indicates the first The first in the running cycle The valid propagation connection flag for each abnormal event is set to 1 if the condition is met, and 0 if the condition is not met. Indicates the first The first in the running cycle The duration of each abnormal event score; Indicates the minimum propagation time; Indicates the maximum propagation time; This represents the minimum event duration, preferably 1 second.

[0137] When there are at least two consecutive valid propagation connections within the same runtime cycle, the corresponding exception event chain is encoded as a valid exception propagation sequence, represented as:

[0138]

[0139] in, Indicates the first An effective anomaly propagation sequence for each running cycle.

[0140] It should also be noted that the window-level anomaly judgment results are converted into anomaly events to avoid directly equating occasional anomalies within a single window with faults. Instantaneous impacts, electromagnetic interference, or local sampling jitter during elevator operation can all cause short-term anomalies. Relying solely on single-window scoring outputs can easily lead to false alarms. Therefore, current anomalies, vibration anomalies, and speed anomalies are each formed into time series, and the same type of anomaly is required to remain abnormal within a window of at least a preset minimum duration, forming effective anomaly segments. Each effective anomaly segment is further encoded as an anomaly event containing the anomaly type, the first anomaly time point, and the duration, preserving the order, duration, and temporal location of the anomaly occurrence. Subsequently, the propagation connection is determined based on the first anomaly time difference between adjacent anomaly events. Only events satisfying positive time lag, upper propagation time limit, and minimum duration are connected as effective propagation chains. The question of "whether multiple anomalies occur simultaneously" is transformed into "whether anomalies propagate in an orderly manner along the motor side, mechanical side, and motion side," reducing the possibility of irrelevant anomalies being combined into fault modes.

[0141] S6: Match the abnormal propagation sequence with the preset fault propagation mode library, and output the corresponding fault level information based on the matching result.

[0142] Furthermore, the corresponding fault level information output based on the matching results includes: constructing a preset fault propagation pattern library; the pattern library includes multiple preset fault propagation patterns, each corresponding to a fault type; the preset fault propagation patterns include combinations of abnormal types, the order of abnormal occurrence, the minimum and maximum allowable propagation time intervals between adjacent abnormal events, and the minimum duration requirement for each abnormal event; converting each abnormal event in the effective abnormal propagation sequence into a standardized actual propagation sequence by mapping fields and reconstructing the order; aligning the standardized actual propagation sequence with the preset fault propagation patterns in order, and calculating the type order matching rate based on the ratio of the number of events of the same type to the number of pattern events; determining whether the actual propagation time interval of each abnormal event falls within the corresponding allowable propagation time interval range, and determining whether the duration is not less than the minimum duration requirement, and calculating the time matching rate based on the proportion of the number of satisfied items to the total number of constraint items; calculating the comprehensive matching score by weighted summation of the type order matching rate and the time matching rate, and outputting the fault type, warning information, or normal status based on the complete matching threshold and the warning matching threshold; calculating the fault severity index by weighted summation of the number of abnormal events in the effective abnormal propagation sequence, the average score of each type of abnormality, and the average duration of abnormal events, and outputting the corresponding fault level information.

[0143] It should be noted that one specific scheme for outputting corresponding fault level information based on the matching results includes, in this embodiment, constructing a preset fault propagation mode library. This mode library includes multiple preset fault propagation modes, each corresponding to a fault type. Each preset fault propagation mode includes anomaly type combinations, anomaly occurrence order, minimum and maximum allowable propagation time intervals between adjacent anomaly events, minimum duration requirements for each anomaly event, and the corresponding fault type. The preset fault propagation mode is represented as follows:

[0144]

[0145] in, Indicates the first A preset fault propagation mode, Indicates the index of fault propagation modes. , Indicates the first The order of exception types in a preset fault propagation pattern. Indicates the first The first preset fault propagation mode One exception type, Indicates the first The number of abnormal events included in each preset fault propagation mode. and They represent the first The lower and upper limits of the allowed propagation time interval between a preset fault propagation mode and events in adjacent modes. Indicates the first The minimum duration requirement for each preset fault propagation mode is preferably 1 second. Indicates the first The fault type corresponding to each preset fault propagation mode.

