CN114358189A - Hydraulic waterproof valve fault diagnosis method based on multi-mode deep residual shrinkage network - Google Patents

Abstract

The invention discloses a hydraulic water-proof valve fault diagnosis method based on a multi-mode deep residual error shrinkage network, which comprises the following steps of: step 1, acquiring original fault signals of a hydraulic waterproof valve by using a plurality of sensors, and preprocessing the original fault signals; step 2, constructing a multi-modal depth residual shrinkage network model, taking the preprocessed signals as input, extracting fault characteristics corresponding to different sensors through the multi-modal depth residual shrinkage network model, and fusing the extracted fault characteristics to obtain a trained multi-modal depth residual shrinkage network model; and 3, carrying out fault diagnosis on the hydraulic waterproof valve by using the trained multi-mode deep residual shrinkage network model. The invention can diagnose different faults of the hydraulic waterproof valve and has the advantages of high fault identification precision and high efficiency.

Description

Hydraulic waterproof valve fault diagnosis method based on multi-mode deep residual shrinkage network

Technical Field

The invention relates to the technical field of fault diagnosis of mechanical equipment, in particular to a fault diagnosis method for a hydraulic waterproof valve based on a multi-mode deep residual shrinkage network.

Background

Many pump trucks have hydraulic waterproof valves installed on the master cylinder, and the hydraulic waterproof valves mainly prevent water in the water tank from adhering to the piston rod of the master cylinder and entering the cylinder barrel of the master cylinder to pollute a hydraulic system and cause system emulsification. The working environment of the hydraulic waterproof valve is extremely harsh, which causes frequent failure of the hydraulic waterproof valve. With the continuous intelligent updating of hydraulic equipment, the intelligent requirements for self-diagnosis of faults of each element of the hydraulic equipment are provided, and the backward self-detection capability of the hydraulic waterproof valve restricts the intelligent development of the hydraulic waterproof valve. The hydraulic valve is wide and important in application, has a plurality of faults and serious damage, can be used for accurately and quickly diagnosing the faults of the hydraulic system, and is favorable for avoiding economic loss and accidents. Therefore, the working state of the hydraulic waterproof valve is monitored in real time, and great benefits are certainly brought to the normal operation of a hydraulic system.

The conventional fault diagnosis method generally comprises 3 steps of feature extraction, feature selection and fault classification. The feature extraction is a key step, and directly influences the quality of fault classification. With the development of a new generation of artificial intelligence, deep learning has been widely introduced into fault diagnosis of mechanical equipment, and is very powerful in feature extraction and classification. However, in the background of strong noise, the fault diagnosis model has problems of low accuracy, overfitting and the like. In recent years, various algorithms for removing noise signals have been developed, such as: wavelet transform, fourier transform, empirical mode decomposition, and the like. In practical engineering application, the hydraulic waterproof valve preferably extracts features through an intelligent method, so that full automation of fault diagnosis is realized. However, when a signal contains a large amount of noise, it is difficult to accurately diagnose the signal only by improving the noise immunity. The fault information contained by a single sensor is always limited. Therefore, to solve this problem, it is necessary to not only improve noise immunity but also enrich the failure information as much as possible.

Disclosure of Invention

The invention aims to provide a hydraulic water-proof valve fault diagnosis method based on a multi-mode deep residual shrinkage network. The invention can diagnose different faults of the hydraulic waterproof valve and has the advantages of high fault identification precision and high efficiency.

The technical scheme of the invention is as follows: the hydraulic water-proof valve fault diagnosis method based on the multi-mode deep residual shrinkage network comprises the following steps:

step 1, acquiring original fault signals of a hydraulic waterproof valve by using a plurality of sensors, and preprocessing the original fault signals;

step 2, constructing a multi-modal depth residual shrinkage network model, taking the preprocessed signals as input, extracting fault characteristics corresponding to different sensors through the multi-modal depth residual shrinkage network model, and fusing the extracted fault characteristics to obtain a trained multi-modal depth residual shrinkage network model;

and 3, carrying out fault diagnosis on the hydraulic waterproof valve by using the trained multi-mode deep residual shrinkage network model.

According to the hydraulic water-proof valve fault diagnosis method based on the multi-mode deep residual shrinkage network, the original fault signals collected by the sensor comprise vibration signals, pressure signals and flow signals.

