Train door system sub-health detection method based on inter-partition threshold value

Abstract

The application discloses a train door system sub-health detection method based on an inter-partition threshold value, which can realize real-time reading of operation data for train door system sub-health monitoring based on a trained model, and a potential representation generation network can extract progressive potential features through multi-stage nonlinear compression and normalization constraint, and then dynamically capture time sequence evolution through a time sequence feature updating function to accurately and robustly express the nature of time sequence data; the automatic quantization of sensor confidence is realized, comprehensive abnormal detection results can be provided according to the abnormal scores of each sensor, and the interference of sensor signals susceptible to external factors on the abnormal detection results is reduced; the threshold values of different stages are determined considering that the sub-health information contained in different stages is different, whether the loss of each stage exceeds the threshold value is respectively evaluated for sub-health detection, the sub-health information contained in each stage is amplified, and therefore the sub-health detection precision of the train door system is improved, and the false positive rate is reduced.

Description

Technical Field

[0001] This invention relates to the field of urban rail transit technology, specifically to a sub-health detection method for train door systems based on interval thresholds. Background Technology

[0002] Train door systems are critical components in urban rail transit and high-speed railway operations. Their operational status directly affects train safety and passenger travel experience. During long-term service, door systems are susceptible to performance degradation and potential malfunctions due to multiple factors, including mechanical wear, motor aging, environmental factors, and frequent opening and closing. Traditional door fault detection relies heavily on explicit failure signals or manual inspection. When the system is still in a sub-healthy stage, its early deterioration characteristics are often subtle and difficult to identify using traditional methods, leading to delayed and reactive maintenance. If repairs are only carried out after the system has completely failed, it will not only significantly increase maintenance costs but also bring huge safety risks. Against this backdrop, sub-health detection of train door systems has received increasing attention. By identifying the sub-healthy state of door systems, it is possible to effectively reduce the accident rate and significantly reduce unplanned maintenance costs.

[0003] In recent years, with the increasing demand for intelligent operation and maintenance of trains, identifying the sub-health status of the train door system before a fault occurs has become a crucial issue for ensuring operational safety and reducing maintenance costs. Common methods include sub-health detection based on signal features, data-driven health indicator modeling, and machine learning methods. However, these methods often require large amounts of labeled data or rely on complex model construction, resulting in high computational overhead, insufficient real-time performance, and limited robustness under different operating conditions. Among existing train door system status detection methods, threshold-based methods have received widespread attention due to their simplicity and strong real-time performance. However, existing threshold-based methods neglect the sub-health detection effects of different stages and different sensors. Given the varying contributions of different sensors to the monitoring tasks, in complex systems, not all sensor signals carry equal amounts of sub-health information. Some variables are highly susceptible to installation standards and operating conditions, and the sub-health information contained in each stage is not the same; it changes with the stage of motion. Existing methods often treat monitoring data as a whole and assign it the same confidence level for sub-health detection tasks, resulting in poor sub-health detection performance. Therefore, how to amplify the sub-health information contained in each sensor at each stage of motion and improve the detection accuracy of sub-health in train door systems has become a key technical problem that urgently needs to be solved in the sub-health detection of train door systems. Summary of the Invention

[0004] In view of this, the present invention provides a sub-health detection method for train door systems based on interval thresholds. It proposes a latent representation generation network trained using normal operation data, and divides the door opening and closing process into five stages. The consistency residual distribution of normal operation data from different sensors in different stages is read separately. Based on this, the door operation data is read in real time and fed into the trained latent representation generation network to calculate the consistency residual of each stage. Based on the judgment index, it is determined whether the door is in a sub-healthy state. This method can be applied to the sub-health detection scenario of train doors and solves the problem of high false judgment rate in the sub-health detection of existing methods.

[0005] To achieve the above objectives, the present invention adopts the following technical solution:

[0006] A sub-health detection method for train door systems based on interval thresholds was designed, including the following steps:

[0007] Step S1: Acquire monitoring signals indicating the normal operating status of the train door system. Monitoring signals This includes signals detected by three types of sensors: motor displacement, current, and speed.

