An elevator door machine system fault diagnosis method and system based on deep learning

Abstract

The application discloses a kind of based on deep learning's elevator door machine system fault diagnosis method and system, it is related to elevator door machine system fault diagnosis technical field.The based on deep learning's elevator door machine system fault diagnosis method, by collecting the running state time series data of each operating stage, input pre-training model extraction feature, construct running health time series index, and in combination with preset health interval judge whether there is fault, realize multi-stage, intelligent door machine fault identification, the application obtains the running state time series data of elevator door machine in each operating stage, in combination with the time series modeling capability of deep learning model, realize the dynamic modeling of door machine system in whole cycle operating state, by multi-head attention mechanism highlight abnormal fragment and extract key operating characteristics, output the stage health index of each operating stage, by comparing the running health time series index of each stage with preset health interval, can realize more fine granularity stage-by-stage fault identification.

Description

Technical Field

[0001] This invention relates to the field of elevator door operator system fault diagnosis technology, specifically to an elevator door operator system fault diagnosis method and system based on deep learning. Background Technology

[0002] As a critical component for elevator operational safety, the elevator door operator system is responsible for opening and closing the car doors and landing doors. Its working condition directly affects passenger safety and equipment operational stability. Most existing elevator door operators use a motor-driven structure, adjusting opening and closing speed, position, and response rhythm through a control module. In actual operation, these systems are susceptible to factors such as mechanical wear, track deformation, electrical fluctuations, and control delays, leading to malfunctions such as door jamming, misalignment, and accidental opening or closing.

[0003] Existing technology, such as the invention patent application with publication number CN116101864B, discloses a method and system for diagnosing elevator door system faults based on sound recognition technology. The method includes the following steps: S1, when the elevator car door opens and closes, sound data of the elevator door system, acceleration data of the car door, and vibration data of the elevator door operator are collected, and the collected sound data, acceleration data, and vibration data are sent to an industrial control computer; S2, the industrial control computer runs a sound algorithm model and periodically analyzes the sound data of the elevator door system. This invention collects sound, acceleration, and vibration data of the elevator door system through multiple sensors and uses sound recognition and fault judgment algorithms to analyze and determine whether there are any abnormalities in the elevator door system, thereby achieving rapid and accurate detection and diagnosis of the health status of the elevator door system. Furthermore, the health status is displayed through a cloud platform, enabling remote monitoring of the elevator door system. If an abnormality occurs in the elevator door system, an automatic alarm is triggered.

[0004] Based on the above findings, the limitations of existing technologies include at least the following problems: Existing methods for fault diagnosis of elevator door systems based on sound recognition have significant structural limitations. This stems from the fact that their fault judgment process highly relies on the abrupt changes in acoustic signal characteristics and the pattern classification of local features at a single moment. First, these methods typically rely on single-channel sound sampling, constructing feature vectors by analyzing the amplitude, spectral characteristics, or kurtosis of the sound signal at a specific time point to determine whether an anomaly has occurred. However, elevator door operators often experience multiple acoustic interferences during actual operation, such as background noise, superimposed control actions, and structural reflection echoes. Furthermore, the sound characteristics of different door operator models vary significantly under normal conditions, making these methods highly sensitive to the on-site environment and equipment type, resulting in poor stability and weak generalization ability. Second, sound recognition methods only rely on short intervals during the door opening and closing process. Transient signals are difficult to systematically capture the state evolution of the gantry crane during multiple stages of operation, such as start-up, acceleration, constant speed, deceleration, and stopping. Judgment is often based on the presence of abnormal sounds, lacking the ability to continuously monitor and infer physical variables such as motor current and voltage changes, control response delays, gate friction trends, and structural offsets. This makes it difficult to identify and warn of structural faults, control chain delay drift, or progressive wear in the early stages. Furthermore, the threshold judgment and classification rules used in existing methods are usually based on empirical settings or specific sample training, failing to build a unified state health expression model. It is also difficult to infer the overall operating status of the gantry crane system through the time series correlation of cross-stage features. As a result, the diagnostic scope is limited to the identification of sudden anomalies, and there is a lack of effective capture capabilities for non-obvious faults such as system performance degradation and behavioral deviations over long periods. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a method and system for diagnosing elevator door operator faults based on deep learning. This solves the problems of existing technologies that rely solely on local sound mutations for static classification and lack the ability to fuse and model multiple physical features throughout the entire door operator operation process and judge temporal health trends.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a deep learning-based method for diagnosing elevator door operator system faults, comprising the following steps: continuously acquiring the time-series data of the operating status of each operating stage of the elevator door operator; inputting the time-series data of the operating status of each operating stage of the elevator door operator into a pre-trained deep learning model for feature extraction processing to obtain the operating sequence features of each operating stage of the elevator door operator; analyzing the operating health time-series index of each operating stage of the elevator door operator based on the operating sequence features of each operating stage of the elevator door operator; comparing the operating health time-series index of each operating stage of the elevator door operator with the corresponding preset stage operating health interval, and considering the operating stage whose operating health time-series index is outside the preset stage operating health interval as an operating fault.

[0007] Furthermore, the operating status timing data includes motor operating current, motor operating voltage, motor operating power, door operator electromagnetic interference intensity, control response delay, door friction load, and door offset angle at several time points.

[0008] Furthermore, the deep learning model includes an input embedding layer, a bidirectional GRU layer, a multi-head attention layer, a stage feature aggregation layer, and a feature output layer. The runtime sequence features include the peak value of the motor operating current, the variance of the motor operating current fluctuation, the number of high-frequency disturbances in the motor operating current, the stability coefficient of the motor operating voltage, the average intensity of the electromagnetic interference of the door operator, the duration of the electromagnetic interference of the door operator, the average delay of the control response, the maximum delay of the control response, the fluctuation rate of the control response delay, the static friction load of the door body, the dynamic friction load of the door body, the slope of the door body friction trend, the abrupt change rate of the door body offset, the average angle of the door body offset, and the door body jitter frequency.