[0146] The mode library includes the following modes: for brake drag or abnormal transmission load, and The preferred values ​​are 0.5s and 15s, respectively, and the mode sequence is set to current anomaly. → Abnormal vibration →Abnormal movement status The mechanism is that the load on the motor side first increases, followed by enhanced mechanical vibration, ultimately affecting the continuity of speed or displacement. This is especially relevant for guide rail wear or mechanical jamming. and The preferred values ​​are 0.5s and 20s, respectively, and the mode sequence is set to vibration anomaly. → Current abnormality →Abnormal movement status The mechanism is that mechanical friction or impact first generates abnormal vibration, followed by increased motor load compensation, ultimately affecting motion stability. For drive control abnormalities, and The preferred values ​​are 0.5s and 10s, respectively, and the mode sequence is set to abnormal motion state. → Current abnormality → Abnormal vibration The mechanism is that speed control abnormalities occur first, followed by increased motor regulation, which further causes abnormal mechanical response. The modes can be configured with parameters according to different elevator models, rated speeds, rated loads, and maintenance experience, but each mode retains the order of abnormality types, time interval ranges, and duration requirements.

[0147] Each exception event in the effective exception propagation sequence is mapped to fields and its order reconstructed, converting it into a standardized actual propagation sequence, represented as:

[0148]

[0149] in, Indicates the first A standardized actual propagation sequence for each operating cycle. Indicates the first The exception type of each abnormal event. Indicates the first The first abnormal time point of an abnormal event. Indicates the first The duration of each abnormal event. Indicates the first The average anomaly score for each anomalous event This represents the number of anomalous events in the standardized actual propagation sequence.

[0150] Standardize the actual propagation sequence With preset fault propagation mode Perform sequence alignment. During sequence alignment, while maintaining the actual chronological order of the abnormal events, start from... Select with The event subsequence with the highest consistency in type order; when multiple optional subsequences exist, the subsequence with the highest overall matching score is selected. If the number of matched events of the same type is... The type order matching rate is then expressed as:

[0151]

[0152] in, Indicates the first The first operating cycle and the first Type order matching rate among fault propagation modes Indicates the first The first operating cycle and the first When matching fault propagation patterns, the number of abnormal events with the same type order is considered. The actual propagation time interval between aligned events is determined to fall within the pattern's allowed range, and the event duration is determined to meet the minimum duration requirement. The time matching rate is expressed as:

[0153]

[0154] in, Indicates the first The first operating cycle and the first Time matching rate of each fault propagation mode Indicates the first The first operating cycle and the first The number of events that meet the time interval requirement after aligning the fault propagation patterns; Indicates the first The first operating cycle and the first After aligning the fault propagation patterns, the number of abnormal events that meet the minimum duration requirement is calculated. Furthermore, the comprehensive matching score is calculated by weighted summation of the type-order matching rate and the time-matching rate, and is expressed as:

[0155]

[0156] in, Indicates the first The first operating cycle and the first The overall matching score of each fault propagation pattern. Indicates the type order matching weight. This indicates the time-matching weight, and This embodiment is preferred. , The type order weight is slightly higher than the time weight because the fault propagation speed can vary significantly under different loads and different aircraft models, while the order of anomaly occurrence usually better reflects the fault mechanism; however, the time constraint is still used to eliminate random combinations of irrelevant anomalies. For all preset fault propagation modes, the mode with the highest comprehensive matching score is selected as the candidate fault mode, denoted as:

[0157]

[0158] in, This represents the fault propagation mode number with the highest overall matching score. Set the exact match threshold. Matching threshold with early warning Preferably, , ;when When the candidate fault mode is selected, the corresponding fault type is output. .when When [the fault type is specified], output a warning message corresponding to the fault type. When the condition is met, output either the normal state or the state where no known fault propagation pattern is matched. Indicates the first The fault propagation mode with the highest comprehensive matching score in each operating cycle The overall matching score, This indicates the fault type corresponding to the fault propagation pattern with the highest overall matching score.

