Anomaly detection and failure prediction method and device for an elevator door operator system

Abstract

The application belongs to the technical field of elevator fault diagnosis, and particularly relates to an abnormality detection and fault prediction method and device for an elevator door machine system, the method comprising: synchronously collecting vibration, current, image and structural strain data in the elevator door machine system, constructing a unified multi-dimensional time sequence after alignment and feature extraction, then calculating the reconstruction error thereof by using an unsupervised autoencoder based on a gated recurrent unit, and comparing the reconstruction error with a dynamic threshold value which is adaptively adjusted with the running cycle of the device, to accurately determine whether the system is abnormal, and once it is determined to be abnormal, an integrated learning classifier is immediately started to output the specific fault type, occurrence probability and key cause feature identification, the application introduces real-time monitoring of the structural quasi-static strain, effectively solving the problem that it is difficult to early perceive and distinguish structural slowly changing faults and realize explainable intelligent diagnosis under the condition of scarce fault samples.

Description

Technical Field

[0001] This invention belongs to the field of elevator fault diagnosis technology, specifically relating to a method and device for anomaly detection and fault prediction of an elevator door operator system. Background Technology

[0002] As the core actuator of elevators, which are frequently started and stopped, the elevator door operator system's operating status directly affects passenger safety and operational efficiency. Traditional scheduled maintenance and reactive repair methods suffer from problems such as delayed response and high costs. Therefore, predictive maintenance based on condition monitoring has become an important development direction.

[0003] Currently, data-driven intelligent diagnostic methods are mainly divided into two categories. The first category is based on supervised deep learning. For example, Chinese patent document CN119557671A, entitled "A Method and System for Predicting and Diagnosing Elevator Operation Anomalies Based on the Internet of Things," describes how this method collects elevator operation audio, performs time-frequency transformation and feature extraction, and then uses a deep convolutional neural network for fault classification. While this type of method can achieve end-to-end fault identification, its performance is highly dependent on a large number of labeled fault samples. In actual elevator operation and maintenance scenarios, severe fault samples are scarce, and labeling costs are extremely high, leading to difficulties in model training. Furthermore, it has weak generalization ability for novel fault types not present in the training set.

[0004] The second category is based on feature rules and dynamic thresholds. As described in Chinese patent document CN120903347A, entitled "A Method for Anomaly Detection and Prediction of Elevator Door Operator System," this method collects data from multiple sensors such as vibration and current sensors, extracts features, and establishes a dynamic benchmark range that changes with the number of operations. Anomalies are determined by comparing real-time features with the benchmark range, and then a predefined fault mode library is used for diagnosis. This method reduces reliance on fault samples, but its diagnostic capabilities are limited by rules and mode libraries predefined by expert experience, making it difficult to characterize complex and coupled fault mechanisms and lacking flexibility. Furthermore, its diagnostic results are usually only fault type codes, lacking in-depth explanations of the judgment criteria, making it difficult for maintenance personnel to quickly locate the root cause of the fault.

[0005] In summary, current solutions lack a mechanism that can adaptively quantify and accurately predict the overall health status of a single piece of equipment based on its performance degradation trajectory throughout its entire lifecycle, without relying on a large number of fault samples. For example, while patent document CN120903347A employs multiple sensors and dynamic thresholds, the monitored vibration and current signals are dynamic operating parameters, primarily used to characterize the instantaneous motion and load of the equipment. This type of method has a fundamental limitation: it cannot effectively perceive and quantify the gradual stress relaxation, installation foundation deformation, or cumulative fatigue that occur in the mechanical structure of the gantry crane system during long-term service. For instance, problems such as installation stress relaxation of the gantry crane beam and slow settlement of the guide rail bracket will first manifest as abnormal distribution or long-term drift of static stress in structural components before causing obvious vibration or current anomalies. Currently, the lack of direct monitoring methods for such quasi-static structural states makes it difficult to provide early warnings of structural faults such as installation frame instability and guide rail alignment deterioration, and also makes it impossible to effectively distinguish whether the overall performance degradation of the equipment stems from normal mechanical wear or the onset of dangerous structural damage. Summary of the Invention

[0006] The technical problem to be solved by this invention is to overcome the difficulty of accurately detecting early anomalies and providing intelligent diagnosis that can guide maintenance in elevator door operator systems under the realistic condition of scarce fault samples.

[0007] To address the aforementioned technical problems, a first aspect of the present invention discloses a method for anomaly detection and fault prediction in an elevator door operator system, the method comprising:

[0008] Simultaneously collect vibration sensing data, current sensing data, image sensing data, structural strain sensing data, and suspension rope early warning data within a single operating cycle of the elevator door operator;

[0009] The collected multi-source sensor data is time-stamp aligned and features are extracted. The extracted numerical features are arranged in chronological order to construct a unified multi-dimensional time series.

[0010] The multidimensional time series is input into an unsupervised autoencoder based on a gated cyclic unit, and its reconstruction error is calculated.

[0011] The reconstruction error is compared with an adaptive dynamic threshold, which is dynamically calculated and determined based on the number of historical operating cycles of the elevator door operator and its corresponding historical reconstruction error sequence; if the reconstruction error exceeds the threshold, it is determined to be abnormal and a diagnosis is triggered.

[0012] In response to the aforementioned anomaly trigger, the corresponding multidimensional time series is input into an ensemble learning classifier, which outputs a fault type prediction, a prediction probability, and the key feature identifier that contributes most to the current classification decision.

[0013] As an optional implementation, in the first aspect of the present invention, the synchronous acquisition of vibration sensing data, current sensing data, and image sensing data within a single operating cycle of the elevator door operator includes:

[0014] The vibration acceleration time-series signal of at least one axis is collected by a vibration acceleration sensor installed on the gantry crane;

[0015] The timing signal of the motor drive current is collected by a current sensor connected in series in the door operator motor circuit;

[0016] An image sensor installed on the side of the door operator collects an image sequence covering the entire door opening and closing process;

[0017] By using fiber optic strain sensors installed on the key load-bearing structure of the gantry crane, the static or quasi-static strain time-series signal of the structure is collected within a single operating cycle.

[0018] It also includes acquiring the warning status data of the suspension rope warning device, which is embedded in the elevator door and used to provide early warnings for elevator safety.

[0019] As an optional implementation, in the first aspect of the present invention, the step of performing timestamp alignment and feature extraction on the collected multi-source sensor data, and arranging the extracted numerical features in chronological order to construct a unified multidimensional time series, includes:

[0020] The vibration acceleration timing signal, the driving current timing signal, and the image sequence are aligned to a unified time axis using timestamp synchronization and interpolation methods.

[0021] For the aligned vibration acceleration signal, extract its root mean square value in the time domain, peak value, and spectral centroid features in the frequency domain;

[0022] For the aligned drive current signal, extract its root mean square value and peak value in the time domain;

[0023] For the aligned strain time series signal, extract its average strain value, strain peak value and strain change gradient characteristics throughout the entire running cycle;

[0024] For the aligned image sequence, the average optical flow amplitude characteristics are calculated using the optical flow method;

[0025] At each synchronization point, all the features extracted above are concatenated into a feature vector, and the feature vectors of all time points are arranged in chronological order to form the multidimensional time series.

[0026] As an optional implementation, in the first aspect of the present invention, the step of inputting the multidimensional time series into an unsupervised autoencoder based on a gated cyclic unit and calculating its reconstruction error includes:

[0027] The encoder of the autoencoder encodes the input multidimensional time series into a context vector;

[0028] The decoder of the autoencoder reconstructs an output sequence based on the context vector;

[0029] The mean square error of each data point between the original input sequence and the reconstructed output sequence is calculated as the reconstruction error.