[0009] Further, the specific steps for obtaining the runtime sequence features of each operating stage of the elevator door operator are as follows: In the input embedding layer of the deep learning model, the runtime state temporal data of each operating stage of the elevator door operator is received, and normalization and feature dimension mapping are performed to obtain a unified dimension runtime sequence vector of each operating stage of the elevator door operator; In the bidirectional GRU layer of the deep learning model, temporal modeling processing is performed on the runtime sequence vector sequence of each operating stage of the elevator door operator to capture the forward and backward dependencies of the operating states, extract the context state encoding at each time point, and obtain the temporal hidden state vector set of each operating stage of the elevator door operator; In the multi-head attention layer of the deep learning model, the electric... The temporal hidden state vector set of each operating stage of the elevator door operator is subjected to parameter importance evaluation and time slice anomaly segment focusing processing. The cross feature representation of high-weight operating segments is extracted to obtain the stage operating state feature vector set of each operating stage of the elevator door operator with attention weighting mechanism. In the stage feature aggregation layer of the deep learning model, the stage operating state feature vector set of each operating stage of the elevator door operator after attention weighting is subjected to global average aggregation processing to form the stage comprehensive representation vector corresponding to the operating stage. In the feature output layer of the deep learning model, the stage comprehensive representation vector of each operating stage of the elevator door operator is mapped to output the operating sequence feature set of each operating stage of the elevator door operator.

[0010] Furthermore, the specific steps for analyzing the operational health time-series index of each operating stage of the elevator door operator are as follows: Based on the operational sequence characteristics of each operating stage of the elevator door operator, the drive system health index, control response stability index, and door structure attitude index of each operating stage of the elevator door operator are analyzed respectively; the drive system health index, control response stability index, and door structure attitude index of each operating stage of the elevator door operator are comprehensively analyzed to obtain the operational health time-series index of each operating stage of the elevator door operator.

[0011] Furthermore, the specific formula for calculating the operational health time series index of an elevator door operator during a certain operating phase is as follows: ;in, This refers to the operational health time series index of an elevator door operator during a specific operational phase. This refers to the health index of the drive system of the elevator door operator during a certain operating phase. It is a natural constant. This refers to the door structure attitude index during a certain operating phase of the elevator door operator. The structural health adjustment coefficients are stored in the database. This refers to the control response stability index of the elevator door operator during a certain operating phase. These are the control and coordination enhancement coefficients stored in the database.

[0012] Furthermore, the specific steps for analyzing the drive system health index of each operating stage of the elevator door operator are as follows: read the peak value of the motor operating current, the variance of the motor operating current fluctuation, the number of high-frequency disturbances in the motor operating current, and the stability coefficient of the motor operating voltage for each operating stage of the elevator door operator; perform comprehensive analysis on the peak value of the motor operating current, the variance of the motor operating current fluctuation, the number of high-frequency disturbances in the motor operating current, and the stability coefficient of the motor operating voltage for each operating stage of the elevator door operator to obtain the drive system health index of each operating stage of the elevator door operator.

[0013] Furthermore, the specific steps for analyzing the control response stability index of each operating stage of the elevator door operator are as follows: Read the average intensity of electromagnetic interference (EMI), duration of EMI, average control response delay, maximum control response delay, and control response delay fluctuation rate of the elevator door operator for each operating stage; perform comprehensive analysis on the average intensity of EMI, duration of EMI, average control response delay, maximum control response delay, and control response delay fluctuation rate of the elevator door operator for each operating stage to obtain the control response stability index of the elevator door operator for each operating stage.

[0014] Furthermore, the specific steps for analyzing the door structure attitude index of each operating stage of the elevator door operator are as follows: Read the door static friction load, door dynamic friction load, door friction trend slope, door offset abrupt change rate, door offset average angle, and door shaking frequency for each operating stage of the elevator door operator; Perform comprehensive analysis on the door static friction load, door dynamic friction load, door friction trend slope, door offset abrupt change rate, door offset average angle, and door shaking frequency for each operating stage of the elevator door operator to obtain the door structure attitude index for each operating stage of the elevator door operator.

[0015] A deep learning-based elevator door operator fault diagnosis system includes: a data acquisition unit for continuously acquiring time-series data of the operating status of each operating stage of the elevator door operator; a feature extraction unit for inputting the time-series data of the operating status of each operating stage of the elevator door operator into a pre-trained deep learning model for feature extraction processing to obtain the operating sequence features of each operating stage of the elevator door operator; an operating health analysis unit for analyzing the operating health time-series index of each operating stage of the elevator door operator based on the operating sequence features of each operating stage of the elevator door operator; and an operating fault analysis unit for comparing the operating health time-series index of each operating stage of the elevator door operator with a corresponding preset stage operating health interval, and considering operating stages with operating health time-series indices outside the preset stage operating health intervals as operating faults.

[0016] The present invention has the following beneficial effects:

[0017] (1) The elevator door machine system fault diagnosis method based on deep learning continuously acquires the time series data of the elevator door machine's operating status in each operating stage, and combines the time series modeling capability of the deep learning model to realize the dynamic modeling of the door machine system's operating status throughout the entire cycle for the first time. Compared with the existing methods that rely on static sound mutations or rule threshold judgments, this scheme uses a pre-trained deep learning model to perform forward and backward modeling of the time series composed of multi-dimensional physical quantities. Furthermore, it highlights abnormal segments and extracts key operating features through a multi-head attention mechanism, and finally outputs the stage health index of each operating stage. By comparing the operating health time series index of each stage with the preset health interval, it can achieve more granular stage-by-stage fault identification, avoid the diagnostic bias caused by ignoring stage differences in the existing methods, and significantly improve the identification accuracy and response time of slowly changing and non-sudden door machine faults.

