A method and system for detecting anomalies in rail transit door systems

By employing decoupling and adversarial training methods, a multi-task loss function is constructed, which improves the anomaly detection accuracy and generalization capability of the rail transit gate system. This solves the problems of high false alarm rate and insufficient model generalization in existing technologies, and achieves stable detection under complex operating conditions.

CN122796596BActive Publication Date: 2026-11-10NANJING KANGNI MECHANICAL & ELECTRICAL
View PDF 2 Cites 0 Cited by

Patent Information

Application Number
CN202611241691.3
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2026-08-17
Publication Date
2026-11-10
Estimated Expiration
2046-08-17

AI Technical Summary

Technical Problem

Existing methods for detecting anomalies in rail transit door systems have a high false alarm rate under load changes, mechanical wear, and environmental interference. Furthermore, they rely on a large number of well-labeled anomaly samples, resulting in insufficient model generalization ability and making it difficult to achieve real-time monitoring and effective detection of complex faults.

Method used

By collecting multi-source time-series operational data, auxiliary classification labels are constructed, decoupling the sub-features related to operating conditions and the sub-features of equipment status, and adversarial training is performed. Combining intra-class distribution preservation loss and random mask reconstruction, a multi-task joint loss function is constructed for anomaly detection.

Benefits of technology

It improves the accuracy and generalization ability of anomaly detection in the gate system, reduces the dependence on abnormal training samples, and enhances the robustness of the model and its ability to detect minor anomalies.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN122796596B_ABST
    Figure CN122796596B_ABST
Patent Text Reader

Abstract

This invention discloses a method and system for anomaly detection in rail transit door systems. The method includes: preprocessing multi-source time-series operational data and labeling it with auxiliary classification tags, including equipment identity tags and operating condition tags; generating a feature matrix containing location awareness; decoupling the encoded feature representation into operating condition-related sub-features and equipment status sub-features, and performing adversarial training using operating condition tags as supervision signals; jointly pre-training the encoder and classifier using equipment identity tags as supervision signals to construct an intra-class distribution preservation loss; constructing a multi-task joint loss function based on intra-class distribution preservation loss, classification loss, and reconstruction loss, calculating the reconstruction error and classification confidence of the test sample, weighting and fusing them into a comprehensive anomaly score, and determining whether there is an anomaly in the door system. This invention improves the accuracy and robustness of anomaly detection in door systems under complex operating conditions.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of rail transit technology, specifically to a method and system for detecting anomalies in rail transit door systems. Background Technology

[0002] Rail transit door systems (such as platform screen doors) are critical equipment for ensuring operational safety and improving traffic efficiency; their reliability directly affects passenger safety and system stability. Currently, the status monitoring and fault detection of door systems mainly rely on traditional methods, including sensor signal judgment based on fixed thresholds, periodic manual inspections, and spectrum analysis based on vibration or sound signals. However, these methods have significant limitations in practical applications: threshold methods cannot adapt to signal fluctuations caused by load changes, mechanical wear, and environmental interference, resulting in a high false alarm rate; manual inspections are inefficient, costly, and difficult to implement in real time; analysis methods based on a single sensing dimension (such as current or sound only) often fail to fully utilize the multi-physics coupling information of the system, limiting their ability to detect complex faults or early latent faults.

[0003] In recent years, with the application of machine learning technology in industrial monitoring, data-driven fault diagnosis methods have emerged, such as using support vector machines and Gaussian mixture models to classify historical fault samples. However, these methods typically rely on a large number of well-labeled anomaly samples for supervised training. In real-world industrial scenarios, anomaly events are rare, fault modes are diverse and difficult to fully cover, resulting in insufficient generalization ability of the models in practical deployments. Furthermore, existing methods mostly focus on classification after a fault occurs, lacking in-depth modeling of the normal operating mechanism of the system, and failing to effectively integrate the dynamic correlation between multi-source time-series signals. Therefore, under complex conditions such as changing operating conditions and environmental noise interference, the detection stability and interpretability remain unsatisfactory. Summary of the Invention

[0004] Purpose of the invention: This invention provides a method and system for detecting anomalies in rail transit door systems, thereby improving the accuracy and generalization ability of door system anomaly detection.

[0005] The abnormal detection method for rail transit door systems of the present invention includes: Collect multi-source time-series operational data generated by the gate system during normal operation, preprocess the multi-source time-series operational data and label it with auxiliary classification tags, the auxiliary classification tags including equipment identity tags and operating condition tags; Numerical features of labeled multi-source time-series running data are extracted and embedded with time-series location information to generate a feature matrix containing location awareness. The feature matrix is ​​input into the encoder. During the encoding process, the encoded feature representation is decoupled into operating condition-related sub-features and equipment status sub-features. The operating condition label is used as a supervision signal for adversarial training. The device state sub-features are input into the classifier, and the encoder and classifier are jointly pre-trained using the device identity label as the supervision signal. An intra-class distribution preservation loss is constructed in the space of the device state sub-features. After the joint pre-training converges, the underlying parameters of the encoder are fixed, and a random mask reconstruction decoder is connected. The device state sub-features are randomly masked to reconstruct reconstructed data with the same dimension as the multi-source time-series running data. The mean square error is used as the reconstruction loss. Based on the intra-class distribution preservation loss, classification loss, and reconstruction loss, a multi-task joint loss function is constructed to calculate the reconstruction error and classification confidence of the test sample. The reconstruction error and classification confidence are weighted and fused into a comprehensive anomaly score, and the presence of anomalies in the gate system is determined based on the comprehensive anomaly score.

[0006] Preferably, the step of preprocessing the multi-source time-series runtime data and labeling it with auxiliary classification tags further includes: The multi-source time-series running data is cleaned to remove invalid and outlier values, and aligned and synchronized based on a unified timestamp; Z-score standardization is performed on each aligned data channel, and a fixed-length sliding window is used to segment the continuous time series into time-series sample segments; Based on the time-series sample fragments, auxiliary classification labels are generated according to the inherent attributes of the data source gate and the operating context information at the time of data collection. The device identity label is generated based on the unique number of the data source gate, and the operating condition label is generated based on the preset operating condition at the time of data collection.

[0007] Preferably, the step of decoupling the operating condition-related sub-features and the equipment state sub-features further includes: A feature decoupling layer is provided at the output of the encoder. The feature decoupling layer includes a learnable first mapping matrix and a second mapping matrix, which respectively map the encoded feature representation to the working condition related sub-feature and the equipment state sub-feature. A condition discriminator is connected after the feature decoupling layer, and the output dimension of the condition discriminator is consistent with the total number of categories of the operating condition label. An adversarial gradient inversion layer is introduced on the connection path between the device state sub-feature and the operating condition discriminator. The adversarial gradient inversion layer outputs the original value of the device state sub-feature during forward propagation and multiplies the gradient from the operating condition discriminator by a preset negative coefficient before passing it to the second mapping matrix during backward propagation.

