Intelligent elevator fault diagnosis method based on factorization gate graph attention network

Abstract

The application discloses an elevator fault intelligent diagnosis method based on a factorization gate graph attention network, and belongs to the technical field of fault diagnosis. The method comprises the following steps: collecting three-phase current, three-axis vibration and displacement signals of an elevator to divide a data set by period slicing; constructing node features by differential feature extraction, and obtaining a graph data set by following physical links to construct an electromechanical causal graph; inputting the training graph data into a network, performing soft mask adaptive weighting and noise reduction by a front gate module, aggregating information by a graph attention network, and performing secondary feature checking by a latent gate module between layers; extracting a central displacement node as a graph-level representation to output a prediction logic value, calculating an error by combining a degradation perception cost-sensitive loss function, updating parameters, obtaining a recognition model, and testing the recognition model. The application effectively fuses multi-source physical information, overcomes feature dilution and noise interference, and significantly improves the accuracy and robustness of fault diagnosis.

Description

Technical Field

[0001] This invention relates to the field of fault diagnosis technology, specifically to an intelligent elevator fault diagnosis method based on factorized gated graph attention networks. Background Technology

[0002] Vertical transportation special equipment plays an irreplaceable role in modern urban three-dimensional transportation networks, and the operational reliability of its key motion actuators is directly related to the overall safety of personnel and equipment. Once an actuator malfunctions or fails, it can easily lead to serious safety accidents such as equipment malfunction or even falls. For a long time, fault diagnosis for such equipment has relied heavily on manual inspection and experience-based judgment. This traditional method not only has significant time lag but also fails to accurately detect subtle degradation risks in the early stages of equipment development, and can no longer meet the urgent demands of modern complex electromechanical systems for high reliability and high safety.

[0003] Currently, data-driven intelligent fault diagnosis methods have gradually become a research hotspot due to their powerful feature extraction capabilities. However, most existing methods mainly focus on extracting time-series or frequency-domain features from signals of a single monitoring dimension, largely ignoring the inherent spatial coupling relationships between multi-source heterogeneous monitoring data. Although some studies have attempted to introduce multi-channel fusion models based on gridded convolution to capture these dependencies, enriching the feature dimensions to some extent, these models are limited by their inherent gridded convolution mechanism, making it difficult to explicitly capture the physical causal correlations between cross-modal sensing information. When performing convolution or stitching operations on multi-source data, it is essentially still limited to calculating the weighted sum between sensing information and the convolution kernel, failing to fully explore and utilize the physical dependencies and interactive evolution laws between multi-dimensional sensing nodes. This results in weak fault features being easily masked by noise, limiting diagnostic accuracy and robustness. Summary of the Invention

[0004] The purpose of this invention is to provide an intelligent elevator fault diagnosis method based on a factorized gated graph attention network, so as to solve the problems mentioned in the background art.

[0005] To address the aforementioned technical problems, this invention provides an intelligent elevator fault diagnosis method based on a factorized gated graph attention network, comprising:

[0006] Sensors were used to collect three-phase drive current, three-axis mechanical vibration and displacement signals of the elevator system. Based on the displacement signals, the data was processed by slicing the reciprocating motion cycle of the door operator to divide it into an initial training dataset and an initial test dataset.

[0007] For periodic slices, the extreme points in the displacement signal are detected to pinpoint the limit positions of the gantry crane's reciprocating motion, and the three-phase drive current and three-axis mechanical vibration signals within the slice period are synchronized. Fault data for the elevator system includes normal operating conditions, varying degrees of door jamming, and broken strands in the wire rope.

[0008] The initial training dataset and the initial test dataset are processed separately. Differential feature extraction is performed based on the physical attributes of the multi-source sensor nodes, and node features are constructed using time-domain analysis and wavelet packet transform.

[0009] Based on this, the three types of sensor nodes are abstracted into a set of nodes in a heterogeneous graph. Following the physical link of power input-mechanical transmission-macroscopic motion, an electromechanical causal graph is constructed, resulting in training and testing graph datasets. The details are as follows:

[0010] Multidimensional time-domain features are extracted from the three-phase drive current signal, including root mean square, average value, peak-to-peak value, standard deviation, skewness, kurtosis, crest factor, waveform factor, logarithm of total energy, volatility, and DC offset.

[0011] After extraction, these feature vectors are filled with zero values ​​to a uniform dimension. .

[0012] Among them, volatility Defined as the ratio of the average of the absolute values ​​of the first-order differences of the signal to the root mean square, the formula is:

[0013] ;

[0014] In the formula, This represents the number of sampling points within the slice period. For the first Current values ​​at each sampling point To prevent extremely small constants with a denominator of zero.

[0015] Logarithmic energy of each subband coefficient The calculation formula is:

[0016] ;

[0017] In the formula, Represented as the first The first sub-band Wavelet packet reconstruction coefficients, This represents the total number of coefficients for that frequency band.