[0008] Step S2: Obtain the first 70% of the monitoring signals under normal train door system conditions. The model is fed into a latent representation generation network for training. The model construction and training steps are as follows:

[0009] Step S2.1: Data preprocessing, signal length... Filter out the signal and reduce the signal length. However, data that is shorter than the target length is padded with edge values ​​until the target length is reached, ensuring that the data length is uniform.

[0010] Step S2.2: The preprocessed data is fed into the latent representation generation network for training. The overall architecture of this network is divided into two parts: the embedding end and the restoration end. The embedding end progressively compresses the input observation features through a multi-level nonlinear projector and extracts progressive latent representations under normalization constraints. Subsequently, the temporal evolution unit applies dynamic updates to the obtained latent representations to explicitly capture the evolutionary trajectory across time periods. Its form can be expressed as:

[0011]

[0012] in, These represent the multi-layer feature transformation weights at the embedding end, and * indicates a computational operation. It is a nonlinear function. For normalization operations, The temporal feature update function for the embedded end is defined by the following formula:

[0013]

[0014] in, For the preceding steps at time step The output, i.e. ; For the embedded temporal feature update function at time step The hidden state of the output; The weight matrix is ​​the input to the gate / candidate state at the embedding end; and These are the update coefficient and reset coefficient for the embedded end, respectively; The updated value is determined by the reset coefficient. Output of the previous time step Input at that moment Calculated; For element-wise multiplication, For the constraint function, It is the hyperbolic tangent function; The weight matrix from the hidden state at the embedding end to the gate / candidate state; For the bias term corresponding to the embedded end;

[0015] At the reconstruction end, the network first renormalizes the latent state using a temporal feature update function, and then gradually reconstructs the original observation domain from the embedding end through a multi-level nonlinear inverse projector and nonlinear mapping. The process is formalized as follows:

[0016]

[0017] in These represent the weights of the inverse transform of multi-layer features. It is a nonlinear function. For normalization operations, The time-series feature update function for the recovery end at time... The output, i.e. , The calculation formula is:

[0018]

[0019] in, For the embedded temporal feature update function at time step The output variable is the time-series feature update function at the restoration end at time step. Input; The time-series feature update function for the recovery end at time step The hidden state of the output; and The update coefficients and reset coefficients are the time series feature update function of the recovery end. The updated value is determined by the reset coefficient. Output of the previous time step Input at that moment Calculated; The weight matrix is ​​input to the gate / candidate state at the restoration end. This is the weight matrix from the hidden state to the gate / candidate state at the restoration end. This is the bias term corresponding to the recovery end;

[0020] Step S2.3: Calculate the consistency residual between the original data and the latent representation signal. The calculation formula is as follows:

[0021]

[0022] in, For the first One monitoring signal value, For the first The hidden representation signal value passed through the latent representation generation network, The number of data points;

[0023] Step S2.4: Continuously update parameters using the Adam optimizer;

[0024] Step S3: Calculate the consistency residual distribution of normal state data, determine the sub-health judgment threshold and parameter weights by interval. During the operation of the car door, the sub-health information contained in different stages is different. Directly using the whole process data for sub-health judgment will bury a lot of key information. Therefore, it is necessary to judge the sub-health threshold in stages. The steps for determining the judgment threshold and calculating the weights are as follows:

[0025] Step S3.1: Divide the motion process of the door system. The threshold calculated using the entire operation process is easy to ignore the small changes in a certain stage. Therefore, it is necessary to divide the motion process of the door system into intervals. The motion process of the door system can be divided into acceleration stage, constant speed stage, deceleration stage, low speed stage and after the door is in place.

