Transmission system fault detection method based on fault sensitive feature depth distance measurement

Abstract

Disclosed is a transmission system fault detection method based on deep distance measurement of fault-sensitive features. The method includes: first, using a vibration sensor to collect vibration signals of bevel gears in normal and faulty states, and extracting the time domain and frequency domain features of the signals to construct a sample set; then, calculating the distance between feature classes to assess their sensitivity to faulty states; inputting the filtered normal state features into a deep support vector data description model for training to construct a hypersphere that can learn the common features of normal samples and set a fault threshold; inputting the filtered normal and fault state features into a Deep-SVDD network to calculate the distance to the center of the hypersphere, and judging the health status of the bevel gears according to a preset threshold. The present invention uses feature evaluation technology to filter bevel gear fault-sensitive features, and trains a Deep-SVDD network based on normal state data to construct a hypersphere, ultimately achieving monitoring of the health status of bevel gears in the transmission system.

Description

Technical Field

[0001] The present invention relates to the field of bevel gear condition monitoring, and in particular to a transmission system fault detection method based on fault sensitive feature depth distance measurement. Background Art

[0002] Monitoring the health of transmission systems in high-end equipment such as artillery, helicopters, and aircraft engines is crucial for the safe operation of the entire machine. Existing mechanical condition monitoring methods primarily extract features from bevel gear vibration signals and set safety thresholds to determine operating status. However, these selected features are not necessarily sensitive to fault conditions, making false alarms a common occurrence in practical applications. Furthermore, existing fault detection models often require data from multiple fault categories, while comprehensive fault databases are rarely available in practice, limiting their application in industrial settings. Summary of the Invention

[0003] To address the shortcomings of existing technologies, the present invention provides a transmission system fault detection method based on deep distance measurement of fault-sensitive features. Vibration signals in normal and faulty states are collected and sliced ​​to construct samples. Time and frequency domain features are extracted from each sample to form a sample feature set. Intra- and inter-class distances are then calculated, and a filtering factor is defined to remove features insensitive to faults. Normal features are then input into a Deep-SVDD (Deep Support Vector Data Description) single-class classification model for training, ultimately enabling bevel gear fault detection.

[0004] To achieve the above objectives, the present invention proposes the following technical approach: A transmission system fault detection method based on fault sensitive feature depth distance measurement has the following steps:

[0005] Step 1: Use a sensor to collect vibration signals of bevel gears in normal and faulty states in the transmission system;

[0006] Step 2: performing time series segmentation on the collected vibration signal to construct a sample set, extracting the time-frequency domain features of each sample and forming a one-dimensional feature row vector to construct a sample feature set;

[0007] Step 3: Calculate the distances between and within the feature classes. The ratio of the distances between the feature classes and within the feature classes is the screening factor. Features with a sensitivity greater than the screening factor are retained. The training set consists of normal samples, and the test set consists of a combination of normal and faulty samples.

[0008] Step 4: Input the training set into the Deep-SVDD model to train and learn the common features of normal samples, determine the center and radius of the hypersphere, and set the fault threshold;

[0009] In step 5, the test set is input into the Deep-SVDD model to obtain a feature vector, and the depth distance is calculated and compared with the fault threshold to detect the fault state of the bevel gear.

[0010] In the transmission system fault detection method based on fault-sensitive feature depth distance measurement, in step 2, the sample feature set includes 16 time domain features and 12 frequency domain features, wherein the time domain features include absolute average value, root mean square value, mean square value, root square amplitude, peak value, peak-to-peak value, variance, standard deviation, maximum value, minimum value, waveform index, peak index, pulse index, margin index, skewness and kurtosis, and the frequency domain features include average value, center of gravity frequency, frequency variance, mean square frequency, root mean square frequency, frequency domain skewness index, frequency amplitude variance, frequency domain kurtosis index, frequency standard deviation, frequency domain skewness, frequency domain kurtosis and square root ratio.