[0018] Logarithm of crest factor The calculation formula is:

[0019] ;

[0020] In the formula, The maximum absolute value of the signal. The root mean square of the signal, and The function is used to truncate the logarithmic input of the crest factor to enhance numerical stability. Additionally, the logarithmic value of the impulse factor, the low-frequency energy ratio, and the kurtosis variation constitute high-dimensional features.

[0021] DC offset is defined as the ratio of the average absolute value of a signal to its root mean square (RMS).

[0022] Performing db4 wavelet basis analysis on triaxial mechanical vibration signals Layer wavelet packet decomposition Each sub-band is used to calculate the logarithmic energy of the coefficients of each sub-band, the energy difference between adjacent bands, the root mean square, peak-to-peak value, standard deviation, skewness, kurtosis, and higher-order surrogate indices of the global signal.

[0023] Higher-order surrogate metrics include the logarithm of the crest factor, the logarithm of the impulse factor, the low-frequency energy ratio, and the kurtosis variation, which together form high-dimensional features.

[0024] For displacement signals, the macroscopic kinematic response of the elevator door operator system is primarily characterized. Based on the physical motion laws of the elevator door operator, the system is divided into three stages according to speed change characteristics: acceleration, constant speed operation, and braking deceleration. Local standard deviations for each stage are extracted, and the dynamic ratio of the standard deviations for the braking and constant speed stages is calculated. A second-order jerk index is obtained by performing two numerical differences on the displacement signal. The absolute maximum value and ratio of the jerk during the acceleration and braking stages, the global average absolute jerk, and the absolute ratio of the average jerk values ​​for both stages are calculated. These are combined with the global range, global standard deviation, and global root mean square to construct a multidimensional feature, which is then filled with zero values ​​to a unified dimension. .

[0025] After feature extraction, independent batch normalization networks are used to align the feature distribution for node features of different physical modes, thereby eliminating dimensional differences between heterogeneous sensors.

[0026] Construct an electromechanical causal graph to obtain training and test graph datasets. Details are as follows:

[0027] Locally fully connected edges are established between three-phase current nodes to learn phase interaction; locally fully connected edges are established between three-axis vibration nodes to learn mechanical vibration coupling in three-dimensional space.

[0028] Using macroscopic displacement nodes as the central nodes of the graph, construct cross-modal physical causal edges with all current and vibration nodes pointing unidirectionally to the central nodes, thus completing the construction of the graph topology matrix.

[0029] The node feature matrices of the initial training dataset and the initial test dataset, after independent batch normalization and alignment, are combined with the graph topology matrix to construct graph structure data, thus obtaining the training graph dataset and the test graph dataset, respectively.

[0030] The training graph dataset is input into a factorized gated graph attention network, and the features of each node are normalized.

[0031] The input pre-gating module performs soft mask adaptive weighting and noise reduction, specifically as follows:

[0032] At the sensor node level, for each node's features, calculate the node weight vector:

[0033] ;

[0034] ;

[0035] In the formula, The feature matrix of each sensor node; This represents the node importance score vector; Node weights; This represents a multilayer perceptron used for node importance assessment; This represents the normalization function.

[0036] It is a learnable temperature parameter used to control the stringency of node selection.

[0037] In the feature dimension, to reduce the parameter complexity of generating a high-dimensional full feature weight matrix and effectively capture the coupling relationship between heterogeneous sensors and statistical indicators of various dimensions, global average pooling is performed on the features of each node to obtain a graph-level summary representation. The summary representation is then input into a multilayer perceptron to generate two latent representation matrices, and low-rank decomposition is used to generate the feature weight matrix.

[0038] ;

[0039] In the formula, Represented as a weight matrix along the feature dimension; Use the Sigmoid activation function; This represents the number of terms in a low-rank decomposition. and They represent the first Two latent representation vectors corresponding to each low-rank decomposition term. This is for outer product operations.

[0040] Final gated output Defined as residual form:

[0041] ;

[0042] In the formula, The input node feature matrix; For gating strength; This indicates element-wise multiplication.

[0043] The refined features are then adaptively calculated using a graph attention network to determine the attention weights between connected nodes, enabling information aggregation and state updates between nodes. The specific steps are as follows:

[0044] Nodes in the diagram The input feature vector is The node features are updated through a graph attention mechanism, utilizing a learnable linear transformation matrix. Map the node features.

[0045] According to the node Its set of neighboring nodes Attention weights are assigned based on the degree of correlation between nodes, and the features of neighboring nodes are aggregated in a weighted manner.

[0046] A multi-head attention mechanism is introduced to concatenate the aggregation results of multiple attention heads to obtain nodes. Update features .

[0047] In inter-layer transmission, to overcome the feature dilution phenomenon that easily occurs during information transmission, a latent gating module is used for secondary feature verification.

[0048] After input-side gating enhancement, a graph attention network is used to adaptively learn the topological dependencies and energy transfer patterns between cross-modal sensor nodes.