[0026] Step S3.2: Based on the trained model, analyze the last 30% of the monitoring signals of the train door system under normal conditions. The average consistency residual of different sensors at each stage is calculated by feeding it into the trained latent representation generation network. and the standard deviation of the consistency residuals The remaining 30% of normal data is used to determine the consistency residual threshold based on the trained model, that is, to determine the fluctuation range of the consistency residual of normal data, which is used to subsequently determine whether anomalies have occurred, and the average consistency residual. and the standard deviation of the consistency residuals The calculation formula is:

[0027]

[0028]

[0029] in, and Indicates the first The first stage, the first The average consistency residuals and standard deviation of the consistency residuals for each parameter; Indicates the first The sample at the th The first stage, the first Consistency residuals of each parameter;

[0030] Step S3.3: Based on the calculated average consistency residual and the standard deviation of the consistency residuals Determine the sub-health assessment thresholds for different sensors at different stages. The calculation formula is:

[0031]

[0032] in, It is usually set to 3, which means that at least 99.73% of the normal data values ​​are within 3 standard deviations of the mean. A parameter with a larger standard deviation means that the corresponding sensor signal distribution is more disturbed and tends to have a negative impact on the sub-health detection results.

[0033] Step S3.4: Determine the weighting of different sensors at different stages. Since different sensors obtain different amounts of information, it is necessary to assess the influence of different parameters on sub-health detection so that comprehensive sub-health detection results can be provided based on the thresholds of different sensors. The weighting calculation formula is as follows:

[0034]

[0035] Step S4; Perform sub-health detection on the train door system based on interval thresholds for the real-time acquired signals. The detection steps are as follows:

[0036] Step S4.1: Based on the real-time acquired monitoring signals of the train door system The loss is calculated by feeding it into the trained latent representation generation network. The loss calculation formula is as follows:

[0037]

[0038] in For monitoring signals In the The first stage, the first Loss of each parameter, For monitoring signal number The sample at the th The first stage, the first The signal values ​​of each parameter, For the monitoring signal The sample at the th The first stage, the first The hidden representation signal values ​​of each parameter;

[0039] Step S4.2: Calculate the judgment index, the first The calculation formula for the indicators for each stage is as follows:

[0040]

[0041] in, For the first The first stage, the first The weights of each parameter; and For the first The first stage, the first The average consistency residuals and standard deviations of the consistency residuals for each parameter; when at any stage At that time, it was determined that the train door system was in a sub-healthy state.

[0042] Furthermore, the signal length mentioned in step S2.1 refers to the number of data points of the signal monitored by the door system.

[0043] Furthermore, the target length mentioned in step S2.1 is a length that can be set by the user. The target length is set to exceed the number of data points being monitored, as long as this condition is met. The target length is set to be greater than the longest data length among all monitored signals.

[0044] Furthermore, the edge value mentioned in step S2.1 refers to the last data of the monitored signal, with the aim of keeping all signals of the same length.

[0045] Furthermore, the nonlinear mapping mentioned in step S2.2 refers to ReLU, which is a nonlinear function.

[0046] Furthermore, the raw data mentioned in step S2.3 is the data input to the network, i.e., the preprocessed data, and the latent representation signal refers to the signal output by the network after passing through the latent representation generation network.

[0047] Furthermore, the trained model mentioned in step S3.1 refers to the latent representation generation network trained using the first 70% of normal data.

[0048] Furthermore, the steps described in S3.2 The term "stage" refers to the door movement stage, which is divided into five stages: acceleration stage, constant speed stage, deceleration stage, low speed stage, and final positioning stage. The values ​​are 1, 2, 3, 4, and 5, representing the five stages of the door system's movement. The parameters refer to the signals monitored by different sensors, including three types of parameters: motor displacement, current, and speed; The first sample refers to the 30% of normal data. One sample.

[0049] Furthermore, in step S3.4, e is a mathematical constant equal to 2.71828, which is called the base of the natural logarithm.

[0050] Furthermore, the steps described in S4.1 The term "stage" refers to the five stages of the door's movement process, namely, the acceleration stage, the constant speed stage, the deceleration stage, the low speed stage, and the final positioning stage. The parameters refer to the signals monitored by different sensors, including three types of parameters: motor displacement, current, and speed; The sample refers to the first sample in the real-time acquired signal. One sample.