[0008] Preferably, the step of conducting the adversarial training further includes: With the goal of minimizing the first cross-entropy loss obtained by the working condition-related sub-features through the working condition discriminator, the first mapping matrix and the parameters of the working condition discriminator are updated so that the working condition-related sub-features are discriminative of the operating conditions. With the goal of maximizing the second cross-entropy loss obtained by the operating condition discriminator after the equipment state sub-features pass through the adversarial gradient inversion layer, the parameters of the second mapping matrix are updated so that the operating condition discriminator cannot identify the operating condition from the equipment state sub-features until the operating condition classification accuracy of the operating condition discriminator for the equipment state sub-features converges to below a preset threshold.

[0009] Preferably, the step of performing the joint pre-training further includes: The classification loss is calculated based on the device identity prediction probability distribution output by the classifier and the device identity label; the class center vector in the device state sub-feature space is calculated for all normal samples under the same device identity label; an intra-class distribution preservation loss is constructed, which is penalized based on the difference between the Euclidean distance of each sample to its class center and the intra-class average distance, and the loss is only generated when the distance of the sample to the class center is less than the product of a preset dispersion lower limit ratio coefficient and the intra-class average distance; and the weighted sum of the cross-entropy classification loss and the intra-class distribution preservation loss is used as the total loss of the joint pre-training.

[0010] Preferably, the step of reconstructing the reconstructed data further includes: randomly setting the device state sub-features to zero according to a preset ratio to generate masked features; and inputting the masked features into the random mask reconstruction decoder to output reconstructed data consistent with the dimension of multi-source time-series running data.

[0011] Preferably, the method further includes: if the comprehensive anomaly score exceeds a preset threshold, then it is determined that the gate system has a fault; if the comprehensive anomaly score does not exceed the preset threshold, then it is determined that the gate system does not have a fault.

[0012] The rail transit door system anomaly detection system of the present invention includes: The data acquisition and labeling unit is used to acquire multi-source time-series operation data generated by the gate system during normal operation, preprocess the multi-source time-series operation data and label it with auxiliary classification tags, the auxiliary classification tags including equipment identity tags and operating condition tags; The feature extraction unit is used to extract numerical features from labeled multi-source time-series running data and embed time-series location information to generate a feature matrix containing location awareness. The encoding decoupling unit is used to input the feature matrix into the encoder, decouple the encoded feature representation into operating condition related sub-features and equipment status sub-features during the encoding process, and use the operating condition label as a supervision signal for adversarial training. The classification pre-training unit is used to input the device state sub-features into the classifier, use the device identity label as the supervision signal to jointly pre-train the encoder and classifier, and construct the intra-class distribution preservation loss in the space of the device state sub-features. The mask-based reconstruction decoding unit is used to fix the low-level parameters of the encoder after the joint pre-training converges, connect to the random mask reconstruction decoder, perform random masking on the device state sub-features, reconstruct reconstructed data with the same dimension as the multi-source time-series running data, and use mean square error as the reconstruction loss. The fault diagnosis unit is used to construct a multi-task joint loss function based on the intra-class distribution preservation loss, classification loss, and reconstruction loss, calculate the reconstruction error and classification confidence of the test sample, weight and fuse the reconstruction error and classification confidence into a comprehensive anomaly score, and determine whether the gate system has a fault based on the comprehensive anomaly score.

[0013] Preferably, the coding decoupling unit further includes: A feature decoupling layer is provided at the output of the encoder. The feature decoupling layer includes a learnable first mapping matrix and a second mapping matrix, which respectively map the encoded feature representation to the working condition related sub-feature and the equipment state sub-feature. A condition discriminator is connected after the feature decoupling layer, and the output dimension of the condition discriminator is consistent with the total number of categories of the operating condition label. An adversarial gradient inversion layer is introduced on the connection path between the device state sub-feature and the operating condition discriminator. The adversarial gradient inversion layer outputs the original value of the device state sub-feature during forward propagation and multiplies the gradient from the operating condition discriminator by a preset negative coefficient before passing it to the second mapping matrix during backward propagation.

[0014] Preferably, the classification pre-training unit further includes: The classification loss is calculated based on the device identity prediction probability distribution output by the classifier and the device identity label; the class center vector in the device state sub-feature space is calculated for all normal samples under the same device identity label; an intra-class distribution preservation loss is constructed, which is penalized based on the difference between the Euclidean distance of each sample to its class center and the intra-class average distance, and the loss is only generated when the distance of the sample to the class center is less than the product of a preset dispersion lower limit ratio coefficient and the intra-class average distance; and the classification loss and the intra-class distribution preservation loss are weighted and summed as the total loss of the joint pre-training.

[0015] Compared with existing technologies, this invention has the following significant advantages: 1. By constructing auxiliary classification labels, it relies solely on normal samples for training, driving the model to learn the normal state representation of the data, reducing dependence on abnormal training samples, and enhancing the model's generalization and recognition capabilities; 2. By decoupling the encoded features into operating condition-related sub-features and equipment state sub-features, and introducing adversarial training to force anomaly judgment to be unaffected by normal operating condition fluctuations such as morning and evening peak hours and load changes, it solves the problem of false alarms caused by confusion between operating condition changes and faults; 3. By constructing intra-class distribution preservation loss during pre-training, it can constrain normal samples of the same equipment under different operating conditions to maintain appropriate dispersion in the feature space, reducing the difficulty of threshold selection; 4. By randomly masking the equipment state sub-features instead of surface memorizing the input, it significantly improves the detection capability of minor anomalies. Attached Figure Description

[0016] Figure 1 This is a schematic flowchart of an anomaly detection method for a rail transit door system provided in an embodiment of the present invention;

[0017] Figure 2 This is a first comparison diagram of the t-SNE dimensionality reduction distribution of device state features before and after feature decoupling provided in this embodiment of the invention;

[0018] Figure 3 This is a second comparison diagram of the t-SNE dimensionality reduction distribution of device state features before and after feature decoupling provided in the embodiments of the present invention;

[0019] Figure 4 This is a first comparison diagram of the feature space structure before and after the introduction of intra-class distribution preservation loss, provided in an embodiment of the present invention;

[0020] Figure 5 This is a second comparison diagram of the feature space structure before and after the introduction of intra-class distribution preservation loss, provided in an embodiment of the present invention;

[0021] Figure 6 This is a first comparison diagram of the reconstruction effect of the door opening speed channel under different mask ratios provided in the embodiments of the present invention;

[0022] Figure 7 This is a second comparison diagram of the reconstruction effect of the door opening speed channel under different mask ratios provided in the embodiments of the present invention;

[0023] Figure 8 This is the third comparison diagram of the reconstruction effect of the door opening speed channel under different mask ratios provided in the embodiments of the present invention. Detailed Implementation

[0024] The embodiments of the present invention will be further described in detail below with reference to the accompanying drawings.