[0018] (2) The elevator door operator system fault diagnosis method based on deep learning takes the time series data of the operation status during the operation phase as the time series input, and no longer relies on non-steady-state signals such as sound and vibration that are easily affected by the environment. After the deep learning feature extraction is completed, the scheme constructs the drive system health index, control response stability index, and door structure posture index to modularly separate and model various core influencing factors, and integrates them to form a unified operation health time series index. At the same time, the adjustment coefficient obtained by training in the database improves the adaptability of the evaluation results to different door operator types and usage scenarios, solves the problems of poor model transferability and weak generalization ability of the existing scheme, and has cross-scenario deployment capability.

[0019] (3) The elevator door machine system fault diagnosis method based on deep learning introduces a well-structured deep learning model, including an input embedding layer, a bidirectional GRU layer, a multi-head attention layer, a stage feature convergence layer, and a feature output layer. This model can not only perform high-dimensional vectorization modeling of the normalized multi-dimensional original operating state, but also enhance the ability to identify short-term sudden change segments and chronic risk trends by means of the attention mechanism. Finally, it outputs a semantic fault feature set of the elevator door machine operation stage. On this basis, the structure adjustment coefficient and control enhancement coefficient are jointly trained by the preset objective function and historical health level data, so that the system has the ability to dynamically adjust the weight of the health index under different usage environments. Compared with the existing static rules or experience judgment schemes, it has significant adaptive learning ability and intelligent fault recognition ability. It can realize online fault level assessment and early warning strategy linkage deployment, and improve the elevator operation safety guarantee capability.

[0020] (4) The elevator door operator fault diagnosis system based on deep learning divides the elevator door operator fault diagnosis process into four functional modules: data acquisition unit, feature extraction unit, operation health analysis unit and operation fault analysis unit. It constructs a modular intelligent diagnosis system architecture with clear boundaries and data interfaces. The modules are decoupled and connected through standardized operation stage index and feature data vector. The system has good functional independence and deployment flexibility. The data acquisition unit can connect to multiple brand door operator systems and collect unified standard physical parameter time series data. The feature extraction unit realizes seamless embedded feature extraction based on pre-trained models. The operation health analysis and fault analysis units realize logical judgment based on standardized indices and preset intervals. The system design simplifies the subsequent model replacement, function upgrade and multi-scenario adaptation work. It solves the problems of high functional coupling, difficulty in module replacement and operation and maintenance separation of the existing fault detection system. It has good engineering implementation and maintenance scalability.

[0021] Of course, any product implementing this invention does not necessarily need to achieve all of the advantages described above at the same time. Attached Figure Description

[0022] Figure 1 This is a flowchart of a fault diagnosis method for elevator door operator system based on deep learning according to the present invention.

[0023] Figure 2 This is a flowchart illustrating the specific steps involved in analyzing the operational health time series index of each operating stage of an elevator door operator in a deep learning-based elevator door operator system fault diagnosis method of the present invention.

[0024] Figure 3 This is a block diagram of a deep learning-based elevator door operator fault diagnosis system according to the present invention. Detailed Implementation

[0025] Please see Figure 1 This invention provides a technical solution: a deep learning-based method for diagnosing elevator door operator system faults, comprising the following steps: continuously acquiring the time-series data of the operating status of each operating stage of the elevator door operator (e.g., start-up, acceleration, constant speed, deceleration, and stop); inputting the time-series data of the operating status of each operating stage of the elevator door operator into a pre-trained deep learning model for feature extraction processing to obtain the operating sequence features of each operating stage of the elevator door operator; analyzing the operating health time-series index of each operating stage of the elevator door operator based on the operating sequence features of each operating stage of the elevator door operator; comparing the operating health time-series index of each operating stage of the elevator door operator with a corresponding preset stage operating health interval, and considering operating stages with operating health time-series indices outside the preset stage operating health intervals as operating faults.

[0026] The operating status timing data includes motor operating current, motor operating voltage, motor operating power, door operator electromagnetic interference intensity, control response delay, door friction load, and door offset angle at several time points.

[0027] Among them, the motor operating current represents the instantaneous current value of the elevator door operator motor during operation, reflecting the energy supply intensity of the drive system to the load. When it is too large or fluctuates violently, it can indicate system overload, abnormal start-up, or electrical control instability. The acquisition steps are as follows: During the data acquisition process, the current signal of each operating stage is continuously sampled by the current acquisition module installed at the input end of the door operator motor. The sampling period is set according to the synchronization signal of the control system (such as every 0.1 seconds). The acquired data is transmitted to the local controller in real time and stored in the database to form a complete current time sequence for subsequent normalization and feature extraction processing.

[0028] The motor operating voltage represents the actual power supply voltage of the elevator door motor during operation. It reflects the stability of the power supply system. Large fluctuations may cause control errors, power deviations, or electrical faults. The acquisition steps are as follows: the power supply circuit of the door motor is sampled at high frequency through a voltage acquisition module synchronized with the current. The voltage value at each time point is recorded and transmitted to the system control module after analog-to-digital conversion. All sampling points are marked with timestamps according to the operating stage and stored in the database for evaluation of voltage stability and power calculation.

[0029] Motor operating power represents the instantaneous output power of a motor at a certain point in time. It is used to measure the motor load matching status and energy consumption, and is one of the basic parameters for judging whether the system is overloaded or operating inefficiently. The steps to obtain it are as follows: the motor operating power is obtained by multiplying the collected motor operating current I(t) and operating voltage V(t) at time points. This calculation is completed in real time at the system control terminal, and the result is bound to the original sampling time and written into the database to form a complete power time series curve.

[0030] The electromagnetic interference intensity of the gantry crane indicates the amplitude of electromagnetic interference detected during the operation of the gantry crane. It can be caused by switching actions, motor commutation, or interference from external equipment. Excessive interference may affect the transmission of control commands and the reliability of sensor acquisition. The acquisition steps are as follows: the interference signal intensity is collected in real time by an electromagnetic interference detection unit (such as a near-field EMI sensor or a shielded loop detector) installed on or around the gantry crane control board. The sampling frequency is set to be synchronized with the main control loop. Each sampling result is filtered and converted to a fixed amplitude and then stored as a time sequence for subsequent calculation.