[0051] This invention enables real-time monitoring of sub-health in train door systems based on trained models and operational data. It utilizes a latent representation generation network to extract essential information from observed data and determines thresholds for judgment within intervals to capture sub-health information, thereby improving the accuracy of sub-health detection in train door systems. The invention features the following characteristics: 1) The latent representation generation network extracts asymptotic latent features through multi-level nonlinear compression and normalization constraints, and then dynamically captures temporal evolution through a temporal feature update function, accurately and robustly representing the essence of temporal data; 2) It achieves automatic quantification of sensor confidence, providing comprehensive anomaly detection results based on the anomaly score of each sensor, helping to reduce interference from sensor signals susceptible to external factors; 3) Considering the different sub-health information contained in different stages, thresholds are determined for different stages, and sub-health detection is performed by assessing whether the loss in each stage exceeds the threshold, amplifying the sub-health information contained in each stage, thereby improving the accuracy of sub-health detection in train door systems and reducing the false alarm rate. Attached Figure Description

[0052] Figure 1 This is a flowchart of the method of the present invention;

[0053] Figure 2A schematic diagram of the framework for a sub-health detection method for train door systems based on interval thresholds;

[0054] Figure 3 This is a performance comparison between the invented method and existing prediction methods. Detailed Implementation

[0055] The present invention will now be described in further detail with reference to the accompanying drawings.

[0056] Reference Figure 1 and Figure 2 A sub-health detection method for train door systems based on interval thresholds includes the following steps:

[0057] Step S1: Acquire monitoring signals indicating the normal operating status of the train door system. Monitoring signals This includes signals detected by three types of sensors: motor displacement, current, and speed.

[0058] Step S2: Obtain the first 70% of the monitoring signals under normal train door system conditions. The model is fed into a latent representation generation network for training. The model construction and training steps are as follows:

[0059] Step S2.1: Data preprocessing, signal length... Filter out the signal and reduce the signal length. However, data that is shorter than the target length is padded with edge values ​​until the target length is reached, ensuring that the data length is uniform.

[0060] Step S2.2: The preprocessed data is fed into the latent representation generation network for training. The overall architecture of this network is divided into two parts: the embedding end and the restoration end. The embedding end progressively compresses the input observation features through a multi-level nonlinear projector and extracts progressive latent representations under normalization constraints. Subsequently, the temporal evolution unit applies dynamic updates to the obtained latent representations to explicitly capture the evolutionary trajectory across time periods. Its form can be expressed as:

[0061]

[0062] in, These represent the multi-layer feature transformation weights at the embedding end, and * indicates a computational operation. It is a nonlinear function. For normalization operations, The temporal feature update function for the embedded end is defined by the following formula:

[0063]

[0064] in, For the preceding steps at time step The output, i.e. ; For the embedded temporal feature update function at time step The hidden state of the output; The weight matrix is ​​the input to the gate / candidate state at the embedding end; and These are the update coefficient and reset coefficient for the embedded end, respectively; The updated value is determined by the reset coefficient. Output of the previous time step Input at that moment Calculated; For element-wise multiplication, For the constraint function, It is the hyperbolic tangent function; The weight matrix from the hidden state at the embedding end to the gate / candidate state; For the bias term corresponding to the embedded end;

[0065] At the reconstruction end, the network first renormalizes the latent state using a temporal feature update function, and then gradually reconstructs the original observation domain from the embedding end through a multi-level nonlinear inverse projector and nonlinear mapping. The process is formalized as follows:

[0066]

[0067] in These represent the weights of the inverse transform of multi-layer features. It is a nonlinear function. For normalization operations, The time-series feature update function for the recovery end at time... The output, i.e. , The calculation formula is:

[0068]

[0069] in, For the embedded temporal feature update function at time step The output variable is the time-series feature update function at the restoration end at time step. Input; The time-series feature update function for the recovery end at time step The hidden state of the output; and The update coefficients and reset coefficients are the time series feature update function of the recovery end. The updated value is determined by the reset coefficient. Output of the previous time step Input at that moment Calculated; The weight matrix is ​​input to the gate / candidate state at the restoration end. This is the weight matrix from the hidden state to the gate / candidate state at the restoration end. This is the bias term corresponding to the recovery end;

[0070] Step S2.3: Calculate the consistency residual between the original data and the latent representation signal. The calculation formula is as follows:

[0071]

[0072] in, For the first One monitoring signal value, For the first The hidden representation signal value passed through the latent representation generation network, The number of data points;