[0025] This invention proposes a cascaded anomaly detection framework (Decouple-Delineate-Reconstruct, DDR). First, the framework decouples the encoded feature representation into condition-sensitive condition-related sub-features and condition-robust equipment state sub-features, thus removing condition interference at the feature level. Second, it introduces intra-class distribution preservation loss into the equipment state sub-feature space, constraining normal samples of the same equipment under different operating conditions to maintain a reasonable, non-overly tight dispersion in the feature space. This enhances the model's tolerance to individual equipment differences and normal aging, and provides a clear decision boundary for anomaly detection. Finally, the equipment state sub-features are randomly masked and reconstructed, forcing the model to learn deep correlations between normal data, rather than simple surface memory, significantly improving the detection sensitivity for minor anomalies and unknown faults. Through the synergistic effect of these three steps, this invention achieves accurate anomaly detection that is robust to changes in operating conditions and sensitive to equipment state.

[0026] Example 1

[0027] See Figure 1 This embodiment provides a method for detecting anomalies in a rail transit door system, including the following steps:

[0028] S1. Collect multi-source time-series operation data generated during the normal operation of the gate system, preprocess the multi-source time-series operation data and label it with auxiliary classification labels, including equipment identity labels and operating condition labels.

[0029] S101. Deploy sensors on the target rail transit gate system, connecting current sensors in series with the motor drive circuit to collect timing data of the motor operating current. A rotary encoder is installed on the motor output shaft to collect motor rotation timing data. Rotational speed time series data are obtained indirectly through differential calculation. , This represents a unit of time. All sensors are timestamped with high precision via a unified timing module to ensure that the data from each channel has a globally consistent timing reference. The data acquisition process should cover all known normal operating cycles of the gate system, including various states such as opening, closing, reopening, and standby.

[0030] S102. Input the collected raw multi-source time series running data into the automated preprocessing process to perform data cleaning, missing value imputation, and remove invalid and outlier values; perform alignment and synchronization based on a unified timestamp; perform Z-score standardization on each aligned data channel, and use a fixed-length sliding window to divide the continuous time series into time series sample segments.

[0031] The preprocessed multi-source time series runtime data input samples are Where B represents the batch size, W represents the number of windows, and L represents the subsequence length. Each time series sample segment can be represented as... w represents the window channel index, and l represents the subsequence index.

[0032] S103. Based on the inherent attributes of the data source gate and the operational context during data acquisition, a set of multi-dimensional auxiliary classification labels is assigned to each normal time-series sample segment, constructing a supervision signal for self-supervised learning. These auxiliary classification labels include device identity labels. and operating condition labels .

[0033] Among them, device identification tags Device identification tag generated based on the unique number of the door. The total number of categories is C; Operating condition labels This is generated based on the preset operating conditions during data acquisition. The preset operating conditions include K categories such as normal operation, high load test, and low load test.

[0034] S2. Extract the numerical features of the labeled multi-source time-series running data and embed the time-series location information to obtain a location-aware feature matrix.

[0035] S201. Each time-series sample segment is fed into an independent feature extraction module for processing. The feature extraction module uses a one-dimensional convolutional neural network to perform end-to-end automated feature extraction from multi-source time-series data. The overall network architecture adopts a highly configurable one-dimensional convolutional stacked structure, supporting flexible adjustment of the following key hyperparameters: N convolutional layers and the number of output channels in each layer within each feature extraction module. kernel size With adjustable filling strategies (including avg, lstm, max, etc.), adaptive extraction of local patterns at different scales can be achieved.

[0036] Each convolutional layer uses the ReLU activation function to perform a non-linear transformation on the convolution result to enhance the model's ability to fit complex working conditions. The convolution operation can be represented as: ;

[0037] in, and This represents the learnable weights and biases of the convolutional layer (containing multiple convolutional kernels), where n represents the convolutional layer index. This represents a subsequence (i.e., a time-series sample segment). It's important to note that the number of convolutional kernels in each convolutional layer is different, and the corresponding bias terms are also different. The bias terms in the formula for each convolutional layer are determined during model training. Through this series of operations, the original high-dimensional time-series signal is transformed into a set of low-dimensional but information-dense local feature vectors. Multiple local feature vectors are combined to form a feature matrix. .

[0038] S202, Obtain the feature matrix Subsequently, temporal positional information is injected into it through rotational positional encoding. Rotational positional encoding fuses absolute and relative positional information into the complex domain space of the feature vector, thereby generating a position-aware feature matrix. In self-attention computation, the absolute position information of the query matrix and key matrix is ​​encoded into the attention computation through a specific rotation transformation operation before they participate in the dot product operation.

[0039] Specifically, index p at each position in the time-series sample segment (where p is the index of each position). Mapped to a rotation matrix This rotation matrix operates on the complex representation of the feature vectors, achieving a geometric rotation and injecting absolute positional information. When these query vectors and key vectors, now endowed with absolute positional information, are used to calculate similarity, the results will reflect the relative distance relationships between elements at different positions.

[0040] This embodiment generates a location-aware feature matrix for each data entry. This dual encoding mechanism not only encodes the semantic content of the feature patterns but also explicitly implies the absolute and relative positional information of the temporal sequence. This dual encoding mechanism can accurately model the spatiotemporal logical relationships of each stage (such as acceleration, constant speed, and deceleration) in the operation of the gate, providing a highly discriminative input representation for subsequent accurate differentiation of working conditions.

[0041] S3. Input the feature matrix with position awareness into the encoder. During the encoding process, decouple the encoded feature representation into operating condition related sub-features and equipment status sub-features, and use the operating condition label as the supervision signal for adversarial training.

[0042] S301, The feature matrix with location awareness capability Input encoder, output deep contextual feature representation It fully integrates the global dependencies of all time steps within the sequence.