[0031] Control response delay, representing the time difference between the system issuing a gate control signal and the gate actually starting to execute the action, is a key parameter for measuring the efficiency and consistency of the control link. It is obtained by recording the timestamp T of each control command issuance.cmd With the corresponding door motion detection feedback timestamp T resp The delay value Δ=T is calculated. resp -T cmd The entire process is completed through controller log recording, encoder signal comparison, or door start sensor determination, and is stored separately according to the operation stage.

[0032] The friction load of the gate body represents the resistance level of the gate body during movement. It is a comprehensive reflection of structural wear, track impurities, and lubrication conditions. Excessive friction will lead to slow operation, abnormal heating, or jamming. The steps to obtain the friction load are as follows: the friction resistance is inferred from the instantaneous current and voltage of the gate machine drive motor and the actual movement state (e.g., based on the motor model or known power-speed relationship), and the operating state is determined by combining the gate body position encoder or acceleration sensor. The system continuously stores the obtained friction values ​​over time and uses them for slope and abnormal change analysis.

[0033] The gate offset angle represents the angle by which the gate deviates from the ideal trajectory during track operation. Continuous deviation may indicate track curvature, loose wheel system, or structural misalignment. The acquisition steps are as follows: acquire the angle change data of the gate during operation by inertial measurement units (IMU) or attitude sensors installed on the left and right guide rails or frame structure of the gate. Each record includes pitch, roll, or yaw angles. The system normalizes the offset angle data and writes it into the database to form an attitude angle time series set.

[0034] Specifically, the deep learning model includes an input embedding layer, a bidirectional GRU layer, a multi-head attention layer, a stage feature aggregation layer, and a feature output layer. The runtime sequence features include the peak value of the motor operating current, the variance of the motor operating current fluctuation, the number of high-frequency disturbances in the motor operating current, the stability coefficient of the motor operating voltage, the average intensity of the electromagnetic interference of the door operator, the duration of the electromagnetic interference of the door operator, the average delay of the control response, the maximum control response delay, the control response delay fluctuation rate, the static friction load of the door body, the dynamic friction load of the door body, the slope of the door body friction trend, the abrupt change rate of the door body offset, the average angle of the door body offset, and the door body jitter frequency.

[0035] The pre-training steps for deep learning models are as follows:

[0036] A large amount of elevator door operator operation phase data covering normal operation status and labeled fault samples is collected. The data includes original physical parameters such as motor current, voltage, power, control delay, EMI intensity, door friction load and offset angle at multiple time points, and is archived according to operation phase labels (such as start, acceleration, deceleration, stop) and health status labels (such as normal / abnormal / minor abnormality).

[0037] The above data is normalized to unify the dimensions and scale; at the same time, the training set, validation set and test set are divided, and each time series data segment is labeled according to the running stage, so that the model can perceive the correspondence between stage information and target output during training.

[0038] Set the model training objective to multi-task loss:

[0039] Run Feature Reconstruction Loss: Uses mean squared error or Huber loss to measure the difference between the physical quantity feature vector output by the model and the true label.

[0040] Normalized time-series data is input into the model structure, and then sequentially processed through input embedding, bidirectional GRU modeling, multi-head attention extraction of key points, stage feature convergence and feature decoding output. The model parameters are trained by gradient descent using Adam or RMSprop optimizers until the loss converges.

[0041] After training is completed, the model structure and parameter weights are exported and deployed in the system as a pre-trained model. It does not need to be retrained in the real running environment and can only be used as a feature extractor. Alternatively, the model can be retrained or fine-tuned at set intervals to improve its adaptability.

[0042] Among them, the peak operating current of the motor represents the maximum current reached by the motor during the operation phase, reflecting the instantaneous intensity of the start-up or load.

[0043] The variance of motor operating current fluctuation represents the degree of current fluctuation throughout the entire operation process; the larger the variance, the worse the power supply stability.

[0044] The number of high-frequency disturbances in the motor operating current indicates the number of times the current jumps frequently in a short period of time, reflecting power supply interference or control signal jitter.

[0045] The motor operating voltage stability coefficient indicates the degree of stability of the voltage around the rated value. The closer it is to 1, the better, as smaller fluctuations indicate a more stable system.

[0046] The average intensity of electromagnetic interference of the gantry crane represents the average amplitude of electromagnetic field interference detected during operation, reflecting the level of electromagnetic pollution.

[0047] The duration of electromagnetic interference (EMI) in the gantry crane indicates the total time during which EMI interference continuously affects the system's control or electrical stability.

[0048] Average control response delay represents the average time between the issuance of a control command and the action response of the gantry crane; the lower the delay, the more agile the crane.

[0049] Maximum control response delay represents the maximum response delay encountered during the entire operation phase, reflecting extreme instability conditions.

[0050] Control response delay volatility represents the degree of fluctuation in control delay, reflecting response consistency and control chain stability.

[0051] The static friction load of the door body represents the resistance load of the door body in a static state before starting, and determines whether starting is difficult or difficult.

[0052] The dynamic friction load of the gate body represents the frictional resistance encountered by the gate body during operation, reflecting track wear or foreign object problems.

[0053] The slope of the door friction trend indicates the trend of frictional force over time; an increase indicates structural wear or deterioration of lubrication.

[0054] The door offset mutation rate represents the frequency of drastic changes in the door's attitude angle over a short period of time, indicating structural loosening or imbalance impact.

[0055] The average angle of gate offset represents the average angle of gate offset during the entire operation process.

[0056] The door vibration frequency indicates the frequency of micro-vibration or mechanical structure vibration during door operation; the higher the frequency, the less stable the structure.