[0011] In the transmission system fault detection method based on fault-sensitive feature depth distance measurement, in step 3, the screening factor is 0.3, 20 fault features are screened from 28 features to represent the sample state, the training set is divided into 400 normal samples, and the test set contains four fault types, with 100 samples of each type.

[0012] The transmission system fault detection method based on fault sensitive feature deep distance measurement is characterized in that in step 4, Deep-SVDD model training includes the following steps:

[0013] The Deep-SVDD model includes an autoencoder model composed of linear layers, where the activation function is set to the ReLU function; the dimensions of the three linear layers in the encoder network are 128, 64, and 32, the dimensions of the three linear layers in the decoder are 32, 64, and 128, and the hidden layer feature encoding dimension is 32; samples are input to the autoencoder and the network is trained using the sample reconstruction loss as the objective function, and the mean of the hidden layer features in the encoder is set to the center of the hypersphere c;

[0014] The loss function of the encoder network for training the autoencoder is:

[0015]

[0016] Where R is the radius of the hypersphere, n is the number of training samples, r is a hyperparameter with a value between 0 and 1, and the max{·,·} function represents the maximum value between the two. i ;θ) Encoder network with parameter θ, λ is the loss balance coefficient, N is the number of training samples, ||·|| F represents the F-norm, L represents the number of network layers, θ lRepresents the network parameters of the lth layer. The first loss objective is to minimize the radius of the hypersphere. The second loss objective is to penalize the distance between features outside the hypersphere and the sphere center c. The third loss is the L2 regularization term to prevent network overfitting. During network optimization, R is updated alternately. The specific method is to fix R for a certain number of training times and then select a new hypersphere radius. The selection method is:

[0017]

[0018] Where n out The number of samples outside the sphere, R old is the radius of the hypersphere before updating, R new is the updated hypersphere radius. After updating R, continue training. This iterative update implements network training. Finally, calculate the distance from the normal feature to the sphere center c to determine the fault alarm threshold.

[0019] In the transmission system fault detection method based on fault-sensitive feature depth distance measurement, in step 5, the depth distance is the Euclidean distance between the feature vector output by the test sample input to the Encoder network and the center point c; if the depth distance calculated by inputting the collected sample features into the model exceeds the preset fault threshold three times in a row, it is considered that the bevel gear has a fault.

[0020] Compared to existing technologies, this method eliminates features insensitive to faults by defining a filtering factor based on intra-class and inter-class distances. The Deep-SVDD model employed in this study can identify bevel gear fault states during the testing phase without requiring training with fault data. This method, independent of expert knowledge, offers high detection efficiency and provides a foundation for subsequent maintenance decisions. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 This is a flow chart of a transmission system fault detection method based on fault sensitive feature depth distance measurement provided by an embodiment of the present disclosure;

[0022] Figure 2 This is a schematic diagram of four bevel gear faults provided by an embodiment of the present disclosure;

[0023] Figure 3 Schematic diagram of sensitivity factors of 28 extracted time-frequency features provided by one embodiment of the present disclosure;

[0024] Figure 4 This is a schematic diagram showing a comparison between the depth distance calculated for a feature-screened test set sample and the true label according to an embodiment of the present disclosure;

[0025] Figure 5 This is a schematic diagram of a comparison between the depth distance calculated for a test set sample that has not undergone feature screening and the true label provided by an embodiment of the present disclosure. DETAILED DESCRIPTION

[0026] In order to better illustrate the technical solution of this application, the following Figures 1 to 5 The present invention is described in more detail. It should be emphasized that the detailed description of the embodiments of the present invention is not intended to limit the scope of the invention claimed, but is for a more thorough understanding of the present invention, and the research scope of the present invention should not be limited by the description here.

