Metro sliding plug door fault detection method based on gram angle field and deep learning

Abstract

This invention discloses a fault detection method for subway sliding doors based on Gram angle field and deep learning, comprising the following steps: preprocessing the current signal of the subway sliding door drive motor; compressing the preprocessed current signal to obtain a compressed current signal; normalizing, transforming coordinates, and calculating the Gram matrix of the compressed current signal, and generating a two-dimensional image from the calculated two-dimensional matrix; and classifying and identifying faults in the two-dimensional image using a deep learning residual network. This method converts the original current data into a two-dimensional image after data compression, and uses a deep learning network to achieve real-time detection of faults such as lower stop pin displacement, pressure roller overpressure, V-shaped inversion faults, and V-shaped dimension positive exceedance faults in the subway door system. It eliminates the need for manual feature extraction, avoiding the difficulty of manual feature extraction in traditional machine learning algorithms, and enables automatic real-time fault detection to ensure train operation safety.

Description

Technical Field

[0001] This invention relates to the field of subway sliding door fault detection, specifically to a subway sliding door fault detection method based on Gram angle field and deep learning. Background Technology

[0002] Displacement of the lower stop pin, overpressure of the pressure roller, V-shaped inversion, and excessive positive V-shaped dimensions in the subway door system are typical faults in subway doors. Once these faults occur, they can cause inconvenience in opening and closing the doors, or even prevent the doors from opening and closing at all, posing a significant threat to subway operation safety and the safety of passengers' lives and property.

[0003] Currently, subway maintenance companies in China primarily rely on manual inspection to detect faults in subway sliding doors, such as displacement of the lower stop pin, overpressure of the pressure roller, V-shaped inversion, and excessive positive V-shaped dimensions. However, traditional manual inspection methods are time-consuming, labor-intensive, and dependent on experience. Furthermore, existing intelligent fault diagnosis technologies for subway sliding doors mainly rely on time-frequency analysis of the drive motor current signal to extract features for fault diagnosis. These methods require manual feature extraction, which presents challenges, and their accuracy heavily depends on the manually selected features, leading to unstable fault detection accuracy. Summary of the Invention

[0004] To address the aforementioned shortcomings in existing technologies, the subway sliding door fault detection method based on Gram angle field and deep learning provided by this invention solves the problems of existing subway sliding door fault detection methods relying on manual feature extraction and having unstable accuracy.

[0005] To achieve the above-mentioned objectives, the technical solution adopted by this invention is as follows: A method for detecting faults in subway sliding doors based on Gram angle fields and deep learning is provided, which includes the following steps: S1. Obtain the current signal of the subway sliding door drive motor and perform preprocessing to obtain the preprocessed current signal; S2. Compress the preprocessed current signal to obtain the compressed current signal; S3. Normalize, transform coordinates and calculate Gram matrix for the compressed current signal, and generate a two-dimensional image from the calculated two-dimensional matrix. S4. Fault classification and recognition of two-dimensional images are performed using deep learning residual networks to complete the fault detection of subway sliding doors.

[0006] Furthermore, the specific method of preprocessing in step S1 is as follows: using a Butterworth filter to suppress noise in the current signal of the subway sliding door drive motor.

[0007] Furthermore, the specific method of step S2 includes the following sub-steps: S2-1. Divide the time series of current of the subway sliding door drive motor into multiple fixed-length subsequence intervals by using a sliding window, and calculate the mean of the data in each subsequence interval. S2-2. Rearrange all the calculated mean values ​​into a new sequence according to time to obtain the compressed current signal.

[0008] Furthermore, the specific method of step S3 includes the following sub-steps: S3-1. Normalize the compressed current signal to obtain the normalized current signal; S3-2. Convert the time series of the normalized current signal into polar coordinates: Encode the values ​​in the normalized current signal using the cosine of the polar angle in polar coordinates; S3-3. Calculate the corresponding Gram angle and field from the polar coordinates obtained by the time series transformation of the normalized current signal, and obtain the corresponding Gram matrix. S3-4. Generate the corresponding two-dimensional image from the Gram matrix.

[0009] Furthermore, the expression encoded in step S3-2 is:

[0010] in This represents the first normalized current signal. i Value The corresponding polar angle is between [0, π / 2]; π is the mathematical constant pi.