[0057] The specific steps for obtaining the runtime sequence features of each operating stage of the elevator door operator are as follows: In the input embedding layer of the deep learning model, the time-series data of the operating state of each operating stage of the elevator door operator are received and normalized and mapped to feature dimensions to obtain a unified-dimensional runtime sequence vector sequence of each operating stage of the elevator door operator. Specifically, the original operating state time-series data of the elevator door operator in each operating stage are received. The time-series data includes motor operating current, motor operating voltage, motor operating power, door operator electromagnetic interference intensity, control response delay, door friction load, and door offset angle at several time points. First, the above original physical data is normalized to eliminate the influence of differences in units, values, or quantities between different physical quantities, making them adaptable to a unified computational space. Then, a set of linear mapping layers or one-dimensional convolutional kernels are used to map the feature dimensions of the parameter vector at each time point, expanding the original 7-dimensional physical parameters to a unified high-dimensional embedding vector (e.g., 64-dimensional), forming a unified-structure runtime sequence vector sequence. This provides a semantically consistent representation foundation for subsequent temporal modeling. In the bidirectional GRU layer of the deep learning model, temporal modeling is performed on the runtime sequence vectors of each operating stage of the elevator door machine to capture the forward and backward dependencies of the operating state, extract the context state encoding at each time point, and obtain the temporal hidden state vector set of each operating stage of the elevator door machine. Specifically, the embedded runtime sequence vectors are received and bidirectional temporal modeling is performed on them based on the gated recurrent neural network structure. On the one hand, the forward GRU unit models the time series from front to back to capture the forward state evolution trend at each time point during the operation of the elevator door machine. On the other hand, the backward GRU unit models the time series from back to front to enhance the model's ability to perceive the impact of subsequent changes on the previous state performance. Finally, the forward and backward hidden states are spliced ​​and fused to form the temporal hidden state vector set of the current operating stage of the elevator door machine, enabling the model to understand the dynamic evolution relationship between the preceding and following states in each time step.In the multi-head attention layer of the deep learning model, the temporal hidden state vector set of each operating stage of the elevator door operator is subjected to parameter importance evaluation and time slice outlier segment focusing processing. Cross-feature representations of high-weight operating segments are extracted to obtain the stage operating state feature vector set of each operating stage of the elevator door operator with an attention weighting mechanism. Specifically, the layer receives the temporal hidden state vector set and uses a multi-head attention mechanism to evaluate the temporal information importance and perform cross-feature modeling. Specifically, this layer maps the temporal hidden state sequence into query vectors, key vectors, and value vectors respectively. The dependency weights between each time point are calculated through a scaled dot product attention mechanism, and the importance distribution under different perspectives is learned based on multiple subspaces (i.e., multi-head structure), thereby improving the model. The model's ability to focus on key segments such as high-risk segments, short-term abrupt changes, or long-term drift trends during operation is demonstrated. Ultimately, multi-head attention outputs are fused to form a sequence of stage operation state feature vectors containing attention-weighted information, providing a structurally enhanced representational foundation for subsequent global aggregation and feature decoding. In the stage feature aggregation layer of the deep learning model, the attention-weighted elevator door operator's stage operation state feature vector set for each operation stage is globally averaged and aggregated to form a stage comprehensive representation vector corresponding to that operation stage. Specifically, this involves globally averaging and aggregating the multi-head attention-processed stage operation state feature vector sequence to extract the feature representation of that operation stage over the overall time span. Specifically, global average pooling is employed. The pooling method compresses the time dimension by averaging the vectors at all time points dimensionally, resulting in a fixed-dimensional stage-comprehensive representation vector. This vector effectively compresses the dynamic evolution information, important state responses, and parameter fluctuation patterns of the current operating stage, possessing global expressive power and serving as a semantic carrier for fault mode analysis, health scoring, and feature decoding. In the feature output layer of the deep learning model, the stage-comprehensive representation vector for each operating stage of the elevator door operator is mapped, outputting the runtime sequence feature set for each operating stage. Specifically, the stage-comprehensive representation vector is input into a multi-channel mapping structure for multi-physical quantity feature processing. Explicit decoding processing: This output layer, through a fully connected network or feature branch module, structurally expands the stage vector, outputting a set of physical feature values ​​highly correlated with the actual operating state. The runtime sequence feature set includes, but is not limited to: peak motor operating current, current fluctuation variance, number of high-frequency current disturbances, voltage stability coefficient, average intensity and duration of electromagnetic interference, average control response delay, maximum delay and delay fluctuation rate, door static and dynamic friction loads, friction trend slope, mean door offset angle, offset abrupt change rate, and door jitter frequency. This feature set, as a highly semantic output, can be directly used for subsequent operational health index modeling and fault level determination.

[0058] In this implementation plan, the proposed deep learning model has a clear structure and distinct layers. It can automatically extract high semantic operational features from the original physical time-series data. Through input embedding, bidirectional GRU modeling, multi-head attention focusing, stage feature convergence, and structured decoding, the model achieves a comprehensive understanding and fine-grained characterization of the state of each operational stage of the elevator door operator. This method does not rely on manually set rules and has good context modeling and critical moment recognition capabilities. The extracted operational time-series features have high interpretability and fault correlation, providing a reliable data foundation for subsequent health index construction and fault level judgment, and significantly improving the intelligence and adaptability of the system.

[0059] Specifically, such as Figure 2 As shown, the specific steps for analyzing the operational health time-series index of each operating stage of the elevator door operator are as follows: Based on the operational sequence characteristics of each operating stage of the elevator door operator, the drive system health index, control response stability index, and door structure attitude index of each operating stage of the elevator door operator are analyzed respectively; the drive system health index, control response stability index, and door structure attitude index of each operating stage of the elevator door operator are comprehensively analyzed to obtain the operational health time-series index of each operating stage of the elevator door operator.