[0159] Furthermore, the fault severity index is calculated based on the number of anomalous events in the effective anomaly propagation sequence, the average score of each type of anomaly, and the average duration of the anomalous events, and is expressed as:

[0160]

[0161] in, Indicates the first Fault severity index for each operating cycle This indicates the number of reference exceptions, preferably 3; Indicates the first The mean of the average anomaly scores for each anomaly event in the effective anomaly propagation sequence of each running cycle. This indicates a reference anomaly score, with 5 being the preferred value; Indicates the first Average duration of each abnormal event in each running cycle This indicates the reference duration, preferably 5 seconds; , and These represent the weights of the number of abnormal events, the average score of abnormal events, and the average duration, respectively, with preferred values ​​of 0.30, 0.45, and 0.25.

[0162] According to the severity index The fault level information is output based on the fault level range it falls within. When... When the condition is normal, output a level 1 status, indicating normal or slight fluctuation; when... When, output a level 2 status, indicating a minor anomaly or warning; when When the time is right, output a level 3 status, indicating a moderate fault risk, requiring inspection; when When the system is in a critical state, a Level 4 status is output, indicating a severe fault risk requiring immediate maintenance or shutdown for inspection. If the overall matching score is lower than the warning matching threshold, but the fault severity index reaches Level 3 or 4, then the message "Unknown mode high severity anomaly" is output. This prevents novel faults not yet covered by the preset fault propagation mode library from being ignored, improving system security in unknown fault scenarios. The final output result is as follows:

[0163]

[0164] in, Indicates the first Fault prediction output results for each running cycle.

[0165] One embodiment of the present invention provides a lift fault prediction system based on multi-source data fusion, including a multi-source data acquisition module, a fusion feature construction module, a working condition bias reduction module, a fault sensitive feature extraction module, an anomaly propagation sequence construction module, and a fault matching output module.

[0166] The system comprises the following modules: a multi-source data acquisition module for collecting multi-source operational data during elevator operation; a fusion feature construction module for performing unified time alignment and segmented feature extraction on the multi-source operational data to form a multi-source fusion feature vector; a working condition bias reduction module for identifying working condition-related features caused by the operating conditions based on the multi-source fusion feature vector and performing reduction processing to construct a state representation without working condition bias; a fault-sensitive feature extraction module for extracting a set of fault-sensitive features representing abnormal operating behavior based on the state representation without working condition bias; an anomaly propagation sequence construction module for converting window-level anomaly judgment results in the fault-sensitive feature set into anomaly events and constructing an anomaly propagation sequence based on the chronological order and time interval between the anomaly events; and a fault matching output module for matching the anomaly propagation sequence with a preset fault propagation pattern library and outputting corresponding fault level information based on the matching results.

[0167] Reference Figure 2 This embodiment also provides a computer device applicable to the elevator fault prediction method based on multi-source data fusion, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the elevator fault prediction method based on multi-source data fusion as proposed in the above embodiment.

[0168] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0169] This embodiment also provides a storage medium storing a computer program, which, when executed by a processor, implements the elevator fault prediction method based on multi-source data fusion as proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

Claims

1. A method for predicting elevator faults based on multi-source data fusion, characterized in that, include: Collect multi-source operational data during the operation of the elevator; Perform unified time alignment processing and working condition segmentation feature extraction processing on multi-source operational data to form a multi-source fusion feature vector; Based on multi-source fusion feature vectors, identify condition-related features caused by operating conditions and perform weakening processing to construct a state representation without operating condition bias. Extract a set of fault-sensitive features that characterize abnormal operating behavior based on the state representation without operating condition bias; The window-level anomaly determination results in the fault-sensitive feature set are converted into anomaly events, and an anomaly propagation sequence is constructed based on the chronological order and time interval between the anomaly events. The abnormal propagation sequence is matched with a preset fault propagation mode library, and the corresponding fault level information is output based on the matching result.