[0011] Furthermore, the expression for the Gram matrix in step S3-3 is:

[0012] in Represents the Gram matrix; This represents the first normalized current signal. n The polar angle corresponding to each value.

[0013] Furthermore, subway sliding door malfunctions include displacement of the lower stop pin, overpressure of the pressure roller, V-shaped inversion, and V-shaped dimension exceeding the standard.

[0014] Furthermore, the deep learning residual network comprises, in sequence, an input layer, a zero-padding layer, a first convolutional layer, a first batch normalization layer, a first ReLU layer, a first max pooling layer, a convolutional module, a first Identity module, a second Identity module, a second convolutional layer, a second batch normalization layer, a second ReLU layer, a second max pooling layer, a third convolutional layer, a third batch normalization layer, a third ReLU layer, a third max pooling layer, a Flatten layer, a fully connected layer, and an output layer. The convolutional module includes a fourth convolutional unit, a fourth batch normalization unit, and a first convolutional unit, a first batch normalization unit, a first ReLU unit, a second convolutional unit, a second batch normalization unit, a second ReLU unit, a third convolutional unit, a third batch normalization unit, a first fusion unit, and a third ReLU unit connected in sequence. The input of the fourth convolutional unit and the input of the first convolutional unit together serve as the input of the convolutional module. The other inputs of the fourth convolutional unit, the fourth batch normalization unit, and the first fusion unit are connected in sequence. The output of the third ReLU unit is the output of the convolutional module. Each Identity module includes a fifth convolutional unit, a fifth batch normalization unit, a fourth ReLU unit, a sixth convolutional unit, a sixth batch normalization unit, a fifth ReLU unit, a seventh convolutional unit, a seventh batch normalization unit, a second fusion unit, and a sixth ReLU unit connected in sequence. The other input of the second fusion unit and the input of the fifth convolutional unit together serve as the input of the current Identity module. The output of the sixth ReLU unit serves as the output of the current Identity module.

[0015] Furthermore, the specific training method for deep learning residual networks is as follows: Obtain fault sample data for subway sliding door lower stop pin displacement, pressure roller overpressure, V-shaped inversion, and positive V-shaped dimension exceeding the standard, and obtain the corresponding two-dimensional images of the fault sample data using the same method as steps S1 to S3; use the fault type as the label, and use the corresponding two-dimensional images of the fault sample data as input to train the deep learning residual network. The learning rate is automatically adjusted during the training process using a cosine annealing algorithm until training is complete; the expression for the learning rate in each iteration is:

[0016] Let be the learning rate at the t-th iteration; The maximum learning rate; Minimum learning rate; This represents the total number of iterations. Pi; It is a cosine function.

[0017] Furthermore, the total number of iterations during training is 100, the maximum learning rate is 0.0001, the minimum learning rate is 0.000001, and the batch size is 4.

[0018] The beneficial effects of this invention are as follows: 1. This method compresses the raw current data and converts it into a two-dimensional image based on the Gram angle field. Based on the converted two-dimensional image, a deep learning network is used to realize the real-time detection of faults such as displacement of the lower stop pin, overpressure of the pressure roller, V-shaped inversion, and positive excess of V-shaped size in the subway door system. No manual feature extraction is required, which avoids the problem of difficulty in manual feature extraction in traditional machine learning algorithms. It can realize automatic real-time fault detection and ensure train operation safety.

[0019] 2. This method introduces a segmented aggregation approximation approach to compress the data while fully preserving the time series features, thereby ensuring the accuracy of classification and recognition.

[0020] 3. This method introduces the cosine annealing algorithm during the training phase to automatically adjust the learning rate, which speeds up the training process while preventing the network from getting stuck in local optima. Attached Figure Description

[0021] Figure 1 This is a flowchart illustrating the method. Figure 2 This is a schematic diagram of the structure of a deep learning residual network; Figure 3 The waveform of the preprocessed current signal of the sliding door drive motor is shown in the example. Figure 4 This is a two-dimensional image generated in a normal state in the embodiment; Figure 5 This is the confusion matrix of the test results in the example. Detailed Implementation

[0022] The specific embodiments of the present invention are described below to enable those skilled in the art to understand the present invention. However, it should be understood that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the present invention as defined and determined by the appended claims. All inventions utilizing the concept of the present invention are protected.