[0060] The specific formula for calculating the operational health time series index of an elevator door operator during a certain operating phase is as follows: ;in, This refers to the operational health time series index of an elevator door operator during a specific operational phase. This refers to the health index of the drive system of the elevator door operator during a certain operating phase. This is a natural constant, and in this embodiment, it is taken as 2.71. This refers to the door structure attitude index during a certain operating phase of the elevator door operator. The structural health adjustment coefficients are stored in the database. This refers to the control response stability index of the elevator door operator during a certain operating phase. These are the control and coordination enhancement coefficients stored in the database.

[0061] It needs to be explained that the structural health adjustment coefficients stored in the database Control coordination enhancement coefficient The specific acquisition steps are as follows: First, based on the elevator door operator fault records and corresponding time series data of drive system health index, control response stability index and door structure attitude index from multiple operating stages in historical samples, a standard operating dataset for parameter training is constructed. Second, this dataset is input into a predefined objective function, which minimizes the deviation between the operating health time series index and the actual manually judged health level. On this basis, a gradient descent optimization algorithm (such as Adam optimizer or adaptive momentum method) is used to jointly iteratively update the structural health adjustment coefficient and the control coordination enhancement coefficient, and finally converges to the coefficient value that best matches the operating health time series index with the actual health state. The parameter acquisition process can be completed in one-time training during the system initialization stage, or it can be periodically and dynamically fine-tuned in combination with real-time operating data to ensure that the operating health time series index evaluation mechanism has adaptability and accuracy under different elevator models or operating conditions.

[0062] The following is a specific implementation example for calculating the operational health time series index of an elevator door operator during a certain operating phase, with the following parameters:

[0063] The health index of the drive system of the elevator door operator during a certain operating phase is approximately 0.742.

[0064] Natural constant The value is 2.71.

[0065] The door structure attitude index of the elevator door operator during a certain operating phase is approximately 0.435.

[0066] The structural health adjustment coefficient stored in the database is approximately 1.215.

[0067] The control response stability index of the elevator door operator during a certain operating phase is approximately 0.668.

[0068] The control coordination enhancement coefficient stored in the database is approximately 1.034.

[0069] Substituting the above data into the specific formula for calculating the operational health time series index of a certain operating stage of the elevator door operator, we obtain:

[0070] The operational health time series index of an elevator door operator at a certain operating stage = (((0.742×2.71^(-0.435))^1.215) + (((0.668^2) / (1+|0.742-0.668|))^1.034))^0.5≈0.902.

[0071] In this implementation scheme, a three-dimensional decomposition mechanism is introduced, comprising the drive system health index, control response stability index, and door structure attitude index, to achieve a structured and multi-dimensional assessment of the elevator door operator's operating status. Simultaneously, the use of structural health adjustment coefficients and control coordination enhancement coefficients trained based on historical sample data enables the calculation of the comprehensive operational health time series index to possess data-driven adaptive capabilities. This method not only enhances the consistency between the index results and the actual fault level but also allows for dynamic fine-tuning based on different elevator models and operating conditions, significantly improving the accuracy, interpretability, and cross-scenario applicability of the assessment indicators. This provides a robust, adjustable, and transferable core support mechanism for intelligent system diagnosis.

[0072] Specifically, the steps for analyzing the health index of the drive system of the elevator door operator at each operating stage are as follows: read the peak value of the motor operating current, the variance of the motor operating current fluctuation, the number of high-frequency disturbances in the motor operating current, and the stability coefficient of the motor operating voltage at each operating stage of the elevator door operator; perform comprehensive analysis on the peak value of the motor operating current, the variance of the motor operating current fluctuation, the number of high-frequency disturbances in the motor operating current, and the stability coefficient of the motor operating voltage at each operating stage of the elevator door operator to obtain the health index of the drive system at each operating stage of the elevator door operator.

[0073] The specific formula for calculating the health index of the drive system of an elevator door operator during a certain operating phase is as follows: ;in, This refers to the health index of the drive system of the elevator door operator during a certain operating phase. It is a natural constant. This refers to the peak operating current of the elevator door operator's motor during a specific operating phase. The peak current adjustment coefficient is stored in the database. This represents the variance of the motor operating current fluctuation during a certain operating phase of the elevator door operator. This is the operating current adjustment factor stored in the database, used to prevent the denominator from being zero. This refers to the motor operating voltage stability coefficient during a certain operating phase of the elevator door operator. The voltage stability adjustment coefficients are stored in the database. This refers to the number of high-frequency disturbances in the motor operating current during a certain operating phase of the elevator door operator.

[0074] It needs to be explained that the current peak adjustment coefficient stored in the database Voltage stability regulation coefficient The specific acquisition steps are as follows: First, extract the complete time series of motor operating current and voltage corresponding to each operating stage from the historical data; second, normalize the current peak and voltage stability in each time period, and calculate the relative fluctuation of the current peak and the magnitude of the voltage deviation from the rated value; then, perform mean processing on multiple operating samples to construct the current peak adjustment coefficient in the database. With voltage stability regulation coefficient These are used as adjustment parameters to determine the degree of influence of current and voltage on the health status of the drive system, respectively.

[0075] In this implementation scheme, a drive system health index constructed based on physical characteristic parameters is introduced. This index systematically integrates key indicators such as peak motor operating current, current fluctuation variance, number of high-frequency current disturbances, and voltage stability. It comprehensively reflects the load pressure and power supply stability of the drive system at each operating stage. At the same time, by using the peak current adjustment coefficient and voltage stability adjustment coefficient obtained from pre-training in the database, adaptive corrections are achieved for different elevator models, motor performances, and on-site operating environments. This method not only enhances the index's ability to represent the real operating state but also improves the consistency and robustness of cross-scenario evaluation through the adjustment coefficient mechanism. It provides accurate and interpretable underlying drive dimension support for subsequent health determination and fault analysis.