[0030] As an optional implementation, in the first aspect of the present invention, the adaptive dynamic threshold is dynamically calculated and determined based on the number of historical operating cycles of the elevator door operator and its corresponding historical reconstruction error sequence. The specific determination method includes:

[0031] Obtain the reconstruction error value of the elevator door operator under N consecutive normal cycles;

[0032] Using the cycle number as the independent variable and the reconstruction error value as the dependent variable, a linear baseline is fitted to characterize the natural trend of its variation with the number of cycles.

[0033] Calculate the standard deviation of the residuals of each historical error value relative to the linear baseline. ;

[0034] For the Each cycle has a corresponding adaptive dynamic threshold. The calculation formula is: ,in For the linear baseline at the 1st The predicted value of the cycle, These are preset coefficients.

[0035] As an optional implementation, in the first aspect of the present invention, the key feature identifier is one or more feature dimension identifiers determined based on the internal decision rules of the gradient boosting decision tree model, which are determined through the following steps:

[0036] After the gradient boosting decision tree model classifies the input abnormal multidimensional time series, it traverses all decision trees in the model and counts the total amount of impurity reduction brought about by each feature dimension when splitting at the nodes of all decision trees, which is used as the global importance score of that feature dimension.

[0037] All feature dimensions are sorted in descending order of global importance score, and the top R feature dimensions are selected, where R is a preset positive integer.

[0038] The identifiers of the selected top R feature dimensions are output as the key feature identifiers.

[0039] As an optional implementation, in the first aspect of the present invention, after constructing the multidimensional time series, a step of dividing the operation phase is further included:

[0040] Based on the waveform characteristics of the driving current timing signal, the start-up, constant speed, and braking phases of a single operating cycle are automatically identified and divided.

[0041] The data points in the multidimensional time series are associated with their respective operational phases.

[0042] As an optional implementation, in the first aspect of the present invention, the key feature identifier is associated with information about the specific operational stage to which it belongs.

[0043] The second aspect of this invention discloses a multi-dimensional anomaly detection and fault prediction device for an elevator door operator system, used to implement the anomaly detection and fault prediction method for an elevator door operator system described in any of the above embodiments, the system comprising:

[0044] The synchronous acquisition module is used to synchronously acquire vibration sensing data, current sensing data, image sensing data, and structural strain sensing data within a single operating cycle of the elevator door operator.

[0045] The time series construction module is used to perform timestamp alignment and feature extraction on the multi-source sensor data collected by the synchronous acquisition module, and arrange the extracted numerical features in chronological order to construct a unified multi-dimensional time series.

[0046] An unsupervised anomaly detection module, which is built into an unsupervised autoencoder based on a gated cyclic unit, is used to receive the multidimensional time series output by the time series construction module, calculate its reconstruction error, and compare the reconstruction error with an adaptive dynamic threshold. If the threshold is exceeded, an anomaly signal is triggered.

[0047] The supervised fault diagnosis module is built into an integrated learning classifier. In response to the abnormal signal triggered by the unsupervised anomaly detection module, it receives the corresponding multidimensional time series and outputs the fault type prediction, prediction probability and key feature identifier.

[0048] The dynamic threshold calculation module is used to dynamically calculate and update the adaptive dynamic threshold used by the unsupervised anomaly detection module based on the number of historical operating cycles of the elevator door operator and its corresponding historical reconstruction error sequence.

[0049] A third aspect of this invention discloses another multi-dimensional anomaly detection and fault prediction device for an elevator door operator system, the system comprising:

[0050] Memory containing executable program code;

[0051] A processor coupled to the memory;

[0052] The processor calls the executable program code stored in the memory to execute the abnormal detection and fault prediction method for an elevator door operator system disclosed in the first aspect of the present invention.

[0053] The fourth aspect of the present invention discloses a computer-readable storage medium storing computer instructions, which, when invoked by a processor, are used to execute an anomaly detection and fault prediction method for an elevator door operator system disclosed in the first aspect of the present invention.

[0054] Compared with the prior art, the beneficial effects of the present invention are:

[0055] 1. By employing an unsupervised autoencoder based on a gated cyclic unit as the anomaly detection module, the system can primarily utilize normal equipment operation data for modeling, thereby reducing the model training's dependence on scarce fault samples and achieving early anomaly perception even in the absence of fault labels. An adaptive dynamic threshold, generated based on the equipment's own historical operating cycle count, is configured for this module. This design allows the anomaly judgment criteria to be synchronously adjusted following the natural wear process of individual equipment, thus improving the early warning system's adaptability to the aging process of individual equipment and reducing false alarms. Vibration, current, and door movement images, among other multi-source data, are explicitly fused, and the preliminary anomaly signal and its corresponding original multi-dimensional features are jointly input into the integrated learning model for fault diagnosis and attribution analysis. This approach enhances the comprehensiveness of diagnosis through multi-sensor information complementarity and provides clear technical basis for maintenance personnel to locate the root cause of the fault through the fault type and key causal feature identifiers output by the model. This makes the diagnostic results directly guide on-site maintenance. Overall, this invention constructs a complete technical path from multi-dimensional perception and adaptive anomaly detection to interpretable fault diagnosis.

[0056] 2. By synchronously acquiring the static strain time-series signal of the key load-bearing structure of the gantry crane, a fourth-dimensional monitoring perspective is introduced into this invention. Vibration signals reflect transient impacts, current signals reflect dynamic loads, while strain signals directly characterize the static loads and cumulative stresses borne by the structural components during long-term operation. The introduction of this dimension of data solves two problems that are difficult to effectively distinguish using traditional methods: first, the slow-developing structural relaxation or foundation settlement caused by long-term use of the equipment; and second, the abnormal stress concentration caused by sudden mechanical overloads or collisions. Attached Figure Description

[0057] The specific embodiments of the present invention will be further described in detail below with reference to the accompanying drawings, wherein:

[0058] Figure 1 This is a flowchart illustrating an abnormality detection and fault prediction method for an elevator door operator system disclosed in an embodiment of the present invention.

[0059] Figure 2 This is a schematic diagram of the structure of a multi-dimensional anomaly detection and fault prediction device for an elevator door operator system disclosed in an embodiment of the present invention;

[0060] Figure 3 This is a schematic diagram of the structure of another multi-dimensional anomaly detection and fault prediction device for an elevator door operator system disclosed in an embodiment of the present invention. Detailed Implementation

[0061] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0062] The terms "first," "second," etc., used in the specification, claims, and accompanying drawings of this invention are used to distinguish different objects, not to describe a specific order. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, apparatus, product, or end that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to these processes, methods, products, or ends.

[0063] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of the invention. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0064] This invention discloses an anomaly detection and fault prediction method and device for an elevator door operator system. By employing an unsupervised autoencoder based on a door control cycle unit as the anomaly detection module, the system can primarily utilize normal operating data of the equipment for modeling, thereby reducing the dependence of model training on scarce fault samples and achieving early anomaly perception even in the absence of fault labels. An adaptive dynamic threshold, generated based on the equipment's own historical operating cycle count, is configured for this module. This design allows the anomaly judgment criteria to be adjusted synchronously with the natural wear and tear process of individual equipment, thereby improving the adaptability of the early warning system to the aging process of individual equipment and reducing false alarms. The invention explicitly integrates multi-source data such as vibration, current, and door movement images, and inputs the preliminary anomaly signal and its corresponding original multi-dimensional features into an integrated learning model for fault diagnosis and attribution analysis. This approach enhances the comprehensiveness of diagnosis through multi-sensor information complementarity and provides clear technical basis for maintenance personnel to locate the root cause of the fault through the fault type and key causal feature identifiers output by the model, making the diagnostic results directly applicable to on-site maintenance.