[0049] However, feature dilution can easily occur during information transmission, causing the weak fault characteristics of critical nodes to be masked by redundant information or macroscopic noise from neighboring nodes.

[0050] To address this, latent factorization gating is embedded between GAT convolutional layers. Unlike the physical feature selection at the input, this gating operates in the abstract high-dimensional space of the graph hidden layer, performing secondary calibration on node importance and feature channels by combining global graph awareness and dual residual mapping. The specific steps are as follows:

[0051] For the hidden layer feature matrix Calculate the node attention vector:

[0052] ;

[0053] ;

[0054] In the formula, This represents the node importance score vector obtained from the hidden layer feature matrix; This represents a multilayer perceptron used for node importance assessment; For the number of nodes, For the hidden layer dimension.

[0055] This is the updated node attention weight vector calculated based on the hidden layer feature matrix; Represents the normalization function; The temperature parameter is learnable; This is the node attention scaling factor.

[0056] To uncover dependencies between feature dimensions, context pooling is used to extract node information from the current graph, thereby obtaining a macroscopic representation of the entire graph. ,Will Input gating network generates feature channel weights The formula is:

[0057] ;

[0058] ;

[0059] In the formula, Indicates the first Hidden feature vectors of each node; Use the Sigmoid activation function; This represents a multilayer perceptron used to generate feature channel weights.

[0060] To prevent gradient vanishing or weight extremes in deep networks, the model introduces residual terms with upper and lower bound constraints to dynamically scale nodes and feature channels respectively:

[0061] ;

[0062] ;

[0063] In the formula, and These represent the dynamic scaling factors for the node and channel dimensions, respectively. This is a truncation function used to restrict the input to a given range; For learnable gating strength.

[0064] The final refined characteristics output are: ;

[0065] in, This represents the output feature matrix after being jointly scaled by the node dimension and the feature channel dimension; The hidden layer feature matrix; This indicates element-wise multiplication.

[0066] The central macroscopic displacement node is used as the final graph-level representation, and the classification prediction logic value of the fault state is output.

[0067] After multi-layer graph attention aggregation and latent refinement, the model abandons the traditional global average pooling and directly extracts the output of macroscopic displacement nodes as the final representation of the state.

[0068] Because this node is located at the center of the heterogeneous topology, it naturally gathers the degradation information of the driving current and mechanical vibration nodes during the graph evolution process. It not only preserves the complete link of fault propagation, but also makes its feature vector have a global view and physical discrimination ability.

[0069] Define the index of the center displacement node in the graph topology as Then the final representation vector at the graph level Represented as:

[0070] ;

[0071] The final graph-level representation vector is input into the classifier, and after nonlinear activation and random deactivation, the classification prediction logistic value of the fault state is calculated. .

[0072] Predict logical values ​​based on classification Calculate the cross-entropy loss term using the predicted and actual fault labels. Obtain the cost penalty term by querying the physical cost matrix based on the correspondence between the predicted and actual labels.

[0073] The cost penalty term is calculated based on a preset physical cost matrix, and the rules for constructing the physical cost matrix are as follows:

[0074] The penalty cost is zero when the predicted label matches the true label.

[0075] When the predicted category and the true category belong to the same fault category and a misjudgment occurs, a basic penalty cost is imposed.

[0076] When a cross-domain misjudgment occurs between the predicted category and the actual category, which belong to different fault categories, the highest penalty cost is imposed.

[0077] For similar faults that exhibit progressive degradation, a penalty cost is set based on the degradation distance between the prediction stage and the actual stage, with the penalty cost for a misjudgment at a skipped stage being higher than the penalty cost for a misjudgment at an adjacent stage.

[0078] By combining the cross-entropy loss term with the cost penalty term, a degradation-aware cost-sensitive loss function is constructed, and the error cost is calculated.

[0079] The gradient is calculated based on the error cost, and the weight parameters of the factorized gated graph attention network are updated through the backpropagation algorithm until the error cost converges, thus obtaining the fault intelligent identification model.

[0080] The test image dataset is input into the fault intelligent identification model to obtain the final classification and identification results, thereby realizing intelligent fault identification of the elevator system. The details are as follows:

[0081] The test dataset is input into the trained fault intelligent identification model, and the predicted logical value vector is output through forward propagation.

[0082] Subsequently, the predicted probability distribution of each fault state category is calculated using the Softmax function to achieve intelligent fault identification of the elevator system.

[0083] Compared with the prior art, the beneficial effects achieved by the present invention are:

[0084] Unlike traditional graph neural networks that employ full connectivity and rely on global average pooling, this invention combines physical causality to construct a graph structure that matches the real fault propagation chain, and extracts the central displacement node as the final state representation, avoiding node interference and giving the model physical interpretability.