[0073] Step S2.4: Continuously update parameters using the Adam optimizer;

[0074] Step S3: Calculate the consistency residual distribution of normal state data, determine the sub-health judgment threshold and parameter weights by interval. During the operation of the car door, the sub-health information contained in different stages is different. Directly using the whole process data for sub-health judgment will bury a lot of key information. Therefore, it is necessary to judge the sub-health threshold in stages. The steps for determining the judgment threshold and calculating the weights are as follows:

[0075] Step S3.1: Divide the motion process of the door system. The threshold calculated using the entire operation process is easy to ignore the small changes in a certain stage. Therefore, it is necessary to divide the motion process of the door system into intervals. The motion process of the door system can be divided into acceleration stage, constant speed stage, deceleration stage, low speed stage and after the door is in place.

[0076] Step S3.2: Based on the trained model, analyze the last 30% of the monitoring signals of the train door system under normal conditions. The average consistency residual of different sensors at each stage is calculated by feeding it into the trained latent representation generation network. and the standard deviation of the consistency residuals The remaining 30% of normal data is used to determine the consistency residual threshold based on the trained model, that is, to determine the fluctuation range of the consistency residual of normal data, which is used to subsequently determine whether anomalies have occurred, and the average consistency residual. and the standard deviation of the consistency residuals The calculation formula is:

[0077]

[0078]

[0079] in, and Indicates the first The first stage, the first The average consistency residuals and standard deviation of the consistency residuals for each parameter; Indicates the first The sample at the th The first stage, the first Consistency residuals of each parameter;

[0080] Step S3.3: Based on the calculated average consistency residual and the standard deviation of the consistency residuals Determine the sub-health assessment thresholds for different sensors at different stages. The calculation formula is:

[0081]

[0082] in, It is usually set to 3, which means that at least 99.73% of the normal data values ​​are within 3 standard deviations of the mean. A parameter with a larger standard deviation means that the corresponding sensor signal distribution is more disturbed and tends to have a negative impact on the sub-health detection results.

[0083] Step S3.4: Determine the weighting of different sensors at different stages. Since different sensors obtain different amounts of information, it is necessary to assess the influence of different parameters on sub-health detection so that comprehensive sub-health detection results can be provided based on the thresholds of different sensors. The weighting calculation formula is as follows:

[0084]

[0085] Step S4; Perform sub-health detection on the train door system based on interval thresholds for the real-time acquired signals. The detection steps are as follows:

[0086] Step S4.1: Based on the real-time acquired monitoring signals of the train door system The loss is calculated by feeding it into the trained latent representation generation network. The loss calculation formula is as follows:

[0087]

[0088] in For monitoring signals In the The first stage, the first Loss of each parameter, For monitoring signal number The sample at the th The first stage, the first The signal values ​​of each parameter, For the monitoring signal The sample at the th The first stage, the first The hidden representation signal values ​​of each parameter;

[0089] Step S4.2: Calculate the judgment index, the first The calculation formula for the indicators for each stage is as follows:

[0090]

[0091] in, For the first The first stage, the first The weights of each parameter; and For the first The first stage, the first The average consistency residuals and standard deviations of the consistency residuals for each parameter; when at any stage At that time, it was determined that the train door system was in a sub-healthy state.

[0092] Furthermore, the signal length mentioned in step S2.1 refers to the number of data points of the signal monitored by the door system.

[0093] Furthermore, the target length mentioned in step S2.1 is a length that can be set by the user. The target length is set to exceed the number of data points being monitored, as long as this condition is met. The target length is set to be greater than the longest data length among all monitored signals.

[0094] Furthermore, the edge value mentioned in step S2.1 refers to the last data of the monitored signal, with the aim of keeping all signals of the same length.

[0095] Furthermore, the nonlinear mapping mentioned in step S2.2 refers to ReLU, which is a nonlinear function.

[0096] Furthermore, the raw data mentioned in step S2.3 is the data input to the network, i.e., the preprocessed data, and the latent representation signal refers to the signal output by the network after passing through the latent representation generation network.

[0097] Furthermore, the trained model mentioned in step S3.1 refers to the latent representation generation network trained using the first 70% of normal data.