[0065] Example 1

[0066] Please see Figure 1 , Figure 1 This is a flowchart illustrating an anomaly detection and fault prediction method for an elevator door operator system disclosed in an embodiment of the present invention. Figure 1 The described method for anomaly detection and fault prediction in an elevator door operator system is applied to a data processing chip, processing terminal, or processing server. The processing server can be a local server or a cloud server; this embodiment of the invention does not limit the specific application. Figure 1 As shown, the multi-dimensional anomaly detection and fault prediction method for the elevator door operator system may include the following operations:

[0067] 101. Synchronously collect vibration sensing data, current sensing data, image sensing data, and structural strain sensing data within a single operating cycle of the elevator door operator.

[0068] Specifically, by simultaneously collecting three types of sensor data with different physical properties—vibration, current, and images—a multi-dimensional and complementary information foundation is built for the status monitoring of elevator door machine systems. Vibration data directly reflects the impact and wear of the mechanical structure, current data characterizes the electrical load and efficiency of the motor drive, and image data provides information on the spatial pose and visual appearance of the door during its movement.

[0069] It is evident that this multi-source synchronous acquisition method enables subsequent analysis to make comprehensive judgments based on both internal signal characteristics and external visual manifestations, providing complete data input for detecting mechanical faults, electrical anomalies, and visible physical obstructions or deformations, thereby supporting more comprehensive fault diagnosis capabilities at the information source level.

[0070] 102. Perform timestamp alignment and feature extraction on the collected multi-source sensor data, arrange the extracted numerical features in chronological order, and construct a unified multidimensional time series.

[0071] Specifically, by aligning the timestamps of multi-source asynchronous sensing data, the problem of data lag and misalignment caused by independent acquisition by different sensors is solved, ensuring that vibration, current and image data have a consistent time reference when characterizing the same running event. The subsequent feature extraction process transforms the synchronized high-dimensional original waveform and image frame sequence into a set of low-dimensional numerical feature indicators with clear physical or statistical significance, thus completing the transformation from the original signal to the state representation quantity.

[0072] As can be seen, the data processing flow outputs a standardized sequence that is structurally regular, multi-dimensionally integrated, and time-aligned. This sequence serves as a unified input format for subsequent unsupervised and supervised learning models, enabling the algorithm to jointly model and analyze device states based on a consistent timeline and comparable feature scales. This provides a reliable data foundation for effectively detecting and diagnosing anomalies from heterogeneous signals.

[0073] 103. Input the multidimensional time series into an unsupervised autoencoder based on a gated cyclic unit and calculate its reconstruction error.

[0074] Specifically, by using an unsupervised autoencoder based on gated cyclic units to model multidimensional time series, the model can learn the temporal characteristics and change patterns of the equipment under normal operating conditions. This method does not rely on labeled fault data, but captures the inherent patterns of health data through the encoding and reconstruction process. The reconstruction error output by the model directly quantifies the degree of deviation between the current data and the learned normal pattern.

[0075] It is evident that this design implements an anomaly perception mechanism that does not rely on prior knowledge of faults. By calculating and evaluating reconstruction errors, the system can sensitively indicate anomalies that deviate statistically from historical normal operating conditions, including unseen or early fault modes, thereby providing effective trigger signals and judgment criteria for subsequent fault diagnosis.

[0076] 104. The reconstruction error is compared with an adaptive dynamic threshold, which is dynamically calculated and determined based on the number of historical operating cycles of the elevator door operator and its corresponding historical reconstruction error sequence; if the reconstruction error exceeds the threshold, it is determined to be abnormal and a diagnosis is triggered.

[0077] Specifically, anomaly detection is achieved by comparing the reconstruction error with an adaptive threshold dynamically calculated based on the device's historical operating data. The calculation of this threshold incorporates the number of historical operating cycles of the device as a key parameter, enabling it to adjust synchronously with the natural aging trajectory of the device's performance, rather than using a fixed value.

[0078] It is evident that this adaptive threshold design based on the equipment lifecycle enables the anomaly judgment criteria to inherently adapt to the performance degradation process of a single device. This improves the accuracy and practicality of the early warning system when distinguishing between normal aging of equipment and sudden real failures, and reduces misjudgments caused by unreasonable threshold settings.

[0079] 105. In response to the aforementioned abnormality trigger, the corresponding multidimensional time series is input into an ensemble learning classifier, which outputs the fault type prediction, the prediction probability, and the key feature identifier that contributes the most to this classification decision.

[0080] Specifically, by inputting the multidimensional time series that triggered the anomaly into the ensemble learning classifier after an anomaly is detected, a decision cascade from anomaly perception to specific fault diagnosis is realized. The classifier can output specific fault type hypotheses and their corresponding probabilities of occurrence based on the input feature sequences, and simultaneously analyze and derive the key feature identifiers that have the greatest impact on this classification decision.

[0081] It is evident that this cascaded diagnostic mechanism, while completing fault classification, provides interpretable output of the judgment basis. The system not only tells "what the possible fault is", but also indicates "what abnormal features were mainly used to make the judgment", which transforms the diagnostic conclusion from a single code label into information containing reasoning basis, thereby enhancing the direct guiding value of the results for on-site maintenance work.

[0082] As an optional embodiment, the step above, namely, synchronously acquiring vibration sensing data, current sensing data, and image sensing data within a single operating cycle of the elevator door operator, includes:

[0083] The vibration acceleration time-series signal of at least one axis is collected by a vibration acceleration sensor installed on the gantry crane;

[0084] The timing signal of the motor drive current is collected by a current sensor connected in series in the door operator motor circuit;

[0085] An image sensor installed on the side of the door operator collects an image sequence covering the entire door opening and closing process;

[0086] By using fiber optic strain sensors installed on the key load-bearing structure of the gantry crane, the static or quasi-static strain time-series signal of the structure is collected within a single operating cycle.

[0087] It also includes acquiring the warning status data of the suspension rope warning device, which is embedded in the elevator door and used to provide early warnings for elevator safety.

[0088] In this embodiment of the invention, by clearly defining the selection, installation location, and acquisition signals of various types of sensors, a precise and complementary physical information source is constructed for subsequent analysis;

[0089] Vibration acceleration sensors, current sensors, and image sensors are explicitly used. The vibration acceleration sensor, installed on the door operator body, directly senses the impact, friction, and vibration of mechanical components during operation. Its acquired time-series signals contain rich mechanical state information. The current sensor, connected in series in the motor drive circuit, accurately measures real-time current changes during motor operation. This signal directly reflects changes in the motor's electrical load, output torque, and encountered resistance. The image sensor is fixed at a position where the door's movement path can be fully observed. Its acquired image sequences visually record the door's displacement, velocity changes, and physical appearance. Fiber optic strain sensors, glued or welded to key load-bearing structures of the elevator door operator (such as the middle of the door operator beam or the surface of the guide rail support), acquire the strain time-series signals of the structure during its operating cycle. The fiber optic strain sensor contains a grating of a specific wavelength. When the sensor undergoes slight tensile or compressive deformation (i.e., strain) along with the structural components, the grating period changes, causing a linear shift in the center wavelength of its reflection or transmission. This wavelength change ( ) and structural strain ( They are directly proportional, and the relationship is: ,in The strain sensitivity coefficient of the sensor is denoted as . Using the initial center wavelength as a reference, the center wavelength of the sensor's reflection spectrum is scanned and calculated in real time by an external fiber optic demodulator, thus obtaining a high-precision micro-strain independent of temperature and electromagnetic interference. The timing signal directly reflects the entire process of "loading, holding, and unloading" of structural stress at the monitoring point during the door opening and closing cycle.

[0090] As can be seen from the perspective of information complementarity, these three types of data depict the same operating process from different physical dimensions. Vibration signals reflect the state from the mechanical structure response level, current signals reflect the load from the electrical drive input level, and image signals directly observe the results from the external visual motion level. This multi-source synchronous acquisition method enables the system to simultaneously acquire the internal state excitation of the equipment, the load feedback of the drive system, and the final appearance and motion performance of the actuator, thus providing the possibility for cross-verification and fault diagnosis from multiple angles.