2. The elevator fault prediction method based on multi-source data fusion as described in claim 1, characterized in that: The multi-source operational data collected during the operation of the elevator includes... Motor load change data are collected by current and voltage sensors installed at the input of the drive motor. Mechanical vibration data is collected by vibration sensors installed on the reducer or guide rail structure; The elevator's operating speed, displacement, and acceleration data are collected using an encoder or position detection device. The running stage marker corresponding to each sampling point is determined based on the running speed data and the acceleration data; The operation phase markers include the start-up phase, stable operation phase, braking phase, and stop phase; When any data source experiences a sudden change, interruption, fixed value, over-range, or continuous missing data, the data for the corresponding time period is marked as valid, and trend compensation is performed based on data from adjacent valid time periods.

3. The elevator fault prediction method based on multi-source data fusion as described in claim 2, characterized in that: The formation of the multi-source fusion feature vector includes, Align the motor load change data, mechanical vibration data, operating speed data, displacement data, and acceleration data according to a unified time axis; Based on the operation phase markers, the stopping phase is used as the boundary condition for adjacent operation phases, and phase features are extracted from the starting phase, stable operation phase, and braking phase respectively. In the startup phase, the current rise rate characteristics and vibration response delay characteristics are extracted. During the stable operation phase, the characteristics of current fluctuation range and vibration amplitude stability are extracted. During the braking phase, the deceleration trend characteristics and vibration attenuation characteristics are extracted. The current rise rate characteristics, vibration response delay characteristics, current fluctuation range characteristics, vibration amplitude stability characteristics, deceleration trend characteristics, and vibration attenuation characteristics are combined in a preset order to form a multi-source fusion feature vector corresponding to the operating cycle.

4. The elevator fault prediction method based on multi-source data fusion as described in claim 3, characterized in that: The construction of the state representation without operating condition bias includes... Based on multi-source fusion feature vectors from multiple operating cycles, a working condition description quantity is constructed. The operating condition description is divided into corresponding operating condition intervals, and the mean, fluctuation range and cross-cycle change trend of each characteristic component are statistically analyzed within the operating condition intervals. Based on the operating condition description, perform operating condition correlation analysis on each feature component; When a certain characteristic component remains stable under the same working conditions, exhibits repeatable regular changes under different working conditions, and does not show sudden abnormal changes, the characteristic component is determined to be a working condition-related characteristic. A set of working condition-related feature indexes is formed based on the working condition-related features, and a trend component representing the impact of working conditions is constructed based on the degree of correlation between each working condition-related feature and the corresponding working condition descriptive quantity. Trend weakening processing is applied to the non-fault-related variation components of operating condition-related characteristics; The condition-related features after trend weakening are combined with the feature components that are not identified as condition-related features to form a state representation without condition bias.

5. The elevator fault prediction method based on multi-source data fusion as described in claim 4, characterized in that: The set of fault-sensitive features extracted to characterize abnormal operating behavior includes, The overall deviation of the operating cycle is calculated based on the state representation after removing the operating condition bias, and the operating cycle is coarsely screened based on the overall deviation. When the overall deviation is less than the preset normal deviation threshold, the corresponding operating cycle will be included in the normal operating benchmark update sample. When the overall deviation is greater than or equal to the preset normal deviation threshold, a sliding window division is performed on the aligned data sequence within the corresponding running cycle; Within each local time window, the normal operating reference is invoked according to the corresponding operating stage; the normal operating reference includes the current reference, vibration reference, and velocity reference. The current anomaly score is calculated based on the difference between the mean current in the window and the current mean benchmark, as well as the current dispersion benchmark. The current change slope is calculated by combining the difference between the first and last currents in the window and the time difference to mark the current anomaly. The vibration anomaly score is calculated based on the difference between the window vibration mean and the vibration mean benchmark, as well as the vibration dispersion benchmark. The vibration dispersion degree is then used to mark the vibration anomaly based on the mean of the squared difference between the window vibration values ​​and the vibration mean. Based on the mean window velocity, the range window velocity, and the increment window displacement, the displacement discontinuity score is calculated and marked as an abnormal motion state by comparing with the corresponding velocity benchmark and dispersion benchmark. Based on the window labeling results of abnormal current, abnormal vibration, and abnormal motion state, a window-level anomaly judgment vector is constructed, and the proportion of abnormal windows and the mean of abnormal scores are calculated respectively to form a set of fault-sensitive features.