[0085] The factorization gating mechanism designed in this invention effectively overcomes the problems of feature dilution and noise interference in heterogeneous sensor fusion. Addressing the issue of large dimensional differences between heterogeneous sensors and the easy subtraction of weak degenerate features, this invention designs a dual factorization gating mechanism at the input end and the hidden layer. This mechanism improves the model's accuracy and robustness in capturing weak fault features through low-rank decomposition and adaptive weighting and residual purification in both sensor nodes and feature channels. Attached Figure Description

[0086] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0087] Figure 1 This is a flowchart illustrating the intelligent elevator fault diagnosis method based on factorized gated graph attention network of the present invention.

[0088] Figure 2 This is a schematic diagram of the triaxial vibration data of the present invention;

[0089] Figure 3 This is a schematic diagram of the three-phase current data of the present invention;

[0090] Figure 4 This is a schematic diagram of laser displacement data of the present invention;

[0091] Figure 5This is a schematic diagram of the confusion matrix of the network recognition results of the present invention;

[0092] Figure 6 This is a comparison chart of diagnostic indicators between the network of this invention and other networks. Detailed Implementation

[0093] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. 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.

[0094] Preliminary progress has been made in the research on multi-source data fault diagnosis based on graph neural networks, but in complex electromechanical systems, there are problems such as weak features being submerged and information interference from irrelevant nodes.

[0095] To further explore the physical and spatiotemporal dependencies among multi-source heterogeneous data, physical constraints and gating mechanisms are integrated. First, a causal topology graph is constructed. Then, a gating network is used to perform soft mask adaptive weighting on the node and feature dimensions to achieve dynamic noise reduction and purification of the data. This allows information to retain fault evolution characteristics along the causal topology path, enhances the ability to capture key fault features, and comprehensively improves diagnostic accuracy.

[0096] Please see Figure 1 This invention provides an intelligent elevator fault diagnosis method based on a factorized gated graph attention network, comprising:

[0097] Sensors were used to collect three-phase drive current, three-axis mechanical vibration and displacement signals of the elevator system. Based on the displacement signals, the data was processed by slicing the reciprocating motion cycle of the door operator to divide it into an initial training dataset and an initial test dataset.

[0098] For periodic slices, the extreme points in the displacement signal are detected to pinpoint the limit positions of the gantry crane's reciprocating motion, and the three-phase drive current and three-axis mechanical vibration signals within the slice period are synchronized. Fault data for the elevator system includes normal operating conditions, varying degrees of door jamming, and broken strands in the wire rope.

[0099] The initial training dataset and the initial test dataset are processed separately. Differential feature extraction is performed based on the physical attributes of the multi-source sensor nodes, and node features are constructed using time-domain analysis and wavelet packet transform.

[0100] Based on this, the three types of sensor nodes are abstracted into a set of nodes in a heterogeneous graph. Following the physical link of power input-mechanical transmission-macroscopic motion, an electromechanical causal graph is constructed, resulting in training and testing graph datasets. The details are as follows:

[0101] Multidimensional time-domain features are extracted from the three-phase drive current signal, including root mean square, average value, peak-to-peak value, standard deviation, skewness, kurtosis, crest factor, waveform factor, logarithm of total energy, volatility, and DC offset.

[0102] After extraction, these feature vectors are filled with zero values ​​to a uniform dimension. .

[0103] Among them, volatility Defined as the ratio of the average of the absolute values ​​of the first-order differences of the signal to the root mean square, the formula is:

[0104] ;

[0105] In the formula, This represents the number of sampling points within the slice period. For the first Current values ​​at each sampling point To prevent extremely small constants with a denominator of zero.

[0106] Logarithmic energy of each subband coefficient The calculation formula is:

[0107] ;

[0108] In the formula, Represented as the first The first sub-band Wavelet packet reconstruction coefficients, This represents the total number of coefficients for that frequency band.

[0109] Logarithm of crest factor The calculation formula is:

[0110] ;

[0111] In the formula, The maximum absolute value of the signal. The root mean square of the signal, and The function is used to truncate the logarithmic input of the crest factor to enhance numerical stability. Additionally, the logarithmic value of the impulse factor, the low-frequency energy ratio, and the kurtosis variation constitute high-dimensional features.

[0112] DC offset is defined as the ratio of the average absolute value of a signal to its root mean square (RMS).

[0113] Performing db4 wavelet basis analysis on triaxial mechanical vibration signals Layer wavelet packet decomposition Each sub-band is used to calculate the logarithmic energy of the coefficients of each sub-band, the energy difference between adjacent bands, the root mean square, peak-to-peak value, standard deviation, skewness, kurtosis, and higher-order surrogate indices of the global signal.

[0114] Higher-order surrogate metrics include the logarithm of the crest factor, the logarithm of the impulse factor, the low-frequency energy ratio, and the kurtosis variation, which together form high-dimensional features.