[0091] Furthermore, it is explicitly required that "at least one axial vibration acceleration timing signal" be collected, rather than the general term "vibration signal".

[0092] Vibration acceleration is a vector with a definite direction. The physical sources and fault modes characterized by vibrations in different directions (axial directions) of an elevator door operator typically differ. For example:

[0093] Vibrations parallel to the direction of door movement may primarily originate from guide rail friction, drive belt wear, or jamming.

[0094] Vibrations perpendicular to the plane of the door may more likely reflect collisions, impacts, or loose parts.

[0095] Vertical vibrations may be related to the installation foundation, suspension structure, or overall resonance.

[0096] Measuring only the overall intensity of the "vibration" will lose this crucial directional information, making it impossible to accurately trace and diagnose the fault. Defining the "axial direction" is to obtain vector information with clear physical meaning, which is the basis for subsequent effective feature analysis and fault attribution.

[0097] Acquiring axial signals is to obtain directional vibration vector information for accurate diagnosis. The axial direction includes at least one direction, which can be X (lateral), Y (direction of motion), and Z (vertical). These three axes constitute a spatial rectangular coordinate system. The X-axis (lateral), Y-axis (direction of motion), and Z-axis (vertical) are used to achieve complete monitoring of the gantry crane's three-dimensional spatial vibration, thereby enabling a more comprehensive characterization of its mechanical state and providing the richest information dimensions for diagnosis.

[0098] Furthermore, in engineering practice, using a current sensor connected in series in the door motor circuit means connecting a current sensor (such as a Hall effect current sensor) in series to the power supply line of the elevator door motor (e.g., any one or all of the three-phase power lines U, V, and W). When the motor is running, current flows through the sensor, and the sensor outputs a voltage or current signal proportional to the real-time current magnitude. This continuous analog signal is then digitized by a data acquisition device at a fixed sampling rate (e.g., several thousand times per second) to obtain a current value sequence arranged in chronological order, i.e., the "motor drive current timing signal".

[0099] Current data directly “characterizes the electrical load and efficiency of the motor drive system.” The motor current is the direct power input that drives the door movement. Its magnitude and changes accurately reflect the motor’s output and load changes. Any mechanical resistance (such as guide rail friction or jamming) or electrical problems (such as voltage instability or winding abnormalities) will be immediately reflected in the current waveform.

[0100] Current waveforms naturally exhibit clear phase characteristics:

[0101] Start-up: The current rises instantaneously to its peak value to overcome static friction and generate starting torque.

[0102] Uniform speed: The current drops and remains relatively stable, maintaining the door leaf at a uniform speed.

[0103] Braking: The current undergoes a step change (dropping to zero or reversing) to achieve deceleration and locking.

[0104] By analyzing the current timing signal, these inflection points can be automatically and accurately identified, thus facilitating subsequent phased and refined analysis.

[0105] It should also be noted that when the system diagnoses a certain fault (such as "poor guide rail lubrication") based on vibration and current characteristics, if the slight shaking or lag of the door leaf at the corresponding stage (such as the braking stage) can be visually observed in the associated image sequence, then the diagnostic result will be supported by intuitive visual evidence, which will enhance the credibility of the diagnostic report and its persuasiveness to maintenance personnel, and extend "interpretable fault diagnosis" from textual feature description to visual on-site evidence.

[0106] The core purpose of acquiring image sequences is to transform direct, high-information visual observations into quantified motion features that can be fused with time-series signals, thereby achieving a closed loop from electromechanical to visual at the information source level to support more comprehensive, reliable, and interpretable intelligent fault diagnosis.

[0107] To further explain, the suspension rope warning device is embedded in the elevator door and is an independent mechanical safety warning module. The device includes a warning suspension rope whose design safety factor is lower than that of the elevator main suspension steel wire rope, a fixing block suspended at the end of the warning suspension rope, and a sensor (such as a position switch or proximity sensor) set at a predetermined position below the fixing block.

[0108] The warning logic is as follows: When the elevator is running normally and smoothly, the warning rope and its suspended fixing block remain stable, and the sensor below is not triggered. When the elevator experiences abnormal and violent shaking, top-rushing, or bottom-falling during operation, the impact force acting on the elevator car will exceed the withstand limit of the warning rope (because its safety factor is lower), causing it to break first. The fixing block then falls and triggers the sensor below, thereby generating a warning signal indicating a "dangerous state." This signal is the "rope warning data."

[0109] By synchronously collecting this mechanical early warning signal and correlating and verifying it with the intelligent anomaly judgment result obtained from multi-source sensor data analysis, the system can achieve dual verification of hardware limit alarm and software early prediction, providing an independent hardware verification guarantee for safety early warning in extreme cases.

[0110] As an optional embodiment, the steps described above, including timestamp alignment and feature extraction of the collected multi-source sensor data, arranging the extracted numerical features in chronological order to construct a unified multidimensional time series, include:

[0111] The vibration acceleration timing signal, the driving current timing signal, and the image sequence are aligned to a unified time axis using timestamp synchronization and interpolation methods.

[0112] For the aligned vibration acceleration signal, extract its root mean square value in the time domain, peak value, and spectral centroid features in the frequency domain;

[0113] For the aligned drive current signal, extract its root mean square value and peak value in the time domain;

[0114] For the aligned strain time series signal, extract its average strain value, strain peak value and strain change gradient characteristics throughout the entire running cycle;

[0115] For the aligned image sequence, the average optical flow amplitude characteristics are calculated using the optical flow method;

[0116] At each synchronization point, all the features extracted above are concatenated into a feature vector, and the feature vectors of all time points are arranged in chronological order to form the multidimensional time series.

[0117] In this embodiment of the invention, by systematically aligning, purifying and reorganizing multi-source heterogeneous sensor data over time, a standardized input that can be directly processed by subsequent machine learning models is constructed. This is a key data preprocessing step for realizing intelligent diagnosis.

[0118] Timestamp alignment solves the problem of asynchronous data acquisition caused by independent acquisition of multiple sensors. By assigning a unified timestamp to all data and using an interpolation algorithm to resample vibration and current data to the time point of the image frame rate, the originally asynchronous physical quantities can maintain strict correspondence when describing the same instant of operation, providing a reliable time reference for the joint analysis of multimodal signals.

[0119] The feature extraction process condenses information in a targeted manner based on the physical meaning of various types of data. For vibration acceleration signals, the root mean square value in the time domain is calculated to reflect the average vibration energy, the peak value is extracted to capture the maximum impact intensity, and the spectral centroid is calculated by Fourier transform to characterize the main frequency distribution of vibration energy. For drive current signals, the root mean square value is also extracted to represent the average load, and the peak value is used to reflect the maximum instantaneous load. For image sequences, the motion vector of pixels between consecutive frames is calculated by optical flow algorithm, and the average optical flow amplitude of the entire door area is obtained, thereby quantifying visual information into a time-series index characterizing the overall movement speed of the door.

[0120] Finally, at each aligned time point, all extracted scalar features are concatenated into a comprehensive feature vector. The feature vectors of all time points are arranged in chronological order to form a well-structured multidimensional time series. This series integrates mechanical vibration, electrical load, and visual motion information, while maintaining complete temporal dynamic characteristics. It provides high-information-density, dimensionally unified, and temporally distinct input data for subsequent intelligent analysis based on time series models.

[0121] It should be noted that the start-up time and sampling frequency (e.g., 1000 vibrations per second, 30 frames per second for images) of the current sensor and camera are usually different and independent. If they are not aligned, the vibration, current and image features that are supposed to be "at the same moment" may actually come from different physical moments that are tens or even hundreds of milliseconds apart. During the rapid start and stop of the elevator door operator, this will lead to completely disordered analysis. For example, any conclusions drawn by matching the vibration data of the "start-up phase" with the image of the "uniform speed phase" will be invalid.