[0076] Specifically, the steps for analyzing the control response stability index of each operating stage of the elevator door operator are as follows: Read the average intensity of electromagnetic interference (EMI), duration of EMI, average control response delay, maximum control response delay, and control response delay fluctuation rate of the elevator door operator for each operating stage; perform comprehensive analysis on the average intensity of EMI, duration of EMI, average control response delay, maximum control response delay, and control response delay fluctuation rate of the elevator door operator for each operating stage to obtain the control response stability index of the elevator door operator for each operating stage.

[0077] The specific formula for calculating the control response stability index of an elevator door operator during a certain operating phase is as follows: ;in, This refers to the control response stability index of the elevator door operator during a certain operating phase. The average delay of the control response during a certain operating phase of the elevator door operator. The control response delay fluctuation rate of the elevator door operator during a certain operating phase. This represents the average electromagnetic interference intensity of the elevator door operator during a certain operating phase. The duration of electromagnetic interference in the elevator door operator during a certain operating phase. This represents the maximum control response delay during a certain operating phase of the elevator door operator. This is a response latency adjustment factor stored in the database to prevent the denominator from being zero.

[0078] In this implementation scheme, by constructing a control response stability index, key control chain parameters such as average control response delay, maximum delay, delay fluctuation rate, and the intensity and duration of electromagnetic interference are integrated into a model. This allows for a comprehensive evaluation of the control chain stability and execution consistency of the elevator door operator at each stage of operation. Furthermore, the design of a stability factor in the denominator of the index formula enhances the computational robustness to extreme values ​​and special operating conditions. Compared to traditional methods that rely solely on a single delay threshold, this index reveals the system's dynamic response capability and internal coordination efficiency under disturbances. It is particularly sensitive to identifying latent problems such as command-action asynchrony and control drift, which helps improve the early warning capability and response accuracy of fault diagnosis.

[0079] Specifically, the steps for analyzing the door structure attitude index of each operating stage of the elevator door operator are as follows: Read the door static friction load, door dynamic friction load, door friction trend slope, door offset abrupt change rate, door offset average angle, and door shaking frequency for each operating stage of the elevator door operator; Perform comprehensive analysis on the door static friction load, door dynamic friction load, door friction trend slope, door offset abrupt change rate, door offset average angle, and door shaking frequency for each operating stage of the elevator door operator to obtain the door structure attitude index for each operating stage of the elevator door operator.

[0080] The specific formula for calculating the operational health time series index of an elevator door operator during a certain operating phase is as follows: ;in, This refers to the door structure attitude index during a certain operating phase of the elevator door operator. This represents the rate of sudden change in door offset during a certain operating phase of the elevator door operator. This refers to the door vibration frequency during a certain operating phase of the elevator door operator. This represents the average angle of door body offset during a certain operating phase of the elevator door operator. This is an offset angle adjustment factor stored in the database, used to prevent the denominator from being zero. This refers to the static friction load on the elevator door body during a certain operating phase of the elevator door operator. This refers to the dynamic friction load on the elevator door body during a certain operating phase of the elevator door operator. This refers to the friction load adjustment factor stored in the database, used to prevent the denominator from being zero. This represents the slope of the door friction trend during a certain operating phase of the elevator door operator.

[0081] The tanh function is specifically represented as follows: ,in, It is a natural constant, and in this example it can be taken as 2.71, with a domain of (−∞, +∞) and a range of (−1, +1).

[0082] In this implementation plan, a gate structure attitude index is constructed to integrate key structural parameters such as static and dynamic friction loads, friction trend slope, offset mutation rate, offset angle, and jitter frequency. This comprehensively characterizes the stability and attitude consistency of the gate during operation. The tanh function is introduced into the formula for friction trend modeling, which can effectively suppress the amplification effect of outliers and highlight the nonlinear response of progressive structural offset and wear risk. At the same time, by setting the denominator stability factor, the numerical robustness of the index under low-speed and low-friction scenarios is improved. This index shows higher sensitivity to hidden structural risks such as gate rail deformation, component loosening, and long-term wear, which helps to achieve early detection and accurate judgment of structural faults.

[0083] Please see Figure 3 This invention provides a technical solution: a deep learning-based elevator door operator fault diagnosis system, comprising: a data acquisition unit for continuously acquiring time-series data of the operating status of each operating stage of the elevator door operator; a feature extraction unit for inputting the time-series data of the operating status of each operating stage of the elevator door operator into a pre-trained deep learning model for feature extraction processing to obtain the operating sequence features of each operating stage of the elevator door operator; an operating health analysis unit for analyzing the operating health time-series index of each operating stage of the elevator door operator based on the operating sequence features of each operating stage of the elevator door operator; and an operating fault analysis unit for comparing the operating health time-series index of each operating stage of the elevator door operator with a corresponding preset stage operating health interval, and considering operating stages with operating health time-series indices outside the preset stage operating health intervals as operating faults.

[0084] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0085] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A fault diagnosis method for elevator door operator systems based on deep learning, characterized in that, Includes the following steps: Continuously acquire the timing data of the operating status of the elevator door operator at each stage of operation; The time-series data of the operating status of each stage of the elevator door operator are input into a pre-trained deep learning model for feature extraction processing to obtain the time-series features of each stage of the elevator door operator. Based on the operational sequence characteristics of each operating stage of the elevator door operator, the operational health time sequence index of each operating stage of the elevator door operator is analyzed. The operational health time series index of each operating stage of the elevator door operator is compared with the corresponding preset stage operational health range for analysis, and the operating stage where the operational health time series index is outside the preset stage operational health range is regarded as an operational failure.

2. The elevator door operator system fault diagnosis method based on deep learning according to claim 1, characterized in that, The operating status timing data includes motor operating current, motor operating voltage, motor operating power, door operator electromagnetic interference intensity, control response delay, door friction load, and door offset angle at several time points.