[0098] Furthermore, the steps described in S3.2 The term "stage" refers to the door movement stage, which is divided into five stages: acceleration stage, constant speed stage, deceleration stage, low speed stage, and final positioning stage. The values ​​are 1, 2, 3, 4, and 5, representing the five stages of the door system's movement. The parameters refer to the signals monitored by different sensors, including three types of parameters: motor displacement, current, and speed; The first sample refers to the 30% of normal data. One sample.

[0099] Furthermore, in step S3.4, e is a mathematical constant equal to 2.71828, which is called the base of the natural logarithm.

[0100] Furthermore, the steps described in S4.1 The term "stage" refers to the five stages of the door's movement process, namely, the acceleration stage, the constant speed stage, the deceleration stage, the low speed stage, and the final positioning stage. The parameters refer to the signals monitored by different sensors, including three types of parameters: motor displacement, current, and speed; The sample refers to the first sample in the real-time acquired signal. One sample.

[0101] Example:

[0102] Taking the rail vehicle door system as an example, the effectiveness of the method of this invention was verified by collecting data during the operation of the door system through the controller. A newly commissioned train was selected, and under the premise of ensuring the normal state of all door systems, normal data generated during the opening and closing of the eight doors of the train were collected. Then, some typical anomalies were constructed on these door systems, and abnormal data were collected, including centering anomalies, V-shaped anomalies, sealing size anomalies, buffer head anomalies, global resistance anomalies, and local resistance anomalies. The collected data included motor displacement signals, motor speed signals, and motor current signals, with a sampling frequency of 100Hz. The data collection of one door opening process lasted for about 3.5 seconds, which included about 350 time steps. The experiment used the method proposed in this invention to detect the sub-health state of the door system and compared it with four sub-health state detection methods (KNN, LOF, PCA, and DBSCAN).

[0103] Table 1 Summary of the basic structure of latent representation generation networks

[0104] Network layer Parameter settings Input layer - Nonlinear projection layer #1 Kernel size: 1×3; Number of kernels: 8; Activation function: ReLU Nonlinear projection layer #2 Kernel size: 1×3; Number of kernels: 16; Activation function: ReLU Batch Normalization Layer #1 - Nonlinear projection layer #3 Kernel size: 1×3; Number of kernels: 32; Activation function: ReLU Temporal Feature Update Layer (Embedded End) - Temporal Feature Update Layer (Restored End) - Nonlinear reverse projector #1 Kernel size: 1×3; Number of kernels: 16; Activation function: ReLU Batch Normalization Layer #2 - Nonlinear reverse projector #2 Kernel size: 1×3; Number of kernels: 8; Activation function: ReLU Nonlinear reverse projector #3 Kernel size: 1×3; Number of kernels: 3; Activation function: ReLU

[0105] Table 2 Summary of Hyperparameters

[0106] Parameter name set up Number of training iterations 100 Optimizer Adam Initial learning rate 0.001 Weight decay 1e-5 Batch size 8

[0107] Table 3 Performance Comparison of Five Methods

[0108] KNN LOF PCA DBSCAN This invention accuracy 90.24 88.46 85.63 91.47 96.25 F1 87.69 85.37 80.71 88.38 92.54

[0109] The basic structure of the latent representation generation network proposed in this paper is shown in Table 1, and the hyperparameter settings are shown in Tables 2 and 3. Figure 3 The accuracy and F1 score of the five methods are compared. The experimental results show that the sub-health detection method for train door systems based on interval threshold proposed in this invention achieves better sub-health detection results, and the detection accuracy is significantly better than the other four existing methods.