[0023] like Figure 1 As shown, the subway sliding door fault detection method based on Gram angle field and deep learning includes the following steps: S1. Obtain the current signal of the subway sliding door drive motor and perform preprocessing to obtain the preprocessed current signal; S2. Compress the preprocessed current signal to obtain the compressed current signal; S3. Normalize, transform coordinates and calculate Gram matrix for the compressed current signal, and generate a two-dimensional image from the calculated two-dimensional matrix. S4. Fault classification and recognition of two-dimensional images are performed using deep learning residual networks to complete the fault detection of subway sliding doors.

[0024] The specific method of preprocessing in step S1 is as follows: use a Butterworth filter to suppress noise in the current signal of the subway sliding door drive motor.

[0025] The specific method of step S2 includes the following sub-steps: S2-1. Divide the time series of current of the subway sliding door drive motor into multiple fixed-length subsequence intervals by using a sliding window, and calculate the mean of the data in each subsequence interval. S2-2. Rearrange all the calculated mean values ​​into a new sequence according to time to obtain the compressed current signal.

[0026] Step S2 can be represented by the following formula:

[0027] in For the first part of the compressed signal i One element; k It is a dimensionality reduction factor, and N = k × n ; N This is the total length of the current signal for the sliding door drive motor. For length is N The first in the current signal of the sliding door drive motor j One element; n This represents the length of the time series after dimensionality reduction.

[0028] The specific method of step S3 includes the following sub-steps: S3-1. Normalize the compressed current signal to obtain the normalized current signal; S3-2. Convert the time series of the normalized current signal into polar coordinates: Encode the values ​​in the normalized current signal using the cosine of the polar angle in polar coordinates; S3-3. Calculate the corresponding Gram angle and field from the polar coordinates obtained by the time series transformation of the normalized current signal, and obtain the corresponding Gram matrix. S3-4. Generate the corresponding two-dimensional image from the Gram matrix.

[0029] The expression encoded in step S3-2 is:

[0030] in This represents the first normalized current signal. i Value The corresponding polar angle is between [0, π / 2]; π is the mathematical constant pi.

[0031] The expression for the Gram matrix in step S3-3 is:

[0032] in Represents the Gram matrix; This represents the first normalized current signal. n The polar angle corresponding to each value.

[0033] Subway sliding door malfunctions include displacement of the lower stop pin, overpressure of the pressure roller, V-shaped inversion, and V-shaped dimension exceeding the standard.

[0034] like Figure 2 As shown, the deep learning residual network includes an input layer, a zero-padding layer, a first convolutional layer, a first batch normalization layer, a first ReLU layer, a first max pooling layer, a convolutional module, a first Identity module, a second Identity module, a second convolutional layer, a second batch normalization layer, a second ReLU layer, a second max pooling layer, a third convolutional layer, a third batch normalization layer, a third ReLU layer, a third max pooling layer, a Flatten layer, a fully connected layer, and an output layer, all connected in sequence. The convolutional module includes a fourth convolutional unit, a fourth batch normalization unit, and a first convolutional unit, a first batch normalization unit, a first ReLU unit, a second convolutional unit, a second batch normalization unit, a second ReLU unit, a third convolutional unit, a third batch normalization unit, a first fusion unit, and a third ReLU unit connected in sequence. The input of the fourth convolutional unit and the input of the first convolutional unit together serve as the input of the convolutional module. The other inputs of the fourth convolutional unit, the fourth batch normalization unit, and the first fusion unit are connected in sequence. The output of the third ReLU unit is the output of the convolutional module. Each Identity module includes a fifth convolutional unit, a fifth batch normalization unit, a fourth ReLU unit, a sixth convolutional unit, a sixth batch normalization unit, a fifth ReLU unit, a seventh convolutional unit, a seventh batch normalization unit, a second fusion unit, and a sixth ReLU unit connected in sequence. The other input of the second fusion unit and the input of the fifth convolutional unit together serve as the input of the current Identity module. The output of the sixth ReLU unit serves as the output of the current Identity module.