[0043] S302. Set a feature decoupling layer at the output of the encoder. The feature decoupling layer contains a learnable first mapping matrix. Second mapping matrix The encoded features are represented respectively. Mapped to operating condition related sub-features and device status sub-features .

[0044] S303. Introduce an adversarial training mechanism to achieve feature decoupling, enabling device state sub-features to... It only includes discriminative information related to the equipment's health status and insensitive to changes in operating conditions. An operating condition discriminator is connected after the feature decoupling layer, and its output dimension is equal to the total number of operating condition categories, K.

[0045] S304. Apply opposite training objectives to the two paths:

[0046] In the working condition discrimination path, the working condition-related sub-features are... The data is directly fed into the operating condition discriminator, which outputs the predicted probability distribution. By minimizing its difference from the actual working condition label. First cross-entropy loss Training the discriminator and the first mapping matrix This makes the operating condition related sub-features It serves as an effective carrier of operating condition information.

[0047] First cross-entropy loss The expression is: ;

[0048] in, express k Actual values ​​of labels for each type of real working condition. express k Predicted values ​​for labels of actual working conditions for each category.

[0049] In the anti-forgetting path, in the device state sub-feature An adversarial gradient inversion layer is introduced between the device condition discriminator and the device condition analyzer. This layer incorporates device state features during forward propagation. The output is sent to the condition discriminator as is to obtain the condition prediction distribution based on state features. And calculate the second cross-entropy loss. .

[0050] Second cross-entropy loss The expression is: ;

[0051] in, This represents the distribution of operating condition predictions based on state characteristics.

[0052] During backpropagation, the adversarial gradient reversal layer will come from the second cross-entropy loss. The gradient is multiplied by a preset negative coefficient. Then, pass it back to the second mapping matrix. This operation makes the second mapping matrix... The parameters are geared towards maximizing the second cross-entropy loss. The direction of the update is changed, thereby erasing the information about the operating conditions at the feature level.

[0053] The two training paths described above are executed alternately until the condition discriminator can no longer identify the equipment state sub-features. The system accurately identifies operating conditions when its classification accuracy converges to below a low preset threshold. At this point, the equipment status sub-features... It is no longer sensitive to changes in operating conditions.

[0054] S4. Input the device state sub-features into the classifier, use the device identity label as the supervision signal to jointly pre-train the encoder and classifier, and construct the intra-class distribution preservation loss in the space of the device state sub-features.

[0055] S401. Based on feature decoupling, the device state sub-features are... Access classifier. Based on device identity tags. To supervise the signal, the encoder, feature decoupling layer, and classifier are jointly pre-trained.

[0056] Calculate the classification loss of the classifier : ;

[0057] Where C represents the total number of device identity label categories, Indicates identity label c The true probability distribution of the devices, This indicates that the identity label output by the classifier is c The predicted probability distribution of the device.

[0058] S402. To enhance the model's robustness to individual differences in door structures and prevent false alarms caused by normal drift such as slow aging during long-term operation, an intra-class distributed preservation loss is introduced. This is used to constrain all normal samples of the same equipment under different operating conditions and operating periods, in the equipment state sub-feature. The feature space maintains a preset, lower-limited dispersion, meaning that normal samples are not allowed to excessively cluster into an extremely narrow point, thus leaving room for tolerance of slight, permissible state fluctuations.

[0059] Specifically, for identity tags c For a given device, calculate the device state sub-features of all its normal samples. Class center vectors in space : ;

[0060] Where i represents the index of the complex pair, This indicates the number of samples corresponding to the device's identity. Represents device state sub-features The i-th complex pair in the equation.

[0061] Calculate the Euclidean distance from each sample to its class center. .

[0062] And calculate the intra-class average distance for all samples in that category. .

[0063] Constructing intra-class distributed preservation loss : ;

[0064] in, This represents the preset lower limit of dispersion ratio coefficient. Indicates the average distance within the class. Represents Euclidean distance. This represents the number of samples corresponding to the device's identity. The mechanism for preserving the distribution within this class is: only when a sample's Euclidean distance to the center is... Less than the preset minimum limit Only when this happens will a penalty be imposed, pushing the sample outward to prevent excessive shrinkage of the intra-class distribution.

[0065] The classification loss of the classifier Preservation loss with intra-class distribution Weighted summation yields the total loss of the joint pre-training. : ;

[0066] in, This represents the preset weighting coefficients, used to minimize the total loss. With the goal of updating relevant network parameters, we can ultimately obtain a device status feature representation that is insensitive to operating conditions and can identify the individual device.

[0067] S5. After the joint pre-training converges, fix the underlying parameters of the encoder, connect the random mask reconstruction decoder, perform random masking on the device state sub-features, and reconstruct complete reconstructed data with the same dimension as the original multi-source time-series running data.

[0068] S501, Equipment Status Sub-features Perform a random mask. Randomly select a portion of its elements according to a preset ratio p and set them to zero, generating a corrupted feature. This process can be represented as: ;

[0069] in, To be related to the device state sub-features Binary mask matrices of the same shape, whose elements take the value 0 with probability p and the value 1 with probability 1-p. This indicates element-wise multiplication.

[0070] S502, masked features Input to the reconstruction decoder, and finally output the reconstruction data. The reconstructed data dimensions are completely consistent with the original input multi-source time-series running data X.

[0071] S503, Use mean squared error as the reconstruction loss for training: ;

[0072] Where N represents the total number of elements in the original multi-source time-series runtime data, and X represents the original input multi-source time-series runtime data. This indicates that the data has been reconstructed.

[0073] S6. Construct a multi-task joint loss function based on intra-class distribution preservation loss, classification loss, and reconstruction loss. Calculate the reconstruction error and classification confidence of the test sample respectively, and then weight and fuse them into a comprehensive anomaly score. If the comprehensive anomaly score exceeds a preset threshold, it is determined that the gate system has an anomaly.

[0074] S601. By organically combining the aforementioned losses, the network is uniformly fine-tuned and optimized, and then a comprehensive anomaly determination is made.

[0075] Multi-task joint loss function for: ;

[0076] in, Let be the weight coefficients for each task, and satisfy . ; Indicates the reconstruction loss. Represents classification loss. This represents the loss for preserving the distribution within the class. It is achieved by minimizing... The model is able to achieve high-precision reconstruction of normal data while maintaining a good structure in the feature space.