Claims

1. A method for detecting sub-health conditions in train door systems based on interval thresholds, characterized in that, Includes the following steps: Step S1: Acquire monitoring signals indicating the normal operating status of the train door system. Monitoring signals This includes signals detected by three types of sensors: motor displacement, current, and speed. Step S2: Obtain the first 70% of the monitoring signals under normal train door system conditions. It is fed into a latent representation generation network for training; Step S3: Calculate the consistency residual distribution of normal state data, determine the sub-health judgment threshold and parameter allocation weights by interval. During the operation of the car door, the sub-health information contained in different stages is different. Directly using the full process data to judge sub-health will drown out a lot of key information. Therefore, it is necessary to judge the sub-health threshold by stage. Step S4: Perform sub-health detection of the train door system by dividing the real-time acquired signals into interval thresholds; Step S2 specifically includes model construction and training, and the specific steps are as follows: Step S2.1: Data preprocessing, signal length... Filter out the signal and reduce the signal length. However, data shorter than the target length is padded with edge values ​​until the target length is reached, ensuring uniform data length. Step S2.2: The preprocessed data is fed into the latent representation generation network for training. The overall architecture of this network is divided into two parts: the embedding end and the restoration end. The embedding end progressively compresses the input observation features through a multi-level nonlinear projector and extracts progressive latent representations under normalization constraints. Subsequently, the temporal evolution unit applies dynamic updates to the obtained latent representations to explicitly capture the evolutionary trajectory across time periods. Its form can be expressed as: in, These represent the multi-layer feature transformation weights at the embedding end, and * indicates a computational operation. It is a nonlinear function. For normalization operations, The temporal feature update function for the embedded end is defined by the following formula: in, For the preceding steps at time step The output, i.e. ; For the embedded temporal feature update function at time step The hidden state of the output; The weight matrix is ​​the input to the gate / candidate state at the embedding end; and These are the update coefficient and reset coefficient for the embedded end, respectively; The updated value is determined by the reset coefficient. Output of the previous time step Input at that moment Calculated; For element-wise multiplication, For the constraint function, It is the hyperbolic tangent function; The weight matrix from the hidden state at the embedding end to the gate / candidate state; For the bias term corresponding to the embedded end; At the reconstruction end, the network first renormalizes the latent state using a temporal feature update function, and then gradually reconstructs the original observation domain from the embedding end through a multi-level nonlinear inverse projector and nonlinear mapping. The process is formalized as follows: in These represent the weights of the inverse transform of multi-layer features. It is a nonlinear function. For normalization operations, The time-series feature update function for the recovery end at time... The output, i.e. , The calculation formula is: in, For the embedded temporal feature update function at time step The output variable is the time-series feature update function at the restoration end at time step. Input; The time-series feature update function for the recovery end at time step The hidden state of the output; and The update coefficients and reset coefficients are the time series feature update function of the recovery end. The updated value is determined by the reset coefficient. Output of the previous time step Input at that moment Calculated; The weight matrix is ​​input to the gate / candidate state at the restoration end. This is the weight matrix from the hidden state to the gate / candidate state at the restoration end. This is the bias term corresponding to the recovery end; Step S2.3: Calculate the consistency residual between the original data and the latent representation signal. The calculation formula is as follows: in, For the first One monitoring signal value, For the first The hidden representation signal value passed through the latent representation generation network, The number of data points; Step S2.4: Continuously update parameters using the Adam optimizer.

2. The method for detecting sub-health conditions in a train door system based on interval thresholds according to claim 1, characterized in that, Step S3 specifically includes determining the threshold and calculating the weights, and the specific steps are as follows: Step S3.1: Divide the motion process of the door system. The threshold calculated using the entire operation process is easy to ignore the small changes in a certain stage. Therefore, it is necessary to divide the motion process of the door system into intervals. The motion process of the door system can be divided into acceleration stage, constant speed stage, deceleration stage, low speed stage and after the door is in place. Step S3.2: Based on the trained model, analyze the last 30% of the monitoring signals of the train door system under normal conditions. The average consistency residual of different sensors at each stage is calculated by feeding it into the trained latent representation generation network. and the standard deviation of the consistency residuals The remaining 30% of normal data is used to determine the consistency residual threshold based on the trained model, that is, to determine the fluctuation range of the consistency residual of normal data, which is used to subsequently determine whether anomalies have occurred, and the average consistency residual. and the standard deviation of the consistency residuals The calculation formula is: in, and Indicates the first The first stage, the first The average consistency residuals and standard deviation of the consistency residuals for each parameter; Indicates the first The sample at the th The first stage, the first Consistency residuals of each parameter; Step S3.3: Based on the calculated average consistency residual and the standard deviation of the consistency residuals Determine the sub-health assessment thresholds for different sensors at different stages. The calculation formula is: in, It is usually set to 3, which means that at least 99.73% of the normal data values ​​are within 3 standard deviations of the mean. A parameter with a larger standard deviation means that the corresponding sensor signal distribution is more disturbed and tends to have a negative impact on the sub-health detection results. Step S3.4: Determine the weighting of different sensors at different stages. Since different sensors obtain different amounts of information, it is necessary to assess the influence of different parameters on sub-health detection so that comprehensive sub-health detection results can be provided based on the thresholds of different sensors. The weighting calculation formula is as follows: The weight calculation results are obtained.