According to the hydraulic water-proof valve fault diagnosis method based on the multi-mode deep residual shrinkage network, the preprocessing comprises signal segmentation and data standardization;

the signal segmentation is to intercept an original fault signal into samples by a slip method, and obtain a new sample by offsetting or sliding a certain distance of a previous sample, wherein the slip length is set according to a formula:

wherein l_pIs the length of the original fault signal data point; l_sIs the sample length; n is_sRepresenting the number of samples;

the data standardization is to map all data samples to a closed interval [0,1] through a normalization strategy, and convert each original fault signal into a standardized format:

wherein the content of the first and second substances,

in order to be in a standardized format,

the ith sample point of the kth sample in the original fault signal; max (x)^k) And min (x)^k) Respectively representing the maximum and minimum traversed in the kth sample.

According to the hydraulic water-proof valve fault diagnosis method based on the multi-mode depth residual shrinkage network, the multi-mode depth residual shrinkage network model is established by firstly establishing a depth residual shrinkage network model; the depth residual shrinkage network model comprises a convolution layer, batch normalization, an activation function, a global average pooling function and a loss function;

the convolutional layer is used for extracting different input characteristics, reducing parameters in network training, avoiding over-fitting and improving the precision of a network model; the mapping relation between the input features of the convolutional layer and the convolutional kernel is expressed as:

in the formula, x_iIs the ith channel of the input feature map, y_jIs the jth channel of the output feature map, k is the convolution kernel, b is the offset, M_jIs a channel set for computing the jth channel of the output feature map;

the batch normalization is a method for normalizing input features in the training of a depth residual shrinkage network model, and a formula of the batch normalization is represented as follows:

wherein x is_nInput representing the nth observed value, y_nOutput representing nth observed value, N_batchRepresenting the size of the minimum lot, ∈ is a common sense close to zero, γ is the parameter value for the scaled distribution, β is the parameter value for the shifted distribution; μ is the average of the observed values measured,

standard deviation of observed values; sigma²Is the variance of the observed value;

the activation function is a nonlinear transformation part of a depth residual shrinkage network model and is used for maintaining the stability of fault characteristics, and the activation function is a Relu activation function and is expressed as follows:

the global average pooling is a method of calculating an average value from each channel of the feature map; applying a cross entropy error as an objective function of minimization in the global mean pooling, and applying a softmax function to strengthen the range of the fault feature to (0,1), wherein the softmax function is expressed as follows:

wherein x is_j、y_jInput and output, x, of the jth neuron, respectively, in the function_iIs the input of the ith neuron, N_classIs the number of neurons;

the cross entropy error for each observation is expressed as:

where t is the target output, t_jIs the actual probability of the j-th observation;

after the cross entropy error is calculated, optimizing the parameters by adopting a gradient descent algorithm, and fully training a depth residual shrinkage network model after multiple iterations;

the depth residual shrinkage network model adopts a soft threshold segmentation method to reduce noise, and a soft threshold function in the soft threshold segmentation method is as follows:

where x and y are input and output, respectively, and τ represents a threshold;

the threshold value setting meets two conditions, namely, the threshold value is positive, and the threshold value cannot be larger than the input maximum value;

after the depth residual error shrinkage network model is constructed, the multiple depth residual error shrinkage network models adopt a parallel structure to form a multi-mode depth residual error shrinkage network model.

According to the hydraulic water-proof valve fault diagnosis method based on the multi-mode depth residual shrinkage network, the depth residual shrinkage network model is provided with a plurality of residual shrinkage building units, and each residual shrinkage building unit comprises two BN layers, two Relu layers, two convolution layers, a threshold module, a soft threshold module and an identity shortcut;

the threshold module consists of an absolute layer, a GAP layer, a BN layer, a ReLU layer and two output FC layers; the threshold module is used for automatically determining a threshold value in a training process of the deep residual shrinkage network model, the process is that the feature mapping is transmitted to a absolute layer, a GAP layer, a BN layer, a ReLU and two output FC layers, the number of channels of the second layer FC layer is the same as that of the feature mapping, finally, the output of the FC network is scaled to the range of (0,1), and the scaling parameters are obtained by a formula:

wherein z and α are characteristic and scale parameters of the neuron, respectively, and c is an index;

the threshold calculation formula is as follows:

τ_c＝α_c·average|x_i，j，c|；

where τ is the threshold of the feature map x; a, b and c are the width, height and channel index of the feature map x, respectively;

after passing through a plurality of residual shrinkage building units, fusing the features extracted by the multiple sensors in a feature fusion layer, inputting the fused features into a Flatten layer, transmitting the fused features to a Dense layer, and finally classifying the fused features by adopting softmax logistic regression.