[0027] It should be noted that certain words are used in the specification and claims to refer to specific components. Those skilled in the art should understand that technicians may use different nouns to refer to the same component. This specification and claims do not use the difference in nouns as a way to distinguish components, but use the difference in the functions of the components as the criterion for distinction. As mentioned throughout the specification and claims, "including" or "comprising" is an open term, so it should be interpreted as "including but not limited to". The subsequent description of the specification is a preferred embodiment of the present invention, but the description is based on the general principles of the specification and is not intended to limit the scope of the invention. The scope of protection of the present invention shall be as defined in the attached claims.

[0028] To facilitate understanding of the embodiments of the present invention, further explanation will be given below using specific embodiments as examples in conjunction with the accompanying drawings, and the accompanying drawings do not constitute a limitation on the embodiments of the present invention.

[0029] In one embodiment, if Figure 1 As shown, a transmission system fault detection method based on fault sensitive feature depth distance measurement includes the following steps:

[0030] In the first step, sensors are used to collect vibration signals of bevel gears in the transmission system in normal and faulty states. Bevel gear fault types include: normal state, tooth surface wear, broken teeth, and small end collapse. The sampling frequency is set to 20480HZ.

[0031] In the second step, the collected signal is segmented into time series, where the number of normal samples is 500 and the number of samples of the other three fault types is 100 each. The length of each sample is 1024. 28 time-frequency domain features are extracted for each sample. The time domain features include absolute mean, root mean square value, mean square value, root square amplitude, peak value, peak-to-peak value, variance, standard deviation, maximum value, minimum value, waveform index, peak index, pulse index, margin index, skewness and kurtosis. The frequency domain features include mean, center of gravity frequency, frequency variance, mean square frequency, root mean square frequency, frequency domain skewness, frequency domain amplitude variance, frequency domain kurtosis index, frequency standard deviation, frequency domain skewness, frequency domain kurtosis and square root ratio.

[0032] In the third step, the average intra-class distance and the average inter-class distance of each feature are calculated, and the screening factor is defined as the ratio of the two. The specific steps are as follows:

[0033] First, define H as the number of fault categories in the sample set, N as the number of samples in each category, and M as the feature dimension. Perform maximum-minimum normalization on the samples in each feature dimension, and then calculate the intra-class distance A of the m-th dimension feature of the h-th category. cm :

[0034]

[0035] Where, f hm (i) and f hm (j) represents the mth eigenvalue of the i-th and j-th samples in the h-th class respectively;

[0036] Continue to calculate the average intra-class distance of each feature:

[0037]

[0038] Calculate the mean intra-class distance of the h-th class m-th dimension feature:

[0039]

[0040] Then we can get the average inter-class distance of each feature:

[0041]

[0042] Finally, the sensitivity evaluation factor α of each feature is obtained m :

[0043] α m =A m / B m

[0044] α m The higher the value, the more sensitive the feature is to faults, and vice versa. It should be noted that the number of fault samples in this example is small, so when calculating α j The number of normal samples can be consistent with it. Select the screening factor based on the feature evaluation factor. In this example, it is set to 0.3, that is, select features with a feature evaluation factor greater than 0.3. Divide the training set according to the filtered features and test set where N train is the number of training samples, N test is the number of test samples, x i is the sample feature, with a dimension of 20; N train and N testBoth are set to 400, but the training samples are 400 normal samples, and the test samples contain four fault types, with 100 samples for each type.

[0045] In the fourth step, the Deep-SVDD single classification model is used to train the training set samples. The specific training steps are as follows:

[0046] An autoencoder model consisting of linear layers is designed, in which the activation function is set to the ReLU function; the activation function is set to the ReLU function; the dimensions of the three linear layers in the encoder are 128, 64, and 32, the dimensions of the three linear layers in the decoder are 32, 64, and 128, and the hidden layer feature encoding dimension is 32; the samples are input into the autoencoder and the network is trained with the reconstruction loss as the objective function, and the mean of the hidden layer features in the encoder is set to the center of the hypersphere c.