[0115] For displacement signals, the macroscopic kinematic response of the elevator door operator system is primarily characterized. Based on the physical motion laws of the elevator door operator, the system is divided into three stages according to speed change characteristics: acceleration, constant speed operation, and braking deceleration. Local standard deviations for each stage are extracted, and the dynamic ratio of the standard deviations for the braking and constant speed stages is calculated. A second-order jerk index is obtained by performing two numerical differences on the displacement signal. The absolute maximum value and ratio of the jerk during the acceleration and braking stages, the global average absolute jerk, and the absolute ratio of the average jerk values ​​for both stages are calculated. These are combined with the global range, global standard deviation, and global root mean square to construct a multidimensional feature, which is then filled with zero values ​​to a unified dimension. .

[0116] After feature extraction, independent batch normalization networks are used to align the feature distribution for node features of different physical modes, thereby eliminating dimensional differences between heterogeneous sensors.

[0117] Construct an electromechanical causal graph to obtain training and test graph datasets. Details are as follows:

[0118] Locally fully connected edges are established between three-phase current nodes to learn phase interaction; locally fully connected edges are established between three-axis vibration nodes to learn mechanical vibration coupling in three-dimensional space.

[0119] Using macroscopic displacement nodes as the central nodes of the graph, construct cross-modal physical causal edges with all current and vibration nodes pointing unidirectionally to the central nodes, thus completing the construction of the graph topology matrix.

[0120] The node feature matrices of the initial training dataset and the initial test dataset, after independent batch normalization and alignment, are combined with the graph topology matrix to construct graph structure data, thus obtaining the training graph dataset and the test graph dataset, respectively.

[0121] The training graph dataset is input into a factorized gated graph attention network, and the features of each node are normalized.

[0122] The input pre-gating module performs soft mask adaptive weighting and noise reduction, specifically as follows:

[0123] At the sensor node level, for each node's features, calculate the node weight vector:

[0124] ;

[0125] ;

[0126] In the formula, The feature matrix of each sensor node; This represents the node importance score vector; Node weights; This represents a multilayer perceptron used for node importance assessment; This represents the normalization function.

[0127] It is a learnable temperature parameter used to control the stringency of node selection.

[0128] In the feature dimension, to reduce the parameter complexity of generating a high-dimensional full feature weight matrix and effectively capture the coupling relationship between heterogeneous sensors and statistical indicators of various dimensions, global average pooling is performed on the features of each node to obtain a graph-level summary representation. The summary representation is then input into a multilayer perceptron to generate two latent representation matrices, and low-rank decomposition is used to generate the feature weight matrix.

[0129] ;

[0130] In the formula, Represented as a weight matrix along the feature dimension; Use the Sigmoid activation function; This represents the number of terms in a low-rank decomposition. and They represent the first Two latent representation vectors corresponding to each low-rank decomposition term. This is for outer product operations.

[0131] Final gated output Defined as residual form:

[0132] ;

[0133] In the formula, The input node feature matrix; For gating strength; This indicates element-wise multiplication.

[0134] The refined features are then adaptively calculated using a graph attention network to determine the attention weights between connected nodes, enabling information aggregation and state updates between nodes. The specific steps are as follows:

[0135] Nodes in the diagram The input feature vector is The node features are updated through a graph attention mechanism, utilizing a learnable linear transformation matrix. Map the node features.

[0136] According to the node Its set of neighboring nodes Attention weights are assigned based on the degree of correlation between nodes, and the features of neighboring nodes are aggregated in a weighted manner.

[0137] A multi-head attention mechanism is introduced to concatenate the aggregation results of multiple attention heads to obtain nodes. Update features .

[0138] In inter-layer transmission, to overcome the feature dilution phenomenon that easily occurs during information transmission, a latent gating module is used for secondary feature verification.

[0139] After input-side gating enhancement, a graph attention network is used to adaptively learn the topological dependencies and energy transfer patterns between cross-modal sensor nodes.

[0140] However, feature dilution can easily occur during information transmission, causing the weak fault characteristics of critical nodes to be masked by redundant information or macroscopic noise from neighboring nodes.

[0141] To address this, latent factorization gating is embedded between GAT convolutional layers. Unlike the physical feature selection at the input, this gating operates in the abstract high-dimensional space of the graph hidden layer, performing secondary calibration on node importance and feature channels by combining global graph awareness and dual residual mapping. The specific steps are as follows:

[0142] For the hidden layer feature matrix Calculate the node attention vector:

[0143] ;

[0144] ;

[0145] In the formula, This represents the node importance score vector obtained from the hidden layer feature matrix; This represents a multilayer perceptron used for node importance assessment; For the number of nodes, For the hidden layer dimension.

[0146] This is the updated node attention weight vector calculated based on the hidden layer feature matrix; Represents the normalization function; The temperature parameter is learnable; This is the node attention scaling factor.