[0122] Furthermore, features are extracted from the aligned strain time-series signal, including:

[0123] Average strain value: Calculates the average strain signal over the entire operating cycle. This characteristic represents the average static stress level experienced by the monitoring point during that cycle. A long-term upward trend in this value may indicate a loss of installation preload or persistent abnormal load.

[0124] Peak strain: Extracts the maximum value of the strain signal. This feature reflects the maximum instantaneous stress within the cycle and can be used to detect stress concentration caused by abnormal impact loads or jamming.

[0125] Strain change (gradient): Calculates the rate of strain change from the start of the cycle to the stable strain value, or from the stable value to the end of the cycle. This feature quantifies the rate of stress loading and unloading. An abnormal rate of change may indicate impeded motion or changes in damping characteristics.

[0126] Unloading residual strain (optional important feature): Calculates the residual strain value relative to the baseline at the end of the cycle. Non-zero residual strain is direct evidence of plastic deformation or loosening of connections in the structure and is a strong indicator for early warning of serious structural failures.

[0127] The final constructed multidimensional time series integrates four dimensions of information: mechanical vibration (dynamic response), driving current (electrical load), visual motion (external manifestation), and structural strain (internal stress). The introduction of the strain dimension enables the series to characterize the health status of the gantry crane system's "skeleton," providing an irreplaceable data foundation for subsequent models to distinguish between "abnormalities on the operating surface" and "damage to the structural body."

[0128] As an optional embodiment, the step described above, of inputting the multidimensional time series into an unsupervised autoencoder based on a gated cyclic unit and calculating its reconstruction error, includes:

[0129] The encoder of the autoencoder encodes the input multidimensional time series into a context vector;

[0130] The decoder of the autoencoder reconstructs an output sequence based on the context vector;

[0131] The mean square error of each data point between the original input sequence and the reconstructed output sequence is calculated as the reconstruction error.

[0132] In this embodiment of the invention, an anomaly perception mechanism that relies solely on historical data of normal device operation for learning is realized by constructing an unsupervised autoencoder model based on a gated recurrent unit. As a type of neural network, the core function of the autoencoder is to learn an efficient representation of the input data and attempt to reconstruct the original input from the representation.

[0133] In this model, the encoder consists of a gated recurrent unit layer. A gated recurrent unit is a special type of recurrent neural network that can effectively process time-series data and capture its long-term dependencies. The encoder receives a multi-dimensional time series as input, which contains fused vibration, current, and image motion features. The encoder processes the sequence step by step and finally generates a fixed-dimensional context vector. This context vector is a condensation and summary of the information contained in the entire input sequence. It encodes the potential features and patterns that the time-series data should have under the normal operating mode learned by the model.

[0134] The decoder section typically consists of an inverse gated recurrent unit layer or a similar structure. It takes the context vector generated by the encoder as the initial state or input and attempts to gradually reconstruct the original multidimensional time series. Ideally, for a normal sequence that conforms to the learned pattern, the decoder should be able to output a reconstructed sequence that is highly similar to the original input.

[0135] The calculation of reconstruction error is crucial to the model output, and this error is quantified using the mean squared error function. Specifically, for the original input sequence and the output sequence reconstructed by the decoder, at each corresponding time step, the values ​​of each feature dimension are compared, the difference between these two values ​​is calculated, and the square of this difference is taken. Then, the squared differences of all feature dimensions at that time step are summed. Next, the summation results from all time steps are accumulated again, and finally, the result is divided by the total number of time steps to obtain a scalar average. This calculated mean squared error value is the reconstruction error.

[0136] During the model training phase, the network's internal parameters are continuously adjusted by inputting a large number of multi-dimensional time series collected and constructed during normal device operation. The goal is to minimize the reconstruction error, enabling the model to learn how to accurately reconstruct the normal operating mode. In the application phase, when a new operating cycle sequence is input, the trained model will attempt to reconstruct it. If the sequence is consistent with the learned normal mode, the reconstruction error will remain at a low level. Conversely, if the sequence contains abnormal modes that the model has not learned, the model will have difficulty reconstructing accurately, leading to an increase in the reconstruction error. Therefore, this reconstruction error value itself becomes a direct indicator that quantitatively reflects the degree to which the current operating state deviates from the historical normal baseline, providing a core basis for subsequent anomaly judgment.

[0137] As an optional embodiment, in the above steps, the adaptive dynamic threshold is dynamically calculated and determined based on the number of historical operating cycles of the elevator door operator and its corresponding historical reconstruction error sequence. The specific determination method includes:

[0138] Obtain the reconstruction error value of the elevator door operator under N consecutive normal cycles;

[0139] Using the cycle number as the independent variable and the reconstruction error value as the dependent variable, a linear baseline is fitted to characterize the natural trend of its variation with the number of cycles.

[0140] Calculate the standard deviation of the residuals of each historical error value relative to the linear baseline. ;

[0141] For the Each cycle has a corresponding adaptive dynamic threshold. The calculation formula is: ,in For the linear baseline at the 1st The predicted value of the cycle, These are preset coefficients.

[0142] In this embodiment of the invention, the adaptive dynamic threshold calculation method is the core of achieving low false alarms and accurate early warning. Its essence is to establish a set of personalized early warning standards for each device that evolve in sync with its own performance degradation process.

[0143] The implementation of this invention begins with data collection. The system first needs to obtain the reconstruction error values ​​corresponding to N consecutive cycles of the elevator door machine that are confirmed to be operating normally within a certain historical period. These error values ​​are calculated by the aforementioned unsupervised autoencoder and constitute a historical error sequence. This sequence is the basis for subsequent calculations and reflects the normal fluctuation level of the equipment in a fault-free state and its possible long-term change trend.

[0144] The next step is baseline fitting. Since the performance of mechanical equipment naturally and slowly degrades with increasing usage, its signal characteristics under normal operating conditions (and the reconstructed error calculated accordingly) may also exhibit a slow changing trend. To quantify this trend, this invention uses the operating cycle number as the independent variable and the corresponding reconstructed error value as the dependent variable, employing the least squares method for linear regression to fit a straight line. Its mathematical expression is typically in the form: Error = Slope × Cycle Number + Intercept. This straight line is called the linear baseline. It describes the natural trend or expected path of the reconstructed error value as the number of operating cycles increases, assuming no sudden failures. The slope characterizes the average rate of change of the error, and the intercept represents the initial value.

[0145] After obtaining the baseline, it is necessary to assess the normal fluctuation range of historical data around that baseline. To do this, the difference between the actual reconstruction error value for each historical period and the predicted value of the linear baseline at that period needs to be calculated; this difference is called the residual. Then, the standard deviation of all these residuals is calculated, denoted as . Standard deviation It quantifies the typical dispersion of historical normal error values ​​from their trend baseline and is a statistical measure of the inherent random fluctuations in normal equipment operation.

[0146] Ultimately, the adaptive dynamic threshold is determined by the following formula: for the future... Each running cycle, and its corresponding dynamic threshold Equal to the baseline forecast value for that period Add a constant with standard deviation The product of, i.e. .in, It is the period sequence number The predicted values ​​obtained by substituting them into the aforementioned linear baseline equation, the coefficients It is a preset, non-negative constant factor used to control the sensitivity of the warning; the larger the value of k, the wider the threshold range and the more conservative the warning; the smaller the value of k, the narrower the threshold range and the more sensitive the warning.