[0047] Then train the Encoder network of the autoencoder, and the loss function is:

[0048]

[0049] Where R is the radius of the hypersphere, n is the number of training samples, r is a hyperparameter with a value between 0 and 1, and the max{·,·} function represents the maximum value between the two. i ;θ) Encoder network with parameter θ, λ is the loss balance coefficient, N is the number of training samples, ||·|| F represents the F-norm, L represents the number of network layers, θ l Represents the network parameters of the lth layer. The first loss goal is to minimize the radius of the hypersphere, that is, we hope that the output features are distributed in a compact space. The second loss goal is to penalize the distance between the features outside the hypersphere and the center c, that is, we hope that the hypersphere can contain most of the normal features. The third loss is the L2 regularization term to prevent the network from overfitting. It can be seen that the first two losses are in conflict. A small R is difficult to guarantee that a large number of normal features are included. Therefore, it is necessary to update R alternately during network optimization. The specific method is to fix R for a certain number of training times, select a new hypersphere radius, and use the depth distance measure R new Defined as:

[0050]

[0051] Where n out The number of samples outside the sphere, R old is the radius of the hypersphere before updating, R new is the updated hypersphere radius. After updating R, continue training, iterating and updating to achieve network training, thereby controlling the output features of most samples within a compact hypersphere. Finally, calculate the distance from the normal feature to the sphere center c to determine the fault alarm threshold.

[0052] In the fifth step, the test sample is fed into the trained encoder network. The distance between the output feature and the sphere center is calculated and compared with the fault alarm threshold. If the distance is greater than the fault threshold, the sample is considered to be in a fault state. Otherwise, the bevel gear is considered to be in a normal state.

[0053] In one embodiment, the bevel gear fault types include: normal state, tooth surface wear, tooth breakage and small end collapse, the sampling frequency is 20480 Hz, and the sample length is 1024.

[0054] In one embodiment, the feature extractors are all BP neural networks. First, the autoencoder is pre-trained to set the mean of the hidden layer vector to the center point c, and then the encoder network of the autoencoder is trained to make the normal samples as close to the center point c as possible.

[0055] In one embodiment, Score is the distance between the feature vector output from the Encoder network when the test sample is input and the center point c.

[0056] Figure 2 The four types of bearing failures used in the bevel gear failure experiment are shown, including normal state, tooth surface wear, tooth breakage and small end collapse.

[0057] Figure 3 The feature sensitivity factors corresponding to the 28 extracted time-frequency domain features are displayed. It is clear that eight features, numbered 11, 12, 13, 14, 15, 23, and 24, are insensitive to faults. Therefore, in this example, a screening factor of 0.3 is set to exclude these eight features. These eight features are specifically the time domain waveform index, time domain peak index, time domain pulse index, time domain margin index, time domain skew, time domain kurtosis, frequency domain skew, and frequency domain kurtosis.

[0058] Figure 4This figure shows the test results of a trained model after inputting a sample feature set after feature screening into a Deep-SVDD network. Label in the figure represents the true sample label, with 0 indicating a normal state and 1 indicating a faulty state. Samples numbered 101 to 200 are experiencing tooth breakage faults, samples numbered 201 to 300 are experiencing tooth surface wear faults, and samples numbered 301 to 400 are experiencing small-segment collapse faults. Score represents the distance between the normalized test sample feature vector and the center c of the hypersphere. To ensure that most normal samples are within the hypersphere, the fault threshold in this example is set to 0.21. As can be seen, some faulty sample features are judged as normal. To address this issue, continuously collected samples can be used to reduce misjudgments. Specifically, the model inputs sample features collected in real time. If the calculated depth distance exceeds the preset fault threshold three times in a row, the bevel gear is considered faulty. The threshold setting can be adjusted based on actual needs.

[0059] Figure 5 The test results of the model after the sample feature set without feature screening is input into the Deep-SVDD network training are shown. The legend setting is the same as Figure 5 It can be seen that the features that are insensitive to faults have an impact on network training, making it difficult to define a reasonable fault threshold to determine normal samples and faulty samples.