[0077] S602. After model training is complete, the online detection phase begins. For each test sample, the model will simultaneously output its reconstruction error. and classification confidence Classification confidence Take the identity prediction probability distribution output by the classifier The maximum value in is expressed as: ;

[0078] Reconstruction error and classification confidence Perform weighted fusion to calculate the comprehensive anomaly score. Its expression is: ;

[0079] in, This indicates the preset fusion weights. When equipment malfunctions, its state characteristics deviate from the normal cluster corresponding to its identity, leading to a significant increase in reconstruction error. Simultaneously, the classifier's confidence in its original identity drops sharply. When the overall anomaly score... When the preset fault determination threshold is exceeded, the system determines that there is a fault in the door system and triggers an alarm.

[0080] Example 2

[0081] This embodiment provides an anomaly detection system for rail transit door systems, including:

[0082] The data acquisition and labeling unit is used to collect multi-source time-series operational data generated during the normal operation of the gate system, preprocess the multi-source time-series operational data and label it with auxiliary classification tags, including equipment identity tags and operating condition tags.

[0083] Specifically, the data acquisition and labeling unit cleans the collected multi-source time-series operational data, removing invalid and outlier values. Then, based on a unified timestamp, the cleaned data is aligned and synchronized to ensure that data from different sensor channels have a consistent time baseline, thus providing accurate temporal correspondences for subsequent feature extraction. Based on the inherent attributes of the data source gate, such as its unique identifier, device identification tags are generated. Simultaneously, based on the preset operating conditions during data acquisition, such as normal operation, high-load testing, or low-load testing, operating condition tags are generated. These tags provide supervisory signals for subsequent adversarial training and classification pre-training, enabling the model to distinguish between different individual devices and different operating conditions.

[0084] The feature extraction unit is used to extract numerical features from the labeled multi-source time-series running data and embed time-series location information to generate a feature matrix containing location awareness.

[0085] Specifically, the feature extraction unit receives multi-source time-series operational data labeled by the data acquisition and labeling unit, and performs numerical feature extraction and temporal position information embedding on it. In some embodiments, a one-dimensional convolutional neural network structure can be used to extract local patterns from the input multi-channel time-series data through multiple convolutional layers, thereby obtaining a highly discriminative numerical feature representation. Simultaneously, to preserve the sequential information in the time-series data, temporal position information is also embedded, for example, by using rotational position encoding or learnable position embedding to inject absolute or relative position information into the feature matrix. Thus, the output feature matrix not only contains the local numerical features of the sensor signals but also explicitly encodes the relative relationships between time steps, thereby possessing positional awareness capabilities and better capturing the dynamic changes at each stage of the gate's operation.

[0086] The encoding decoupling unit is used to input the feature matrix into the encoder, decouple the encoded feature representation into operating condition related sub-features and equipment status sub-features during the encoding process, and use the operating condition label as a supervision signal for adversarial training. Specifically, the encoding-decoupling unit includes an encoder, a feature decoupling layer, a condition discriminator, and an adversarial gradient inversion layer. The encoder further encodes the feature matrix output by the feature extraction unit into a deep contextual feature representation. The feature decoupling layer, located at the encoder's output, contains a learnable first mapping matrix and a second mapping matrix. The first mapping matrix maps the encoded feature representation to condition-related sub-features, which primarily contain information related to the operating condition. The second mapping matrix maps the encoded feature representation to equipment status sub-features, which aim to retain discriminative information related to the equipment's health status. The condition discriminator, connected after the feature decoupling layer, has an output dimension consistent with the total number of categories for the operating condition labels, and is used to classify the condition-related sub-features. The adversarial gradient inversion layer is introduced into the connection path between the equipment status sub-features and the condition discriminator. During forward propagation, it outputs the original value; during backward propagation, it multiplies the gradient from the condition discriminator by a preset negative coefficient and passes it to the second mapping matrix, thereby achieving adversarial training and forcing the equipment status sub-features to be insensitive to changes in operating conditions.

[0087] The classification pre-training unit is used to input device state sub-features into the classifier, use device identity labels as supervision signals to jointly pre-train the encoder and classifier, and construct the intra-class distribution preservation loss in the space of device state sub-features.

[0088] Specifically, the classification pre-training unit takes the device state sub-features output by the encoding decoupling unit as input and outputs a device identity prediction probability distribution. The classification loss is calculated based on the device identity prediction probability distribution output by the classifier and the actual device identity labels, and used to train the encoder and classifier to accurately identify different door devices. For all normal samples under the same device identity label, its class center vector in the device state sub-feature space is calculated, and an intra-class distribution preservation loss is constructed. This loss constrains normal samples of the same device under different operating conditions to maintain a preset dispersion range in the feature space, preventing excessive aggregation of normal samples, thus providing a tolerance margin for fluctuations in normal device states and reducing the false alarm rate.

[0089] The masked reconstruction decoding unit is used to fix the low-level parameters of the encoder after the joint pre-training converges, and connect to the random masked reconstruction decoder. After performing random masking on the device state sub-features, it reconstructs the reconstructed data with the same dimension as the multi-source time-series running data, and uses the mean square error as the reconstruction loss.

[0090] Specifically, after the joint pre-training converges, the encoder's underlying parameters are fixed, and the device state sub-features are randomly set to zero according to a preset ratio to generate masked features. For example, 15% of the elements in the device state sub-features can be randomly selected and set to zero, thereby destroying some feature information. Based on the incomplete masked features, a series of decoding operations are used to reconstruct complete reconstructed data with dimensions consistent with the original multi-source time-series running data. This random masking reconstruction mechanism forces the model to learn the deep correlation patterns between sensor channels in normal data, rather than simply memorizing the input, thereby improving the ability to detect minor anomalies.

[0091] The fault diagnosis unit is used to construct a multi-task joint loss function based on intra-class distribution preservation loss, classification loss, and reconstruction loss, calculate the reconstruction error and classification confidence of the test sample, and weight and fuse the reconstruction error and classification confidence into a comprehensive anomaly score. Based on the comprehensive anomaly score, it is determined whether there is a fault in the gate system.

[0092] Specifically, the multi-task joint loss function combines intra-class distribution preservation loss, classification loss, and reconstruction loss, and sums them using preset weighting coefficients. During training, the goal is to minimize this multi-task joint loss function, simultaneously optimizing reconstruction accuracy and classification accuracy. This allows the model to maintain a good structure in the feature space while achieving high-precision reconstruction capabilities for normal data. After model training, the reconstruction error and classification confidence are calculated for each test sample, where the classification confidence is the maximum value in the device identity prediction probability distribution output by the classifier. The reconstruction error and classification confidence are weighted and fused into a comprehensive anomaly score. When this comprehensive anomaly score exceeds a preset fault judgment threshold, a fault is determined in the gate system, triggering a corresponding alarm or maintenance prompt.