[0147] As can be seen, through this series of calculations, the threshold for anomaly detection is no longer a fixed number, but a dynamic value that rises in sync with the equipment's performance degradation baseline as its operating cycle increases. This allows the early warning system to automatically adapt to the aging process of the equipment: for new equipment, the baseline value is lower, the threshold range is narrower, and the standard is stricter; for older equipment, the baseline value rises due to natural degradation, and the threshold range also rises accordingly, with the standard being reasonably relaxed. This method effectively distinguishes between normal performance degradation and sudden anomalies, thus maintaining the sensitivity to detect real faults without generating excessive false alarms.

[0148] As an optional embodiment, in the above steps, the key feature identifier is one or more feature dimension identifiers determined based on the internal decision rules of the gradient boosting decision tree model, which are determined through the following steps:

[0149] After the gradient boosting decision tree model classifies the input abnormal multidimensional time series, it traverses all decision trees in the model and counts the total amount of impurity reduction brought about by each feature dimension when splitting at the nodes of all decision trees, which is used as the global importance score of that feature dimension.

[0150] All feature dimensions are sorted in descending order of global importance score, and the top R feature dimensions are selected, where R is a preset positive integer.

[0151] The identifiers of the selected top R feature dimensions are output as the key feature identifiers.

[0152] In this embodiment of the invention, by designing a feature importance quantification and selection mechanism based on the internal decision rules of the model for the fault diagnosis module, the key transformation of the diagnosis results from "black box" judgment to "white box" interpretation is realized, thereby improving the direct guiding value of the output results for maintenance work.

[0153] The core of this mechanism lies in using the structural characteristics of the gradient boosting decision tree model itself for post-hoc attribution analysis. The gradient boosting decision tree is composed of multiple decision trees, and each tree makes classification decisions through a series of node splitting rules. Whenever the model splits at a node of the tree using a specific dimension (i.e., a certain feature) of the input feature sequence, this split will reduce the classification impurity. The amount of impurity reduction can be intuitively understood as how much the model improves the clarity of the fault category based on the judgment of this feature at this point.

[0154] In the specific calculation process, after the model completes the classification prediction of an abnormal sequence, the system will traverse all decision trees in the ensemble and perform contribution statistics for each input feature dimension. For a certain feature dimension, the system accumulates the sum of all impurity reductions caused by it when all nodes of all trees split. This accumulated sum is defined as the global importance score of the feature dimension. This score is a quantitative indicator. The higher the value, the greater the contribution of the feature dimension to correctly distinguishing different fault categories in the collective decision-making process of the entire model. In other words, it is likely to be the key signal on which this judgment is based.

[0155] Subsequently, the system sorts all input feature dimensions in descending order according to the calculated global importance score. After sorting, it selects the top three feature dimensions based on a preset number R, for example, R=3. Finally, the system outputs the identifiers of these selected feature dimensions as the key feature identifiers used for this diagnosis.

[0156] As can be seen, through this process, when the model outputs a fault type prediction, such as "poor guide rail lubrication," the system can simultaneously output several key features on which this judgment is based, such as "high center of gravity of Y-axis vibration spectrum during braking" and "insufficient peak starting current." This allows maintenance personnel not only to know "what the fault is," but also to understand "why the model makes this judgment." Consequently, they can conduct targeted on-site verification and validation of these specific abnormal feature indicators, greatly improving the credibility of diagnostic conclusions and the accuracy of maintenance actions.

[0157] As an optional embodiment, the above steps, after constructing the multidimensional time series, further include a runtime phase division step:

[0158] Based on the waveform characteristics of the driving current timing signal, the start-up, constant speed, and braking phases of a single operating cycle are automatically identified and divided.

[0159] The data points in the multidimensional time series are associated with their respective operational phases.

[0160] In this embodiment of the invention, by introducing an automatic segmentation step based on the characteristics of the driving current waveform, the periodic working process of the elevator door operator is finely deconstructed and labeled, thereby enhancing the contextual awareness and directionality of subsequent state analysis. The driving current timing signal directly represents the operating state of the motor. The starting current rising segment, the uniform speed running smooth segment, and the braking current falling segment that appear sequentially in its waveform have clear physical meanings and easily identifiable morphological features. By automatically detecting these feature inflection points through the algorithm, the continuous door operator operating cycle can be divided into three physical stages: starting, uniform speed, and braking. The entire process does not require manual intervention.

[0161] Based on this, each data point in the aforementioned multidimensional time series is associated with its specific operational stage, which is equivalent to labeling all vibration, current, and image features with a temporal context. This allows subsequent anomaly detection and fault diagnosis analysis to be built on a stage-specific, refined model. The system can learn the independent normal behavior patterns of the equipment at different stages, thereby improving the sensitivity to stage-specific anomalies. When an anomaly is detected, the system can clearly indicate which operational stage the anomaly mainly occurs in, such as abnormal vibration in the startup stage or abnormal current in the braking stage. This stage-related information provides maintenance personnel with a key spatiotemporal dimension reference for locating the root cause of the fault, making the diagnostic conclusion not only "what kind of fault" but also indicating "which work process the fault occurred in," significantly improving the operability and guiding value of the diagnostic results.

[0162] As an optional embodiment, in the above steps, the key feature identifier is associated with information about the specific operational stage to which it belongs.

[0163] In this embodiment of the invention, by binding and associating the identified key feature identifiers with their specific operational stages (start-up, constant speed, braking) in the original data, the final mapping of diagnostic information from the abstract feature space to the specific physical operation context is completed, which is a key step in improving the operability of diagnostic results.

[0164] Once the system completes fault diagnosis and outputs key feature identifiers (such as "Y-axis vibration spectrum centroid"), the feature identifier itself only indicates "which signal indicator is abnormal". By associating it with the specific operating stage information, the diagnostic report is further enriched to "which signal indicator is abnormal in which operating stage". For example, the output format is upgraded from "key feature: Y-axis vibration spectrum centroid" to "key feature: Y-axis vibration spectrum centroid during braking stage".

[0165] The addition of this related information provides on-site maintenance personnel with crucial and directly actionable context. The dynamic characteristics and stress states of elevator door operators vary significantly at different operating stages. Anomalies during the start-up stage may indicate insufficient starting torque or excessive initial resistance; anomalies during the constant speed stage may reflect issues with guide rail flatness or continuous friction; and anomalies during the braking stage are directly related to brake performance or collisions at the closing end. Knowing the specific stage at which the anomaly occurs allows maintenance personnel to instantly narrow down the troubleshooting scope to the specific mechanical components and control logic corresponding to that stage. This greatly improves the efficiency and accuracy of fault diagnosis, avoids blind, large-scale repairs, and ultimately forms a closed-loop, actionable maintenance decision support information system.

[0166] Furthermore, we will use a specific simulated scenario as an example to demonstrate its role in actual elevator operation and maintenance.

[0167] Scenario: A split-type elevator door operator in a high-rise hotel, which has been running for more than 80,000 times, has recently experienced a slow increase in the reconstruction error of its monitoring system, occasionally exceeding the adaptive threshold. However, no obvious abnormal patterns have been observed in the vibration and current characteristics, causing maintenance personnel to worry about false alarms or natural aging.

[0168] 1. Four-dimensional data acquisition and feature extraction:

[0169] During one operating cycle, the system simultaneously collected four types of data. The newly added fiber optic grating sensor, installed on the support seat of the right door leaf guide rail, extracted an "average strain value" of +152. (Compared to baseline one month ago +50) The "Residual Strain After Unloading" value is +12. (Historical baseline close to 0) Meanwhile, the vibration and current characteristics fluctuated within their respective dynamic thresholds, and no obvious obstruction was found in the optical flow of the image.

[0170] 2. Unsupervised anomaly detection

[0171] After inputting the four-dimensional time series into the unsupervised autoencoder, the calculated reconstruction error is 0.78, which is slightly higher than the current period adaptive threshold of 0.75. The system triggers a "minor anomaly" alarm. The main difficulty in reconstructing the model is the pattern of the strain features.