[0035] The specific training method for deep learning residual networks is as follows: Obtain fault sample data for subway sliding door lower stop pin displacement, pressure roller overpressure, V-shaped inversion, and positive V-shaped dimension exceeding the standard, and obtain the corresponding two-dimensional images of the fault sample data using the same method as steps S1 to S3; use the fault type as the label, and use the corresponding two-dimensional images of the fault sample data as input to train the deep learning residual network. The learning rate is automatically adjusted during the training process using a cosine annealing algorithm until training is complete; the expression for the learning rate in each iteration is:

[0036] Let be the learning rate at the t-th iteration; The maximum learning rate; Minimum learning rate; This represents the total number of iterations. Pi; It is a cosine function.

[0037] The total number of iterations during training was 100, the maximum learning rate was 0.0001, the minimum learning rate was 0.000001, and the batch size was 4.

[0038] In one embodiment of the present invention, the data types and quantities used in the examples are shown in Table 1.

[0039] Table 1: Experimental Data

[0040] The actual current value of the slug door drive motor was processed using a Butterworth filter bandpass filter to remove noise from the original signal. The waveform of the slug door drive motor current signal after noise reduction and other preprocessing is shown below. Figure 3 As shown.

[0041] The pre-processed current signal of the sliding door drive motor X ={ x 1, x 2, ... x N} divided into lengths of k The data is divided into segments of 10, and the mean of each segment is calculated to obtain the approximate data. X '={ x 1', x 2', ... x n '}.

[0042] The current signal of the subway sliding door drive motor is normalized and scaled to the range [0,1]. The normalization formula is as follows:

[0043] In the formula x'' The normalized value is min( x' ) represents the minimum value of the motor current signal before normalization, max( x' () represents the maximum value of the motor current signal before normalization, in A.

[0044] The normalized current time series is converted into polar coordinates. The normalized current values ​​are then encoded using the cosine of the polar angle, and their corresponding timestamps are encoded as polar coordinate radii. The encoding formula is as follows:

[0045] In the formula, φ i for x'' The corresponding polar angle in polar coordinates, r i for x'' The corresponding polar radius. It should be noted that although this embodiment also calculates the polar radius, the polar radius is only one parameter of the polar coordinate system, and this method does not use the polar radius parameter.

[0046] The Gram angle and field (GASF) of the current signal after polar coordinate transformation are calculated to obtain the corresponding Gram matrix, and a two-dimensional image is generated. The generated two-dimensional image corresponding to the normal state time series is shown below. Figure 4 As shown in the figure. The trained deep learning residual network was tested using actual current data from the drive motors of subway sliding doors. The confusion matrix of the test results is shown in the figure. Figure 5 As shown.

[0047] Depend on Figure 5 It can be seen that the detection accuracy of this method reaches 100% under normal and under the fault conditions of lower pin displacement, pressure roller overpressure, V-shaped dimension positive excess, and V-shaped inversion, achieving the optimal result. From the above results, it can be seen that this method is effective and feasible for dealing with the faults of lower pin displacement, pressure roller overpressure, V-shaped dimension positive excess, and V-shaped inversion in subway sliding doors.

[0048] In summary, this invention utilizes Gram angle field to convert one-dimensional time-series signals into two-dimensional images. Compared to other methods for converting one-dimensional signals into two-dimensional images, it has advantages such as low computational complexity and strong feature representation capabilities. After conversion to two-dimensional images, this invention uses deep learning residual networks for classification and recognition, eliminating the need for manual feature extraction. This enables automatic real-time detection of these faults, ensuring driving safety during operation and demonstrating significant application value.