[0093] Example 3

[0094] This embodiment aims to verify the effectiveness of the method of the present invention in forcing the device state sub-features to be insensitive to changes in operating conditions through encoder synchronization decoupling and adversarial training.

[0095] Three gate systems—Line 1 (143A+139B), Line 2 (106A+191B), and Line 3 (105A+97B)—of a certain subway line in actual operation were selected as data collection objects. Multi-source time-series operational data were collected for each gate system under both Condition A (normal operation) and Condition B (high-load testing), including drive motor current, gate vibration amplitude, and door opening / closing travel time. The collected data underwent cleaning, timestamp alignment, and sliding window slicing. Each sample segment was T=64 time steps long, with an overlap rate of 21.88% between adjacent windows. A total of 781 samples were collected from the three gate systems under the two conditions, and corresponding auxiliary classification labels were added according to equipment identity and operating condition.

[0096] The contrast model does not introduce feature decoupling or adversarial training; it directly uses the encoder's original encoded features. As a feature of equipment status, the model of this invention synchronously decouples the encoded features into operating condition-related sub-features. and device status sub-features Adversarial training is conducted using operating condition labels as supervisory signals, and hyperparameters are used. λ Set to 0.2. Input the device state features extracted from the two models into the t-SNE dimensionality reduction algorithm, map them to a two-dimensional space, and obtain a feature distribution comparison map.

[0097] Figure 2 It can be clearly observed that the original coding features contain a large amount of operating condition-related information, and the normal state representation of the equipment is interfered with by operating condition fluctuations. Figure 3 It can be seen that the feature points of the same gate under the two working conditions A and B are highly clustered according to the sub-features of the gate equipment, and no longer exhibit the phenomenon of mixed distribution affected by the working conditions.

[0098] The above results demonstrate that this invention successfully extracts device state sub-features through feature decoupling and adversarial training. The system decouples operational condition information, making it less sensitive to changes in operating conditions and retaining only discriminative features related to the equipment's health status. This lays the feature foundation for stable, low-false-alarm anomaly detection under varying operating conditions.

[0099] Example 4

[0100] This embodiment aims to verify the method of the present invention, which uses device state sub-features. When constructing a distribution-aware classification pre-training for input, the effectiveness of maintaining the loss constraint through intra-class distribution is similar to maintaining a reasonable dispersion of normal samples in the feature space.

[0101] Using the samples collected in Example 3, the comparison configuration was set to use only the classification loss. The encoder and classifier are pre-trained without introducing an intra-class distribution preservation loss term; this embodiment configures the classification loss... Based on this, an intra-class distributed preservation loss is introduced. Together, they constitute the total loss of joint pre-training. Among them, the dispersion lower limit proportional coefficient Set to 0.3, weighting coefficient Set to 0.01.

[0102] To quantitatively assess the structural quality of the feature space, the intra-class average distance is introduced to reflect the clustering degree of samples of the same class: for each device, the vector from all its samples to the class center is calculated. The mean of the Euclidean distances is calculated, and then averaged over all devices. The separation is introduced as the ratio of inter-class dispersion to intra-class dispersion. Here, inter-class dispersion is the sum of squared distances from each class center to the global center, and intra-class dispersion is the sum of squared distances from each sample to its own class center.

[0103] A comparison of the feature space structure before and after introducing the intra-class distribution preservation loss is presented. Figure 4 As can be seen, in the comparison configuration, the average intra-class distance is 7.91, and the average inter-class distance is 28.01. There is a significant overlap in the probability densities of intra-class and inter-class distances, resulting in a separation of only 3.54. At this point, the feature distribution is relatively loose, and samples from different devices are too close together in the feature space, easily leading to device identity confusion. Introducing intra-class distribution preservation loss... Figure 5 It can be seen that the peak probability density of intra-class distance shifts significantly to the left and the distribution narrows, the average intra-class distance decreases to 2.08, the clustering of samples of the same type of device is greatly enhanced, the separation degree increases from 3.54 to 5.08, the overlapping area between clusters is reduced, and the structure of the feature space is substantially optimized.

[0104] The results show that by introducing intra-class distribution preservation loss, this invention can effectively compress intra-class dispersion and optimize inter-class separation while maintaining reasonable dispersion of samples from the same device, thus constructing a device state sub-feature space with a better structure and providing a high-quality feature distribution state for subsequent anomaly detection.

[0105] Example 5

[0106] This embodiment aims to verify the feasibility and effectiveness of the method of the present invention, which involves connecting a random mask reconstruction decoder after joint pre-training convergence, performing random masking on the device state sub-features, and having the decoder reconstruct the complete original data.

[0107] According to the method provided by this invention, after the joint pre-training converges, the low-level parameters of the encoder are fixed, and a random mask reconstruction decoder is connected. This involves the device state sub-features. All elements are randomly set to zero according to a preset ratio p, and the features after masking are obtained. The features after masking Input random mask reconstruction decoder, output reconstructed data The reconstructed data dimensions are consistent with the original multi-source time-series runtime data X.

[0108] To evaluate the impact of different mask ratios on reconstruction accuracy, three mask ratios (p=15%, p=30%, and p=40%) were set for experimental comparison. For ease of visualization, the door opening speed, a representative sensor channel, was selected, and comparison graphs of the original and reconstructed curves under different mask ratios were plotted.

[0109] like Figure 6 As shown, when the mask ratio p=15%, the reconstructed curve closely matches the original curve. The reconstructed value accurately tracks the trend and amplitude of the true value at each stage of speed change, with only slight deviations in local details. This indicates that even with only 15% of the device state sub-feature information lost, the random mask reconstruction decoder can fully reconstruct the complete original multi-source time-series data based on the remaining 85% of the feature information. When the mask ratio is increased to 30%, as... Figure 7 As shown, the reconstructed curve generally follows the trend of the original curve, but the fitting deviation increases in the region of sudden changes in motor speed, the high-speed range, and the braking range. Some high-frequency fluctuation details are lost, and the reconstruction accuracy decreases to some extent. When the mask ratio is further increased to 40%, as... Figure 8 As shown, although the reconstructed curve still maintains the ability to track the main trend of the original curve, the deviation in amplitude and phase is further increased, especially the reconstruction error at the signal inflection point is significantly increased, and some detailed features can no longer be accurately restored.

[0110] The results show that as the mask ratio increases, the reconstruction accuracy gradually decreases, but at a reasonable mask ratio (e.g., 15%), the model still maintains high reconstruction accuracy. This verifies that the technical solution of performing random masking at the device state sub-feature level followed by reconstruction by the decoder is feasible.