[0172] 3. Supervised fault diagnosis and attribution

[0173] Anomaly sequences are input into a gradient boosting decision tree model. The model outputs the following after comprehensive four-dimensional feature analysis:

[0174] Fault type prediction: Loose guide rail mounting bracket.

[0175] Predicted probability: 76%.

[0176] Key feature identifiers:

[0177] 1. (Corresponding to the average strain value of the right guide rail bracket)

[0178] 2. (Residual strain corresponding to unloading of the right guide rail bracket)

[0179] 3. (Corresponding to the peak starting current, its contribution is far lower than the former two)

[0180] Phase correlation: Key strain characteristics persist throughout the entire operating cycle and are not completely unloaded during the "braking phase".

[0181] 4. Interpretable diagnostic reports and results:

[0182] The system pushed out a diagnostic report, which clearly pointed out: "The root cause of the abnormality points to the mechanical structure, rather than the drive or moving parts. The continuously increasing static stress and clear residual deformation were detected at the right door guide rail bracket, indicating that the bracket connection is loose, which leads to a change in the stress state of the guide rail, and may cause the door to deviate or make abnormal noise in the long term."

[0183] On-site verification: Based on the precise positioning, the maintenance personnel inspected the right guide rail bracket and found that two fixing bolts were loose. After tightening, subsequent monitoring data showed that the average strain value and residual strain returned to the normal baseline. In this case, traditional vibration and current monitoring failed to provide a clear alarm, while the newly added strain monitoring directly captured early, static evidence of structural loosening, preventing the fault from escalating.

[0184] Example 2

[0185] Please see Figure 2 , Figure 2 This is a schematic diagram of the structure of a multi-dimensional anomaly detection and fault prediction device for an elevator door operator system disclosed in an embodiment of the present invention. Figure 2 The multi-dimensional anomaly detection and fault prediction device for an elevator door operator system described herein can be applied to a data processing chip, processing terminal, or processing server, and the processing server can be a local server or a cloud server; this embodiment of the invention does not limit the application. Figure 2 As shown, the multi-dimensional anomaly detection and fault prediction device for the elevator door operator system may include the following operations:

[0186] The synchronous acquisition module 201 is used to synchronously acquire vibration sensing data, current sensing data, image sensing data, and structural strain sensing data within a single operating cycle of the elevator door operator.

[0187] Specifically, by simultaneously collecting three types of sensor data with different physical properties—vibration, current, and images—a multi-dimensional and complementary information foundation is built for the status monitoring of elevator door machine systems. Vibration data directly reflects the impact and wear of the mechanical structure, current data characterizes the electrical load and efficiency of the motor drive, and image data provides information on the spatial pose and visual appearance of the door during its movement.

[0188] It is evident that this multi-source synchronous acquisition method enables subsequent analysis to make comprehensive judgments based on both internal signal characteristics and external visual manifestations, providing complete data input for detecting mechanical faults, electrical anomalies, and visible physical obstructions or deformations, thereby supporting more comprehensive fault diagnosis capabilities at the information source level.

[0189] The time series construction module 202 is used to perform timestamp alignment and feature extraction on the collected multi-source sensor data, and arrange the extracted numerical features in chronological order to construct a unified multi-dimensional time series.

[0190] Specifically, by aligning the timestamps of multi-source asynchronous sensing data, the problem of data lag and misalignment caused by independent acquisition by different sensors is solved, ensuring that vibration, current and image data have a consistent time reference when characterizing the same running event. The subsequent feature extraction process transforms the synchronized high-dimensional original waveform and image frame sequence into a set of low-dimensional numerical feature indicators with clear physical or statistical significance, thus completing the transformation from the original signal to the state representation quantity.

[0191] As can be seen, the data processing flow outputs a standardized sequence that is structurally regular, multi-dimensionally integrated, and time-aligned. This sequence serves as a unified input format for subsequent unsupervised and supervised learning models, enabling the algorithm to jointly model and analyze device states based on a consistent timeline and comparable feature scales. This provides a reliable data foundation for effectively detecting and diagnosing anomalies from heterogeneous signals.

[0192] The unsupervised anomaly detection module 203 is used to input the multidimensional time series into an unsupervised autoencoder based on a gated cyclic unit and calculate its reconstruction error.

[0193] Specifically, by using an unsupervised autoencoder based on gated cyclic units to model multidimensional time series, the model can learn the temporal characteristics and change patterns of the equipment under normal operating conditions. This method does not rely on labeled fault data, but captures the inherent patterns of health data through the encoding and reconstruction process. The reconstruction error output by the model directly quantifies the degree of deviation between the current data and the learned normal pattern.

[0194] It is evident that this design implements an anomaly perception mechanism that does not rely on prior knowledge of faults. By calculating and evaluating reconstruction errors, the system can sensitively indicate anomalies that deviate statistically from historical normal operating conditions, including unseen or early fault modes, thereby providing effective trigger signals and judgment criteria for subsequent fault diagnosis.

[0195] The supervised fault diagnosis module 204 is built into an integrated learning classifier. In response to the abnormal signal triggered by the unsupervised anomaly detection module, it receives the corresponding multidimensional time series and outputs the fault type prediction, prediction probability and key feature identifier.

[0196] Specifically, anomaly detection is achieved by comparing the reconstruction error with an adaptive threshold dynamically calculated based on the device's historical operating data. The calculation of this threshold incorporates the number of historical operating cycles of the device as a key parameter, enabling it to adjust synchronously with the natural aging trajectory of the device's performance, rather than using a fixed value.

[0197] It is evident that this adaptive threshold design based on the equipment lifecycle enables the anomaly judgment criteria to inherently adapt to the performance degradation process of a single device. This improves the accuracy and practicality of the early warning system when distinguishing between normal aging of equipment and sudden real failures, and reduces misjudgments caused by unreasonable threshold settings.

[0198] The dynamic threshold calculation module 205 is used to dynamically calculate and update the adaptive dynamic threshold used by the unsupervised anomaly detection module based on the number of historical operating cycles of the elevator door operator and its corresponding historical reconstruction error sequence.

[0199] Specifically, by inputting the multidimensional time series that triggered the anomaly into the ensemble learning classifier after an anomaly is detected, a decision cascade from anomaly perception to specific fault diagnosis is realized. The classifier can output specific fault type hypotheses and their corresponding probabilities of occurrence based on the input feature sequences, and simultaneously analyze and derive the key feature identifiers that have the greatest impact on this classification decision.

[0200] It is evident that this cascaded diagnostic mechanism, while completing fault classification, provides interpretable output of the judgment basis. The system not only tells "what the possible fault is", but also indicates "what abnormal features were mainly used to make the judgment", which transforms the diagnostic conclusion from a single code label into information containing reasoning basis, thereby enhancing the direct guiding value of the results for on-site maintenance work.

[0201] Example 3

[0202] Please see Figure 3 , Figure 3 This is a schematic diagram of the structure of another multi-dimensional anomaly detection and fault prediction device for an elevator door operator system disclosed in an embodiment of the present invention. Figure 3 As shown, the device may include:

[0203] Memory 301 storing executable program code;

[0204] Processor 302 coupled to memory 301;

[0205] The processor 302 calls the executable program code stored in the memory 301 to execute some or all of the steps in the abnormal detection and fault prediction method of the elevator door operator system disclosed in Embodiment 1 of the present invention.

[0206] Example 4

[0207] This invention discloses a computer storage medium storing computer instructions. When these computer instructions are invoked, they are used to execute some or all of the steps in the abnormal detection and fault prediction method for an elevator door operator system disclosed in Embodiment 1 of this invention.