Claims

1. A method for detecting faults in subway sliding doors based on Gram angle fields and deep learning, characterized in that, Includes the following steps: S1. Obtain the current signal of the subway sliding door drive motor and perform preprocessing to obtain the preprocessed current signal; S2. Compress the preprocessed current signal to obtain the compressed current signal; S3. Normalize, transform coordinates and calculate Gram matrix for the compressed current signal, and generate a two-dimensional image from the calculated two-dimensional matrix. S4. Fault classification and recognition of two-dimensional images are performed using deep learning residual networks to complete the fault detection of subway sliding doors; The specific method of step S3 includes the following sub-steps: S3-1. Normalize the compressed current signal to obtain the normalized current signal; S3-2. Convert the time series of the normalized current signal into polar coordinates: Encode the values ​​in the normalized current signal using the cosine of the polar angle in polar coordinates; S3-3. Calculate the corresponding Gram angle and field from the polar coordinates obtained by the time series transformation of the normalized current signal, and obtain the corresponding Gram matrix. S3-4. Generate the corresponding two-dimensional image from the Gram matrix; The expression encoded in step S3-2 is: in This represents the first normalized current signal. i Value The corresponding polar angle has a value between [0, π / 2]; π is the mathematical constant pi. The expression for the Gram matrix in step S3-3 is: in Represents the Gram matrix; This represents the first normalized current signal. n The polar angles corresponding to each value; Subway sliding door malfunctions include displacement of the lower stop pin, overpressure of the pressure roller, V-shaped inversion, and V-shaped dimension exceeding the standard.

2. The subway sliding door fault detection method based on Gram angle field and deep learning according to claim 1, characterized in that, The specific method of preprocessing in step S1 is as follows: use a Butterworth filter to suppress noise in the current signal of the subway sliding door drive motor.

3. The subway sliding door fault detection method based on Gram angle field and deep learning according to claim 1, characterized in that, The specific method of step S2 includes the following sub-steps: S2-1. Divide the time series of current of the subway sliding door drive motor into multiple fixed-length subsequence intervals by using a sliding window, and calculate the mean of the data in each subsequence interval. S2-2. Rearrange all the calculated mean values ​​into a new sequence according to time to obtain the compressed current signal.

4. The subway sliding door fault detection method based on Gram angle field and deep learning according to claim 1, characterized in that, A deep learning residual network consists of an input layer, a zero-padding layer, a first convolutional layer, a first batch normalization layer, a first ReLU layer, a first max pooling layer, a convolutional module, a first Identity module, a second Identity module, a second convolutional layer, a second batch normalization layer, a second ReLU layer, a second max pooling layer, a third convolutional layer, a third batch normalization layer, a third ReLU layer, a third max pooling layer, a Flatten layer, a fully connected layer, and an output layer, all connected in sequence. The convolutional module includes a fourth convolutional unit, a fourth batch normalization unit, and a first convolutional unit, a first batch normalization unit, a first ReLU unit, a second convolutional unit, a second batch normalization unit, a second ReLU unit, a third convolutional unit, a third batch normalization unit, a first fusion unit, and a third ReLU unit connected in sequence. The input of the fourth convolutional unit and the input of the first convolutional unit together serve as the input of the convolutional module. The other inputs of the fourth convolutional unit, the fourth batch normalization unit, and the first fusion unit are connected in sequence. The output of the third ReLU unit is the output of the convolutional module. Each Identity module includes a fifth convolutional unit, a fifth batch normalization unit, a fourth ReLU unit, a sixth convolutional unit, a sixth batch normalization unit, a fifth ReLU unit, a seventh convolutional unit, a seventh batch normalization unit, a second fusion unit, and a sixth ReLU unit connected in sequence. The other input of the second fusion unit and the input of the fifth convolutional unit together serve as the input of the current Identity module. The output of the sixth ReLU unit serves as the output of the current Identity module.

5. The subway sliding door fault detection method based on Gram angle field and deep learning according to claim 4, characterized in that, The specific training method for deep learning residual networks is as follows: Obtain fault sample data for subway sliding door lower stop pin displacement, pressure roller overpressure, V-shaped inversion, and positive V-shaped dimension exceeding the standard, and obtain the corresponding two-dimensional images of the fault sample data using the same method as steps S1 to S3; use the fault type as the label, and use the corresponding two-dimensional images of the fault sample data as input to train the deep learning residual network. The learning rate is automatically adjusted during the training process using a cosine annealing algorithm until training is complete; the expression for the learning rate in each iteration is: Let be the learning rate at the t-th iteration; The maximum learning rate; Minimum learning rate; This represents the total number of iterations. Pi; It is a cosine function.

6. The subway sliding door fault detection method based on Gram angle field and deep learning according to claim 5, characterized in that, The total number of iterations during training was 100, the maximum learning rate was 0.0001, the minimum learning rate was 0.000001, and the batch size was 4.

Images