[0147] To uncover dependencies between feature dimensions, context pooling is used to extract node information from the current graph, thereby obtaining a macroscopic representation of the entire graph. ,Will Input gating network generates feature channel weights The formula is:

[0148] ;

[0149] ;

[0150] In the formula, Indicates the first Hidden feature vectors of each node; Use the Sigmoid activation function; This represents a multilayer perceptron used to generate feature channel weights.

[0151] To prevent gradient vanishing or weight extremes in deep networks, the model introduces residual terms with upper and lower bound constraints to dynamically scale nodes and feature channels respectively:

[0152] ;

[0153] ;

[0154] In the formula, and These represent the dynamic scaling factors for the node and channel dimensions, respectively. This is a truncation function used to restrict the input to a given range; For learnable gating strength.

[0155] The final refined characteristics output are: ;

[0156] in, This represents the output feature matrix after being jointly scaled by the node dimension and the feature channel dimension; The hidden layer feature matrix; This indicates element-wise multiplication.

[0157] The central macroscopic displacement node is used as the final graph-level representation, and the classification prediction logic value of the fault state is output.

[0158] After multi-layer graph attention aggregation and latent refinement, the model abandons the traditional global average pooling and directly extracts the output of macroscopic displacement nodes as the final representation of the state.

[0159] Because this node is located at the center of the heterogeneous topology, it naturally gathers the degradation information of the driving current and mechanical vibration nodes during the graph evolution process. It not only preserves the complete link of fault propagation, but also makes its feature vector have a global view and physical discrimination ability.

[0160] Define the index of the center displacement node in the graph topology as Then the final representation vector at the graph level Represented as:

[0161] ;

[0162] The final graph-level representation vector is input into the classifier, and after nonlinear activation and random deactivation, the classification prediction logistic value of the fault state is calculated. .

[0163] Predict logical values ​​based on classification Calculate the cross-entropy loss term using the predicted and actual fault labels. Obtain the cost penalty term by querying the physical cost matrix based on the correspondence between the predicted and actual labels.

[0164] The cost penalty term is calculated based on a preset physical cost matrix, and the rules for constructing the physical cost matrix are as follows:

[0165] The penalty cost is zero when the predicted label matches the true label.

[0166] When the predicted category and the true category belong to the same fault category and a misjudgment occurs, a basic penalty cost is imposed.

[0167] When a cross-domain misjudgment occurs between the predicted category and the actual category, which belong to different fault categories, the highest penalty cost is imposed.

[0168] For similar faults that exhibit progressive degradation, a penalty cost is set based on the degradation distance between the prediction stage and the actual stage, with the penalty cost for a misjudgment at a skipped stage being higher than the penalty cost for a misjudgment at an adjacent stage.

[0169] By combining the cross-entropy loss term with the cost penalty term, a degradation-aware cost-sensitive loss function is constructed, and the error cost is calculated.

[0170] The gradient is calculated based on the error cost, and the weight parameters of the factorized gated graph attention network are updated through the backpropagation algorithm until the error cost converges, thus obtaining the fault intelligent identification model.

[0171] The test image dataset is input into the fault intelligent identification model to obtain the final classification and identification results, thereby realizing intelligent fault identification of the elevator system. The details are as follows:

[0172] The test dataset is input into the trained fault intelligent identification model, and the predicted logical value vector is output through forward propagation.

[0173] Subsequently, the predicted probability distribution of each fault state category is calculated using the Softmax function to achieve intelligent fault identification of the elevator system.

[0174] Example 1: This example uses an elevator system test platform to collect multi-source heterogeneous sensor data to verify the effectiveness of the method.

[0175] The platform is equipped with three types of sensors: a three-phase current sensor at the drive motor to characterize electromagnetic torque fluctuations, a three-axis accelerometer at the door to capture minor mechanical impacts, and a laser displacement sensor to record the system's operating trajectory in real time.

[0176] In this embodiment, the system's fault data covers a total of 11 working modes: 1 healthy state; 7 jamming states of different severity, caused by slider wear, foreign objects in the door slot, etc., manifested as displacement hysteresis or nonlinear peak reduction; and 3 wire rope continuous degradation and strand breakage states, with 1-2 strands broken in the early stage, 3-4 strands broken in the middle stage, and 5-6 strands broken in the late stage.

[0177] Please see Figure 2 , Figure 3 and Figure 4 The system displays the vibration, current, and displacement signals collected by the platform during a specific continuous operating period.

[0178] Among them, characteristic vibration data reflects the mechanical impact and structural response of the door system during operation, current data characterizes the load changes and electromagnetic torque fluctuations of the drive motor, and displacement data depicts the trajectory changes and dynamic behavior of the door system during motion. These three types of signals jointly describe the operating characteristics from three levels: mechanical, electrical, and motion state.