[0208] Example 5

[0209] This invention discloses a computer program product, which includes a non-transitory computer-readable storage medium storing a computer program, and the computer program is operable to cause a computer to perform the steps of an abnormality detection and fault prediction method for an elevator door operator system as described in Embodiment 1.

[0210] The system embodiments described above are merely illustrative. The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical modules; that is, they may be located in one place or distributed across multiple network modules. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0211] Through the detailed description of the above embodiments, those skilled in the art can clearly understand that each implementation method can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, including read-only memory (ROM), random access memory (RAM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), one-time programmable read-only memory (OTPROM), electrically-erasable programmable read-only memory (EEPROM), compact disc read-only memory (CD-ROM) or other optical disc storage, disk storage, magnetic tape storage, or any other computer-readable medium that can be used to carry or store data.

[0212] Finally, it should be noted that the method and apparatus for anomaly detection and fault prediction of an elevator door operator system disclosed in the embodiments of the present invention are merely preferred embodiments of the present invention and are only used to illustrate the technical solutions of the present invention, not to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for anomaly detection and fault prediction in an elevator door operator system, characterized in that, The method includes: Simultaneously collect vibration sensing data, current sensing data, image sensing data, structural strain sensing data, and suspension rope early warning data within a single operating cycle of the elevator door operator; The collected multi-source sensor data is time-stamp aligned and features are extracted. The extracted numerical features are arranged in chronological order to construct a unified multi-dimensional time series. The multidimensional time series is input into an unsupervised autoencoder based on a gated cyclic unit, and its reconstruction error is calculated. The reconstruction error is compared with an adaptive dynamic threshold, which is dynamically calculated and determined based on the number of historical operating cycles of the elevator door operator and its corresponding historical reconstruction error sequence; if the reconstruction error exceeds the threshold, it is determined to be abnormal and a diagnosis is triggered. In response to the aforementioned anomaly trigger, the corresponding multidimensional time series is input into an ensemble learning classifier, which outputs a fault type prediction, a prediction probability, and the key feature identifier that contributes most to the current classification decision.

2. The method for anomaly detection and fault prediction of an elevator door operator system according to claim 1, characterized in that, The synchronous acquisition of vibration sensing data, current sensing data, and image sensing data within a single operating cycle of the elevator door operator includes: The vibration acceleration time-series signal of at least one axis is collected by a vibration acceleration sensor installed on the gantry crane; The timing signal of the motor drive current is collected by a current sensor connected in series in the door operator motor circuit; Image sensors installed on the side of the door operator are used to collect image sequences covering the entire door opening and closing process; By using fiber optic strain sensors installed on the key load-bearing structure of the gantry crane, the static or quasi-static strain time-series signal of the structure is collected within a single operating cycle. It also includes acquiring the warning status data of the suspension rope warning device, which is embedded in the elevator door and used to provide early warnings for elevator safety.

3. The method for anomaly detection and fault prediction of an elevator door operator system according to claim 1, characterized in that, The process involves aligning timestamps and extracting features from the collected multi-source sensor data, arranging the extracted numerical features in chronological order, and constructing a unified multidimensional time series, including: The vibration acceleration timing signal, the driving current timing signal, and the image sequence are aligned to a unified time axis using timestamp synchronization and interpolation methods. For the aligned vibration acceleration signal, extract its root mean square value in the time domain, peak value, and spectral centroid features in the frequency domain; For the aligned drive current signal, extract its root mean square value and peak value in the time domain; For the aligned strain time series signal, extract its average strain value, strain peak value and strain change gradient characteristics throughout the entire running cycle; For the aligned image sequence, the average optical flow amplitude characteristics are calculated using the optical flow method; At each synchronization point, all the features extracted above are concatenated into a feature vector, and the feature vectors of all time points are arranged in chronological order to form the multidimensional time series.

4. The method for anomaly detection and fault prediction of an elevator door operator system according to claim 1, characterized in that, The step of inputting the multidimensional time series into an unsupervised autoencoder based on a gated cyclic unit and calculating its reconstruction error includes: The encoder of the autoencoder encodes the input multidimensional time series into a context vector; The decoder of the autoencoder reconstructs an output sequence based on the context vector; The mean square error of each data point between the original input sequence and the reconstructed output sequence is calculated as the reconstruction error.

5. The method for anomaly detection and fault prediction of an elevator door operator system according to claim 1, characterized in that, The adaptive dynamic threshold is dynamically calculated and determined based on the number of historical operating cycles of the elevator door operator and its corresponding historical reconstruction error sequence. The specific determination method includes: Obtain the reconstruction error value of the elevator door operator under N consecutive normal cycles; Using the cycle number as the independent variable and the reconstruction error value as the dependent variable, a linear baseline is fitted to characterize the natural trend of its variation with the number of cycles. Calculate the standard deviation of the residuals of each historical error value relative to the linear baseline. ; For the Each cycle has a corresponding adaptive dynamic threshold. The calculation formula is: ,in For the linear baseline at the 1st The predicted value of the cycle, These are preset coefficients.

6. The method for anomaly detection and fault prediction of an elevator door operator system according to claim 1, characterized in that, The key feature identifier is one or more feature dimension identifiers determined based on the internal decision rules of the gradient boosting decision tree model, which are determined through the following steps: After the gradient boosting decision tree model classifies the input abnormal multidimensional time series, it traverses all decision trees in the model and counts the total amount of impurity reduction brought about by each feature dimension when splitting at the nodes of all decision trees, which is used as the global importance score of that feature dimension. All feature dimensions are sorted in descending order of global importance score, and the top R feature dimensions are selected, where R is a preset positive integer. The identifiers of the selected top R feature dimensions are output as the key feature identifiers.

7. The method for anomaly detection and fault prediction of an elevator door operator system according to claim 1, characterized in that, After constructing the multidimensional time series, the process also includes a phase division step: Based on the waveform characteristics of the driving current timing signal, the start-up, constant speed, and braking phases of a single operating cycle are automatically identified and divided. The data points in the multidimensional time series are associated with their respective operational phases.

8. The method for anomaly detection and fault prediction of an elevator door operator system according to claim 7, characterized in that, The key feature identifier is associated with information about the specific operational stage to which it belongs.

9. A multi-dimensional anomaly detection and fault prediction device for an elevator door operator system, used to implement the anomaly detection and fault prediction method for an elevator door operator system as described in any one of claims 1-8, characterized in that, The system includes: The synchronous acquisition module is used to synchronously acquire vibration sensing data, current sensing data, image sensing data, structural strain sensing data, and suspension rope early warning data within a single operating cycle of the elevator door operator. The time series construction module is used to perform timestamp alignment and feature extraction on the multi-source sensor data collected by the synchronous acquisition module, and arrange the extracted numerical features in chronological order to construct a unified multi-dimensional time series. An unsupervised anomaly detection module, which is built into an unsupervised autoencoder based on a gated cyclic unit, is used to receive the multidimensional time series output by the time series construction module, calculate its reconstruction error, and compare the reconstruction error with an adaptive dynamic threshold. If the threshold is exceeded, an anomaly signal is triggered. The supervised fault diagnosis module is built into an integrated learning classifier. In response to the abnormal signal triggered by the unsupervised anomaly detection module, it receives the corresponding multidimensional time series and outputs the fault type prediction, prediction probability and key feature identifier. The dynamic threshold calculation module is used to dynamically calculate and update the adaptive dynamic threshold used by the unsupervised anomaly detection module based on the number of historical operating cycles of the elevator door operator and its corresponding historical reconstruction error sequence.

10. A multi-dimensional anomaly detection and fault prediction device for an elevator door operator system, characterized in that, The system includes: Memory containing executable program code; A processor coupled to the memory; The processor calls the executable program code stored in the memory to execute the abnormal detection and fault prediction method for an elevator door operator system as described in any one of claims 1-8.

Images