3. The method for detecting sub-health conditions in a train door system based on interval thresholds according to claim 1, characterized in that, The specific detection steps in step S4 are as follows: Step S4.1: Based on the real-time acquired monitoring signals of the train door system The loss is calculated by feeding it into the trained latent representation generation network. The loss calculation formula is as follows: in For monitoring signals In the The first stage, the first Loss of each parameter, For monitoring signal number The sample at the th The first stage, the first The signal values ​​of each parameter, For the monitoring signal The sample at the th The first stage, the first The hidden representation signal values ​​of each parameter; Step S4.2: Calculate the judgment index, the first The calculation formula for the indicators for each stage is as follows: in, For the first The first stage, the first The weights of each parameter; and For the first The first stage, the first The average consistency residuals and standard deviations of the consistency residuals for each parameter; when at any stage At that time, it was determined that the train door system was in a sub-healthy state.

4. The method for detecting sub-health conditions in a train door system based on interval thresholds according to claim 1, characterized in that, The signal length mentioned in step S2.1 refers to the number of data points of the signal monitored by the door system.

5. The method for detecting sub-health conditions in a train door system based on interval thresholds according to claim 1, characterized in that, The target length mentioned in step S2.1 is a length that can be set by the user. The target length is set to exceed the number of data points being monitored. As long as this condition is met, the target length should be set to be greater than the longest data length among all monitored signals.

6. The method for detecting sub-health conditions in a train door system based on interval thresholds according to claim 1, characterized in that, The edge value mentioned in step S2.1 refers to the last data of the monitored signal, with the aim of keeping all signals of the same length.

7. The method for detecting sub-health conditions in a train door system based on interval thresholds according to claim 1, characterized in that, The nonlinear mapping mentioned in step S2.2 refers to ReLU, which is a nonlinear function.

8. The method for detecting sub-health conditions in a train door system based on interval thresholds according to claim 1, characterized in that, The raw data mentioned in step S2.3 is the data input to the network, that is, the preprocessed data. The latent representation signal refers to the signal output by the network after passing through the latent representation generation network.

9. The method for detecting sub-health conditions in a train door system based on interval thresholds according to claim 2, characterized in that, The trained model mentioned in step S3.1 refers to the latent representation generation network trained using the first 70% of normal data.

10. A method for detecting sub-health conditions in a train door system based on interval thresholds according to claim 2, characterized in that, The steps described in step S3.2 The term "stage" refers to the door movement stage, which is divided into five stages: acceleration stage, constant speed stage, deceleration stage, low speed stage, and final positioning stage. The values ​​are 1, 2, 3, 4, and 5, representing the five stages of the door system's movement. The parameters refer to the signals monitored by different sensors, including three types of parameters: motor displacement, current, and speed; The first sample refers to the 30% of normal data. One sample.

11. The method for detecting sub-health conditions in a train door system based on interval thresholds according to claim 2, characterized in that, In step S3.4, e is a mathematical constant, approximately equal to 2.71828, and is called the base of the natural logarithm.

12. The method for detecting sub-health conditions in a train door system based on interval thresholds according to claim 3, characterized in that, The steps described in step S4.1 The term "stage" refers to the five stages of the door's movement process, namely, the acceleration stage, the constant speed stage, the deceleration stage, the low speed stage, and the final positioning stage. The parameters refer to the signals monitored by different sensors, including three types of parameters: motor displacement, current, and speed; The sample refers to the first sample in the real-time acquired signal. One sample.

Images