[0060] The above are only preferred embodiments of the present disclosure and are not intended to limit the implementation methods and protection scope of the present disclosure. Those skilled in the art should be aware that any solutions obtained by equivalent substitutions and obvious changes made using the contents of the present disclosure should be included in the protection scope of the present disclosure.

Claims

1. A transmission system fault detection method based on fault sensitive feature depth distance measurement, characterized in that: The method comprises the following steps: Step 1: Use a sensor to collect vibration signals of bevel gears in normal and fault states in the transmission system; Step 2: performing time series segmentation on the collected vibration signal to construct a sample set, extracting the time-frequency domain features of each sample and forming a one-dimensional feature row vector to construct a sample feature set; Step 3: Calculate the distances between and within the feature classes. The ratio of the distances between the feature classes and within the feature classes is the screening factor. Features with a sensitivity greater than the screening factor are retained. The training set consists of normal samples, and the test set consists of a combination of normal and faulty samples. Step 4: Input the training set into the Deep-SVDD model to train and learn the common features of normal samples, determine the center and radius of the hypersphere, and set the fault threshold; Step 5: The test set is input into the Deep-SVDD model to obtain a feature vector, and the depth distance is calculated and compared with the fault threshold to detect the fault state of the bevel gear; In step 4, Deep-SVDD model training includes the following steps: The Deep-SVDD model includes an autoencoder model composed of linear layers, where the activation function is set to the ReLU function; the dimensions of the three linear layers in the encoder network are 128, 64, and 32, the dimensions of the three linear layers in the decoder are 32, 64, and 128, and the hidden layer feature encoding dimension is 32; samples are input to the autoencoder and the network is trained using the sample reconstruction loss as the objective function, and the mean of the hidden layer features in the encoder is set to the center of the hypersphere c; The loss function of the encoder network for training the autoencoder is: ， Where R is the radius of the hypersphere, is the number of training samples, is a hyperparameter with a value between 0 and 1, The function represents the maximum value between the two, The parameters are Encoder network, is the loss balance coefficient, N is the number of training samples, represents the F-norm, Indicates the number of network layers, Indicates the The network parameters of the layer, the first loss goal is to minimize the radius of the hypersphere, the second loss goal is to penalize the distance between the features outside the hypersphere and the sphere center c, and the third loss is the L2 regularization term to prevent the network from overfitting. During network optimization, R is updated alternately. The specific method is to fix R for a certain number of training times and then select a new hypersphere radius. The selection method is: ， In the formula The number of samples outside the sphere, is the radius of the hypersphere before updating, is the updated hypersphere radius. After updating R, continue training. The network training is achieved by iterative updating. Finally, the distance from the normal feature to the sphere center c is calculated to determine the fault alarm threshold. In step 5, the depth distance is the Euclidean distance between the feature vector output by the test sample input to the Encoder network and the center point c; if the depth distance calculated by inputting the collected sample features into the model exceeds the preset fault threshold three times in a row, the bevel gear is considered to be faulty.

2. A transmission system fault detection method based on fault sensitive feature depth distance measurement according to claim 1, characterized in that: In step 2, the sample feature set includes 16 time domain features and 12 frequency domain features, wherein the time domain features include absolute mean value, root mean square value, mean square value, root square amplitude, peak value, peak-to-peak value, variance, standard deviation, maximum value, minimum value, waveform index, peak index, pulse index, margin index, skewness and kurtosis, and the frequency domain features include mean value, center of gravity frequency, frequency variance, mean square frequency, root mean square frequency, frequency domain skewness index, frequency amplitude variance, frequency domain kurtosis index, frequency standard deviation, frequency domain skewness, frequency domain kurtosis and square root ratio.

3. The transmission system fault detection method based on fault sensitive feature depth distance measurement according to claim 1, characterized in that: In step 3, the screening factor is 0.3, and 20 fault features are selected from the 28 features to represent the sample status. The training set is divided into 400 normal samples, and the test set contains four fault types, with 100 samples for each type.

Images