[0179] Please see Figure 5 The diagram displays the confusion matrix of the diagnostic results from this invention. It also shows a comparison of diagnostic metrics (Predicted Label and True Label) for different network methods. The network proposed in this invention (based on Factorized Gating Graph Attention Network, FGGAT) achieves an optimal classification accuracy of 94.89%. Because this invention combines physical causality to construct a topological structure and further introduces a factorized dual-gating mechanism, it effectively overcomes the deficiency of traditional networks in mining spatial physical topological relationships. It effectively suppresses macroscopic kinematic noise and purifies weak continuous degradation features, achieving a significant improvement in recognition accuracy.

[0180] Please see Figure 6The model demonstrates excellent performance across four key metrics: macro precision (Macro-P), macro recall (Macro-R), macro average F1 score (Macro-F1), and Kappa coefficient. Particularly noteworthy is its macro-F1 score of 95.45% for progressive degradation faults with extremely similar features in a minority class, showcasing its ability to decouple from high-frequency, weak impacts.

[0181] Meanwhile, the high macro-recall rate of 95.87% significantly reduces the false negative rate of hidden faults, and the Kappa coefficient also verifies the statistical consistency of its predictions. In summary, the method described in this invention has significant engineering application value in multimodal feature extraction and industrial predictive maintenance.

[0182] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0183] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An intelligent elevator fault diagnosis method based on factorized gated graph attention network, characterized in that: The method includes: Step 1: Collect the three-phase drive current, three-axis mechanical vibration and displacement signals of the elevator system. Based on the displacement signals, perform slicing of the door operator reciprocating motion cycle to divide the data into initial training dataset and initial test dataset. Step 2: Process the initial training dataset and the initial test dataset respectively, use differential feature extraction to construct node features, construct electromechanical causal graph, and obtain training graph dataset and test graph dataset respectively; Step 3: Input the training graph dataset into the factorization-gated graph attention network, normalize the features of each node, input them into the pre-gating module for soft mask adaptive weighting and noise reduction, and then perform information aggregation and state update through the graph attention network. The steps for soft mask adaptive weighting and noise reduction in the input pre-gating module are as follows: At the sensor node level, for each node's features, calculate the node weight vector: ； ； In the formula, The feature matrix of each sensor node; This represents the node importance score vector; Node weights; This represents a multilayer perceptron used for node importance assessment; Represents the normalization function; The temperature parameter is learnable; In the feature dimension, global average pooling is performed on the features of each node to obtain a graph-level aggregated representation. The aggregated representation is then input into a multilayer perceptron to generate two latent representation matrices, and low-rank decomposition is used to generate the feature weight matrix. ； In the formula, Represented as a weight matrix along the feature dimension; Use the Sigmoid activation function; This represents the number of terms in a low-rank decomposition. and They represent the first Two latent representation vectors corresponding to each low-rank decomposition term. This is an outer product operation; Final gated output Defined as residual form: ； In the formula, The input node feature matrix; For gating strength; This indicates element-wise multiplication; In the inter-layer transmission, the latent gating module is used for secondary feature verification. The central macroscopic displacement node is used as the final graph-level representation. The classification prediction logic value of the fault state is output, the degradation perception cost sensitive loss function is constructed, the error cost is calculated by combining the classification prediction logic value and the real fault label, and the network parameters are updated through the backpropagation algorithm to obtain the fault intelligent identification model. The specific steps for performing secondary feature verification using the latent gating module are as follows: For the hidden layer feature matrix Calculate the node attention vector: ； ； In the formula, This represents the node importance score vector obtained from the hidden layer feature matrix; This represents a multilayer perceptron used for node importance assessment; For the number of nodes, For the hidden layer dimension; This is the updated node attention weight vector calculated based on the hidden layer feature matrix; Represents the normalization function; The temperature parameter is learnable; This is the node attention scaling factor; Context pooling is used to extract node information from the current graph to obtain a macroscopic representation of the entire graph. ,Will Input gating network generates feature channel weights The formula is: ； ； In the formula, Indicates the first Hidden feature vectors of each node; Use the Sigmoid activation function; This represents a multilayer perceptron used to generate feature channel weights; Introduce residual terms with upper and lower bound constraints to dynamically scale nodes and feature channels respectively: ； ； In the formula, and These represent the dynamic scaling factors for the node and channel dimensions, respectively. This is a truncation function used to restrict the input to a given range; For learnable gating strength; The final refined characteristics output are: ; in, This represents the output feature matrix after being jointly scaled by the node dimension and the feature channel dimension; The hidden layer feature matrix; This indicates element-wise multiplication; Step 4: Input the test image dataset into the fault intelligent identification model to obtain the final classification and identification results, thereby realizing the intelligent identification of faults in the elevator system.

2. The elevator fault intelligent diagnosis method based on factorized gated graph attention network according to claim 1, characterized in that: In step 1, for periodic slices, the extreme points in the displacement signal are detected to locate the limit position of the gantry crane's reciprocating motion, and the three-phase drive current and three-axis mechanical vibration signal within the slice period are synchronized.