[0111] Example 6

[0112] This embodiment aims to quantitatively evaluate the independent contribution of the proposed cascaded anomaly detection framework to the final detection performance through ablation experiments, and to verify the comprehensive detection capability of the complete framework on a large-scale validation set.

[0113] The training data in this embodiment comes from actual operation data collected from a subway line, covering normal operation samples of multiple gate systems under various operating conditions to ensure that the model learns sufficient normal state patterns. The validation set is constructed independently from different time periods and different gates, containing 5766 normal samples and 231 abnormal samples, with an abnormal to normal ratio of approximately 1:25, which is consistent with the actual situation of rare abnormal samples in industrial scenarios.

[0114] Four progressive model configurations are set up: Configuration A uses only the basic autoencoder structure to directly reconstruct the original multi-source time series data, using the reconstruction error as the anomaly score; Configuration B introduces feature decoupling and adversarial training on the baseline; Configuration C further introduces intra-class distribution preservation loss. Configure D-access random mask reconstruction decoder for device state sub-features. After random masking, reconstruction is performed, and anomaly detection is conducted using a multi-task joint loss function and a comprehensive anomaly score. All configurations are trained on the same training set and evaluated on the same validation set. The masking ratio is uniformly set to 15%, and the optimal values ​​for hyperparameters such as fusion weights are determined through grid search on the validation set.

[0115] This embodiment uses four metrics—precision, recall, F2 score, and AUC-PR—to evaluate anomaly detection performance. The definitions of each metric are as follows: ;

[0116] Wherein, TP represents the number of fault samples correctly identified as abnormal, FP represents the number of normal samples incorrectly identified as abnormal, and FN represents the number of fault samples incorrectly identified as normal.

[0117] This is suitable for the extremely imbalanced scenario in this validation set, where the ratio of anomalous to normal classes is approximately 1:25. AUC-PR, the area under the precision-recall curve, is sensitive to class imbalance and can more realistically reflect the model's overall discriminative ability under imbalanced data.

[0118] The experimental results are shown in Table 1. It can be seen that the model shows a monotonically increasing improvement trend in all four indicators. Finally, the complete cascaded anomaly detection framework achieved an excellent level of false alarm rate of 1.09% and false negative rate of 1.73%.

[0119] Table 1 Experimental Results Model Configuration AUC-PR F2 score Accuracy Recall rate Configuration A 0.8500 0.7575 0.4419 0.9221 Configuration B 0.8813 0.7937 0.4762 0.9524 Configuration C 0.9407 0.9083 0.7450 0.9610 Configuration D 0.9621 0.9349 0.7828 0.9827

[0120] In the baseline model of configuration A, the AUC-PR was only 0.8500, and the F2 score was 0.7575. After introducing feature decoupling and adversarial training (configuration B), the AUC-PR improved to 0.8813, the F2 score improved to 0.7937, and the recall improved from 0.9221 to 0.9524, indicating that the ability of equipment state features to detect faults is enhanced after removing operating condition interference. Further introducing intra-class distribution preservation loss (configuration C), the AUC-PR jumped significantly to 0.9407, the F2 score reached 0.9083, the precision significantly improved from 0.4762 to 0.7450, and the recall continued to rise to 0.9610. This fully verifies the dual role of normalized feature space structure in suppressing false alarms and improving fault identification accuracy. Ultimately, after being integrated with the random mask reconstruction decoder, the complete cascaded anomaly detection framework (configuration D) of this invention achieved excellent performance with an AUC-PR of 0.9621, an F2 score of 0.9349, and a recall rate as high as 0.9827.

[0121] It is important to note that the precision values ​​for all configurations in this embodiment are generally lower than the recall values. This is due to the sample distribution characteristics of the validation set itself. The ratio of anomalous samples to normal samples in this validation set is approximately 1:25, representing extremely imbalanced data typical of industrial scenarios. Under imbalanced conditions, even if the model misclassifies a very small number of normal samples as anomalous, the sheer number of false positives (FPs) will be significant due to the large base of normal samples, thus lowering the precision metric. This phenomenon is prevalent in the field of anomaly detection and aligns with engineering realities: in actual operation of rail transit gate systems, normal operating time far exceeds fault time, resulting in a naturally scarce pool of anomalous samples. The validation set must reflect this true distribution to ensure the on-site representativeness of the evaluation results.

[0122] Although data imbalance inevitably suppresses precision, observing the complete results of configuration D shows that the precision of the proposed cascaded anomaly detection framework reaches 0.7828, an improvement of over 77% compared to 0.4419 in configuration A, still meeting the practical requirements for engineering deployment under extremely imbalanced data. Meanwhile, the recall rate is as high as 0.9827, indicating that only about 4 out of 231 anomalous samples were missed. The AUC-PR reaches 0.9621, which, as a comprehensive indicator based on the precision-recall curve, also fully demonstrates the excellent ranking and discrimination performance of the anomaly score in highly imbalanced scenarios.

[0123] In summary, despite the unavoidable class imbalance inherent in real industrial data, the cascaded anomaly detection framework proposed in this invention still comprehensively surpasses baseline methods in all key metrics, including precision, recall, F2 score, and AUC-PR. Ultimately, with an extremely low false negative rate and engineering-acceptable precision, it fully verifies the advanced nature of the proposed solution and its significant application value in anomaly detection of actual rail transit door systems.

Claims

1. A method for detecting anomalies in a rail transit door system, characterized in that, include: Collect multi-source time-series operational data generated by the gate system during normal operation, preprocess the multi-source time-series operational data and label it with auxiliary classification tags, the auxiliary classification tags including equipment identity tags and operating condition tags; Numerical features of labeled multi-source time-series running data are extracted and embedded with time-series location information to generate a feature matrix containing location awareness. The feature matrix is ​​input into the encoder. During the encoding process, the encoded feature representation is decoupled into operating condition-related sub-features and equipment status sub-features. The operating condition label is used as a supervision signal for adversarial training. The device state sub-features are input into the classifier, and the encoder and classifier are jointly pre-trained using the device identity label as the supervision signal. An intra-class distribution preservation loss is constructed in the space of the device state sub-features. After the joint pre-training converges, the underlying parameters of the encoder are fixed, and a random mask reconstruction decoder is connected. The device state sub-features are randomly masked to reconstruct reconstructed data with the same dimension as the multi-source time-series running data. The mean square error is used as the reconstruction loss. Based on the intra-class distribution preservation loss, classification loss, and reconstruction loss, a multi-task joint loss function is constructed to calculate the reconstruction error and classification confidence of the test sample. The reconstruction error and classification confidence are weighted and fused into a comprehensive anomaly score, and the presence of anomalies in the gate system is determined based on the comprehensive anomaly score.