3. The elevator door operator system fault diagnosis method based on deep learning according to claim 1, characterized in that, The deep learning model includes an input embedding layer, a bidirectional GRU layer, a multi-head attention layer, a stage feature aggregation layer, and a feature output layer. The runtime sequence features include peak motor operating current, variance of motor operating current fluctuation, number of high-frequency disturbances in motor operating current, stability coefficient of motor operating voltage, average intensity of electromagnetic interference of the door operator, duration of electromagnetic interference of the door operator, average control response delay, maximum control response delay, control response delay fluctuation rate, static friction load of the door body, dynamic friction load of the door body, slope of the door body friction trend, abrupt change rate of door body offset, average angle of door body offset, and door body shaking frequency.

4. The elevator door operator system fault diagnosis method based on deep learning according to claim 1, characterized in that, The specific steps to obtain the runtime sequence characteristics of each operating stage of the elevator door operator are as follows: In the input embedding layer of the deep learning model, the time sequence data of the operation status of each operation stage of the elevator door machine is received, and normalization and feature dimension mapping are performed to obtain a time sequence vector sequence of each operation stage of the elevator door machine with a unified dimension. In the bidirectional GRU layer of the deep learning model, the runtime sequence vector sequence of each operating stage of the elevator door machine is processed by temporal modeling to capture the forward and backward dependencies of the operating state, extract the context state encoding at each time point, and obtain the temporal hidden state vector set of each operating stage of the elevator door machine. In the multi-head attention layer of the deep learning model, the parameter importance evaluation and time slice abnormal segment focusing processing are performed on the temporal hidden state vector set of each operating stage of the elevator door machine. The cross feature representation of high-weight operating segments is extracted to obtain the stage operating state feature vector set of each operating stage of the elevator door machine with attention weighting mechanism. In the stage feature aggregation layer of the deep learning model, the stage operation state feature vector set of each operation stage of the elevator door machine after attention weighting is globally averaged and aggregated to form a stage comprehensive representation vector corresponding to that operation stage. In the feature output layer of the deep learning model, the stage comprehensive representation vector of each operating stage of the elevator door machine is mapped and processed to output the runtime sequence feature set of each operating stage of the elevator door machine.

5. The elevator door operator system fault diagnosis method based on deep learning according to claim 1, characterized in that, The specific steps for analyzing the operational health time series index of each operating stage of the elevator door operator are as follows: Based on the runtime sequence characteristics of each operating stage of the elevator door operator, the drive system health index, control response stability index, and door structure attitude index of each operating stage of the elevator door operator are analyzed respectively. The health index of the drive system, the stability index of the control response, and the attitude index of the door structure are comprehensively analyzed for each operating stage of the elevator door operator to obtain the operating health time series index of each operating stage of the elevator door operator.

6. The elevator door operator system fault diagnosis method based on deep learning according to claim 5, characterized in that, The specific formula for calculating the operational health time series index of an elevator door operator during a certain operating phase is as follows: ； in, , , , The following are, in order: the operational health timing index of the elevator door operator at a certain operating stage, the drive system health index, the door structure attitude index, and the control response stability index. It is a natural constant. , These are, in order, the structural health adjustment coefficient and the control coordination enhancement coefficient stored in the database.

7. The elevator door operator system fault diagnosis method based on deep learning according to claim 5, characterized in that, The specific steps for analyzing the health index of the drive system at each stage of elevator door operator operation are as follows: Read the peak value of motor operating current, the variance of motor operating current fluctuation, the number of high-frequency disturbances in motor operating current, and the stability coefficient of motor operating voltage for each operating stage of the elevator door operator; The peak value of motor operating current, variance of motor operating current fluctuation, number of high-frequency disturbances in motor operating current, and stability coefficient of motor operating voltage in each operating stage of the elevator door operator are comprehensively analyzed to obtain the health index of the drive system of the elevator door operator in each operating stage.

8. The elevator door operator system fault diagnosis method based on deep learning according to claim 5, characterized in that, The specific steps for analyzing the control response stability index of each operating stage of the elevator door operator are as follows: Read the average intensity of electromagnetic interference (EMI) of the elevator door operator, the duration of EMI, the average delay of control response, the maximum delay of control response, and the fluctuation rate of control response delay for each stage of elevator door operator operation. The average intensity of electromagnetic interference, duration of electromagnetic interference, average delay of control response, maximum delay of control response, and fluctuation rate of control response are comprehensively analyzed for each operating stage of the elevator door operator to obtain the control response stability index for each operating stage of the elevator door operator.

9. The elevator door operator system fault diagnosis method based on deep learning according to claim 5, characterized in that, The specific steps for analyzing the door structure attitude index at each stage of elevator door operator operation are as follows: Read the door static friction load, door dynamic friction load, door friction trend slope, door offset abrupt change rate, door offset average angle, and door vibration frequency for each operating stage of the elevator door operator; The static friction load, dynamic friction load, friction trend slope, abrupt change rate of door offset, average offset angle, and shaking frequency of the elevator door machine in each operating stage are comprehensively analyzed to obtain the door structure attitude index of the elevator door machine in each operating stage.

10. A deep learning-based elevator door operator system fault diagnosis system, employing the deep learning-based elevator door operator system fault diagnosis method according to any one of claims 1-9, characterized in that, include: The data acquisition unit is used to continuously acquire the time-series data of the operating status of the elevator door operator at each operating stage; The feature extraction unit is used to input the time-series data of the operating status of each operating stage of the elevator door machine into the pre-trained deep learning model for feature extraction processing, so as to obtain the operating sequence features of each operating stage of the elevator door machine. The operational health analysis unit is used to analyze the operational health time series index of each operational stage of the elevator door operator based on the operational sequence characteristics of each operational stage. The operation fault analysis unit is used to compare the operation health time series index of each operation stage of the elevator door operator with the corresponding preset stage operation health interval, and to regard the operation stage where the operation health time series index is outside the preset stage operation health interval as an operation fault.

Images