3. The elevator fault intelligent diagnosis method based on factorized gated graph attention network according to claim 1, characterized in that: In step 2, the differential feature extraction is as follows: Multidimensional time-domain features are extracted from the three-phase drive current signal, including root mean square, average, peak-to-peak value, standard deviation, skewness, kurtosis, crest factor, waveform factor, logarithm of total energy, volatility, and DC offset. After extraction, these feature vectors are filled with zero values ​​to a uniform dimension. ; DC offset is defined as the ratio of the average absolute value of the signal to the root mean square value. Performing db4 wavelet basis analysis on triaxial mechanical vibration signals Layer wavelet packet decomposition Each sub-band is used to calculate the logarithmic energy of the coefficients of each sub-band, the energy difference between adjacent bands, the root mean square, peak-to-peak value, standard deviation, skewness, kurtosis, and higher-order surrogate indices of the global signal. Higher-order surrogate metrics include the logarithm of the crest factor, the logarithm of the impulse factor, the low-frequency energy ratio, and the kurtosis variation, which together form high-dimensional features. The displacement signal is divided into three stages based on velocity change characteristics: acceleration, constant speed, and deceleration. Local standard deviations are extracted for each stage, and the dynamic ratio of the standard deviations for the braking and constant speed stages is calculated. A second-order jerk index is obtained by performing two numerical differences on the displacement signal. The absolute maximum value and ratio of the jerk during the acceleration and deceleration stages, the global average absolute jerk, and the absolute ratio of the average jerk values ​​for both stages are calculated. These are combined with the global range, global standard deviation, and global root mean square to construct a multi-dimensional feature, which is then filled with zero values ​​to a unified dimension. ; After feature extraction, separate batch normalization networks are used to align the feature distribution for node features of different physical modes.

4. The intelligent elevator fault diagnosis method based on factorized gated graph attention network according to claim 3, characterized in that: Volatility Defined as the ratio of the average of the absolute values ​​of the first-order differences of the signal to the root mean square, the formula is: ； In the formula, This represents the number of sampling points within the slice period. For the first Current values ​​at each sampling point For the first Current values ​​at each sampling point To prevent extremely small constants with a denominator of zero; Logarithmic energy of each subband coefficient The calculation formula is: ； In the formula, Represented as the first The first sub-band Wavelet packet reconstruction coefficients, This represents the total number of coefficients for that frequency band. Logarithm of crest factor The calculation formula is: ； In the formula, The maximum absolute value of the signal. The root mean square of the signal.

5. The intelligent elevator fault diagnosis method based on factorized gated graph attention network according to claim 3, characterized in that: In step 2, the electromechanical cause-effect graph is constructed as follows: Establish fully connected edges between three-phase current nodes; establish fully connected edges between three-axis vibration nodes; Using macroscopic displacement nodes as the central nodes of the graph, construct cross-modal physical causal edges with all current and vibration nodes pointing unidirectionally to the central nodes, thus completing the construction of the graph topology matrix; The node feature matrices of the initial training dataset and the initial test dataset, after independent batch normalization and alignment, are combined with the graph topology matrix to construct graph structure data, thus obtaining the training graph dataset and the test graph dataset, respectively.

6. The intelligent elevator fault diagnosis method based on factorized gated graph attention network according to claim 1, characterized in that: In step 3, the specific steps for information aggregation and state updating using a graph attention network are as follows: Nodes in the diagram The input feature vector is The node features are updated through a graph attention mechanism, utilizing a learnable linear transformation matrix. Map node features; According to the node Its set of neighboring nodes Attention weights are assigned based on the degree of correlation between nodes, and the features of neighboring nodes are aggregated in a weighted manner. A multi-head attention mechanism is introduced to concatenate the aggregation results of multiple attention heads to obtain nodes. Update features .

7. The intelligent elevator fault diagnosis method based on factorized gated graph attention network according to claim 1, characterized in that: In step 3, the specific steps for using the central macroscopic displacement node as the final graph-level representation are as follows: Define the index of the center displacement node in the graph topology as The final representation vector at the graph level Represented as: ； The final graph-level representation vector is input into the classifier, and after nonlinear activation and random deactivation, the classification prediction logistic value of the fault state is calculated. .

8. The intelligent elevator fault diagnosis method based on factorized gated graph attention network according to claim 1, characterized in that: In step 3, the degradation-aware cost-sensitive loss function includes a cross-entropy loss term and a cost penalty term. The cost penalty term is calculated based on a preset physical cost matrix, and the construction rules for the physical cost matrix are as follows: The penalty cost is zero when the predicted label matches the true label. When the predicted category and the true category belong to the same fault category and a misjudgment occurs, a basic penalty cost is imposed. When a cross-domain misjudgment occurs between the predicted category and the actual category, which belong to different fault categories, the highest penalty cost is imposed. For similar faults that exhibit progressive degradation, a penalty cost is set based on the degradation distance between the prediction stage and the actual stage, with the penalty cost for misjudging a level that skips a stage being higher than the penalty cost for misjudging an adjacent stage.

Images