2. The method according to claim 1, characterized in that, The steps of preprocessing the multi-source time-series runtime data and labeling it with auxiliary classification tags also include: The multi-source time-series running data is cleaned to remove invalid and outlier values, and aligned and synchronized based on a unified timestamp; Z-score standardization is performed on each aligned data channel, and a fixed-length sliding window is used to segment the continuous time series into time-series sample segments; Based on the time-series sample fragments, auxiliary classification labels are generated according to the inherent attributes of the data source gate and the operating context information at the time of data collection. The device identity label is generated based on the unique number of the data source gate, and the operating condition label is generated based on the preset operating condition at the time of data collection.

3. The method according to claim 1, characterized in that, The step of decoupling the operating condition-related sub-features and the equipment state sub-features further includes: A feature decoupling layer is provided at the output of the encoder. The feature decoupling layer includes a learnable first mapping matrix and a second mapping matrix, which respectively map the encoded feature representation to the working condition related sub-feature and the equipment state sub-feature. A condition discriminator is connected after the feature decoupling layer, and the output dimension of the condition discriminator is consistent with the total number of categories of the operating condition label. An adversarial gradient inversion layer is introduced on the connection path between the device state sub-feature and the operating condition discriminator. The adversarial gradient inversion layer outputs the original value of the device state sub-feature during forward propagation and multiplies the gradient from the operating condition discriminator by a preset negative coefficient before passing it to the second mapping matrix during backward propagation.

4. The method according to claim 1, characterized in that, The steps for conducting the adversarial training also include: With the goal of minimizing the first cross-entropy loss obtained by the working condition-related sub-features through the working condition discriminator, the first mapping matrix and the parameters of the working condition discriminator are updated so that the working condition-related sub-features are discriminative of the operating conditions. With the goal of maximizing the second cross-entropy loss obtained by the operating condition discriminator after the equipment state sub-features pass through the adversarial gradient inversion layer, the parameters of the second mapping matrix are updated so that the operating condition discriminator cannot identify the operating condition from the equipment state sub-features until the operating condition classification accuracy of the operating condition discriminator for the equipment state sub-features converges to below a preset threshold.

5. The method according to claim 1, characterized in that, The steps for performing the joint pre-training also include: The classification loss is calculated based on the device identity prediction probability distribution output by the classifier and the device identity label; the class center vector in the device state sub-feature space is calculated for all normal samples under the same device identity label; an intra-class distribution preservation loss is constructed, which is penalized based on the difference between the Euclidean distance of each sample to its class center and the intra-class average distance, and the loss is only generated when the distance of the sample to the class center is less than the product of a preset dispersion lower limit ratio coefficient and the intra-class average distance; and the classification loss and the intra-class distribution preservation loss are weighted and summed as the total loss of the joint pre-training.

6. The method according to claim 1, characterized in that, The steps for reconstructing the reconstructed data further include: randomly setting the device state sub-features to zero according to a preset ratio to generate masked features; and inputting the masked features into the random mask reconstruction decoder to output reconstructed data consistent with the dimensions of the multi-source time-series running data.

7. The method according to claim 1, characterized in that, The method further includes: if the overall anomaly score exceeds a preset threshold, then the gate system is determined to have a fault; if the overall anomaly score does not exceed the preset threshold, then the gate system is determined to have no fault.

8. An anomaly detection system for a rail transit door system, characterized in that, include: The data acquisition and labeling unit is used to acquire multi-source time-series operation data generated by the gate system during normal operation, preprocess the multi-source time-series operation data and label it with auxiliary classification tags, the auxiliary classification tags including equipment identity tags and operating condition tags; The feature extraction unit is used to extract numerical features from labeled multi-source time-series running data and embed time-series location information to generate a feature matrix containing location awareness. The encoding decoupling unit is used to input the feature matrix into the encoder, decouple the encoded feature representation into operating condition related sub-features and equipment status sub-features during the encoding process, and use the operating condition label as a supervision signal for adversarial training. The classification pre-training unit is used to input the device state sub-features into the classifier, use the device identity label as the supervision signal to jointly pre-train the encoder and classifier, and construct the intra-class distribution preservation loss in the space of the device state sub-features. The mask-based reconstruction decoding unit is used to fix the low-level parameters of the encoder after the joint pre-training converges, connect to the random mask reconstruction decoder, perform random masking on the device state sub-features, reconstruct reconstructed data with the same dimension as the multi-source time-series running data, and use mean square error as the reconstruction loss. The fault diagnosis unit is used to construct a multi-task joint loss function based on the intra-class distribution preservation loss, classification loss, and reconstruction loss, calculate the reconstruction error and classification confidence of the test sample, weight and fuse the reconstruction error and classification confidence into a comprehensive anomaly score, and determine whether the gate system has a fault based on the comprehensive anomaly score.

9. The system according to claim 8, characterized in that, The coding decoupling unit further includes: A feature decoupling layer is provided at the output of the encoder. The feature decoupling layer includes a learnable first mapping matrix and a second mapping matrix, which respectively map the encoded feature representation to the working condition related sub-feature and the equipment state sub-feature. A condition discriminator is connected after the feature decoupling layer, and the output dimension of the condition discriminator is consistent with the total number of categories of the operating condition label. An adversarial gradient inversion layer is introduced on the connection path between the device state sub-feature and the operating condition discriminator. The adversarial gradient inversion layer outputs the original value of the device state sub-feature during forward propagation and multiplies the gradient from the operating condition discriminator by a preset negative coefficient before passing it to the second mapping matrix during backward propagation.

10. The system according to claim 8, characterized in that, The classification pre-training unit also includes: The classification loss is calculated based on the device identity prediction probability distribution output by the classifier and the device identity label; the class center vector in the device state sub-feature space is calculated for all normal samples under the same device identity label; an intra-class distribution preservation loss is constructed, which is penalized based on the difference between the Euclidean distance of each sample to its class center and the intra-class average distance, and the loss is only generated when the distance of the sample to the class center is less than the product of a preset dispersion lower limit ratio coefficient and the intra-class average distance; and the classification loss and the intra-class distribution preservation loss are weighted and summed as the total loss of the joint pre-training.

Citation Information

Patent Citations

  • Deep learning based rail-transit platform-edge door fault diagnosis system

    CN110262463A

  • Method and system for predicting abrasion degree of mechanical part of rail transit vehicle door system

    CN118940117A