6. The elevator fault prediction method based on multi-source data fusion as described in claim 5, characterized in that: The process of constructing the abnormal propagation sequence includes: The window-level anomaly determination vector is input into the fault discrimination model in the order of time windows, and the average value of the first and last sampling times of each time window is taken as the window center time of the anomaly state. Sequences of abnormal current states, abnormal vibration states, and abnormal motion states were established separately. When the same type of abnormal state is abnormal for a continuous time window of no less than the preset minimum duration window number, the corresponding continuous window segment is determined as a valid abnormal segment. For each valid abnormal segment, the center time of the first abnormal window is taken as the first abnormal time point, and the difference between the center time of the last abnormal window and the center time of the first abnormal window plus the duration of a single window is taken as the duration. Each valid abnormal segment is encoded into an abnormal event containing the abnormality type, the time point of the first abnormality, and the duration, and sorted according to the time point of the first abnormality to obtain the candidate abnormality propagation sequence; Calculate the difference between the first abnormal time points of adjacent abnormal events in the candidate abnormal propagation sequence as the propagation time interval, and determine the effective propagation connection based on the positive time lag, the upper limit of the propagation time, and the minimum duration. When there are at least two consecutive valid propagation connections within the same runtime cycle, the corresponding abnormal event chain is encoded into a valid abnormal propagation sequence according to the abnormality type, the time of the first abnormality, and the duration.

7. The elevator fault prediction method based on multi-source data fusion as described in claim 6, characterized in that: The step of outputting corresponding fault level information based on the matching results includes... Construct a library of preset fault propagation modes; the library includes multiple preset fault propagation modes, each mode corresponding to a fault type. The preset fault propagation mode includes combinations of anomaly types, the order in which anomalies occur, the minimum and maximum allowable propagation time intervals between adjacent anomaly events, and the minimum duration requirement for each anomaly event. By mapping fields and reconstructing the order of each anomalous event in the effective anomalous propagation sequence, it is converted into a standardized actual propagation sequence. The standardized actual propagation sequence is sequentially aligned with the preset fault propagation pattern, and the type sequence matching rate is calculated by the ratio of the number of events of the same type to the number of pattern events. Determine whether the actual propagation time interval of each abnormal event falls within the corresponding allowed propagation time interval range, and determine whether the duration is not less than the minimum duration requirement. Calculate the time matching rate based on the proportion of the number of satisfied items to the total number of constraint items. The comprehensive matching score is calculated by weighted summation of the type order matching rate and the time matching rate, and the fault type, warning information or normal status is output according to the complete matching threshold and the warning matching threshold. The severity index is calculated by weighting and summing the number of abnormal events in the effective abnormal propagation sequence, the average score of each type of abnormality, and the average duration of the abnormal events, and the corresponding fault level information is output.

8. A multi-source data fusion-based elevator fault prediction system, employing the multi-source data fusion-based elevator fault prediction method as described in any one of claims 1 to 7, characterized in that: It includes a multi-source data acquisition module, a fusion feature construction module, a working condition bias reduction module, a fault-sensitive feature extraction module, an anomaly propagation sequence construction module, and a fault matching output module; The multi-source data acquisition module is used to collect multi-source operating data during the operation of the elevator. The fusion feature construction module is used to perform unified time alignment processing and working condition segmentation feature extraction processing on multi-source operating data to form a multi-source fusion feature vector. The operating condition bias reduction module is used to identify operating condition-related features caused by operating conditions based on multi-source fusion feature vectors and perform reduction processing to construct a state representation without operating condition bias. The fault-sensitive feature extraction module is used to extract a set of fault-sensitive features that characterize abnormal operating behavior based on the state representation without operating condition bias. The anomaly propagation sequence construction module is used to convert the window-level anomaly judgment results in the fault-sensitive feature set into anomaly events, and construct an anomaly propagation sequence according to the order and time interval between the anomaly events. The fault matching output module is used to match the abnormal propagation sequence with a preset fault propagation pattern library and output the corresponding fault level information based on the matching result.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the elevator fault prediction method based on multi-source data fusion as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the elevator fault prediction method based on multi-source data fusion as described in any one of claims 1 to 7.

Images