Compared with the prior art, the invention utilizes a plurality of sensors to collect the original fault signals of the hydraulic waterproof valve and preprocesses the original fault signals; then, a multi-modal depth residual error shrinkage network model is built, preprocessed signals are used as input, fault features corresponding to different sensors are extracted through the multi-modal depth residual error shrinkage network model, and the extracted fault features are fused to obtain a trained multi-modal depth residual error shrinkage network model; and finally, performing fault diagnosis on the hydraulic waterproof valve by using the trained multi-mode deep residual shrinkage network model. Therefore, the invention adopts a multi-sensor data fusion method to overcome the problem of uncertain fault description in a single-sensor method, namely, overcome the uncertainty of a single sensor. In the aspect of robustness of fault diagnosis, the multi-sensor data fusion method can be more stable than a single-sensor method. According to the method, on one hand, noise components in the signals are automatically reduced through an attention mechanism and soft thresholding, manual extraction of characteristics of a plurality of sensors is not needed, manpower is not liberated, on the other hand, uncertainty and inaccuracy of data of a single sensor are avoided, and fault characteristics can be better mined by combining information of a plurality of sensors. Compared with other multi-sensor data methods, the method is more efficient, can obtain higher diagnosis precision, automatically reduces noise, extracts features, and is suitable for engineering machinery.

Drawings

FIG. 1 is a schematic flow diagram of the present invention;

FIG. 2 is a schematic illustration of a slip process;

FIG. 3 is a schematic diagram of a residual shrinkage building block;

FIG. 4 is a schematic diagram of a multi-modal depth residual shrinkage network model;

FIG. 5 is a confusion matrix-the accuracy of the present invention for each type of fault diagnosis;

fig. 6 is a diagram illustrating a comparison result of various hydraulic water-proof valve diagnosis methods.

Detailed Description

The invention is further illustrated by the following figures and examples, which are not to be construed as limiting the invention.

Example (b): the hydraulic water-proof valve fault diagnosis method based on the multi-mode deep residual shrinkage network, as shown in fig. 1, comprises the following steps:

step 1, acquiring original fault signals of a hydraulic waterproof valve by using a plurality of sensors, and preprocessing the original fault signals; the method comprises the following steps that original fault signals collected by a sensor comprise vibration signals, pressure signals and flow signals, wherein the vibration signals are obtained by an acceleration sensor, the pressure signals are obtained by a pressure sensor, the flow signals are obtained by a flow sensor, and the preprocessing comprises signal segmentation and data standardization;

the signal segmentation is to intercept an original fault signal into samples by a slip method, as shown in fig. 2, a new sample is obtained by offsetting or sliding a certain distance of a previous sample, wherein the slip length is set according to a formula:

wherein l_pIs the length of the original fault signal data point; l_sIs the sample length; n is_sRepresenting the number of samples;

the data standardization is to map all data samples to a closed interval [0,1] through a normalization strategy, and convert each original fault signal into a standardized format:

wherein the content of the first and second substances,

in order to be in a standardized format,

the ith sample point of the kth sample in the original fault signal; max (x)^k) And min (x)^k) Respectively representing the maximum and minimum traversed in the kth sample.

Step 2, constructing a multi-modal depth residual shrinkage network model, taking the preprocessed signals as input, extracting fault characteristics corresponding to different sensors through the multi-modal depth residual shrinkage network model, and fusing the extracted fault characteristics to obtain a trained multi-modal depth residual shrinkage network model;

specifically, the multi-mode depth residual shrinkage network model is established by establishing a depth residual shrinkage network model; the depth residual shrinkage network model comprises a convolution layer, batch normalization, an activation function, a global average pooling function and a loss function;

the convolutional layer is used for extracting different input characteristics, reducing parameters in network training, avoiding over-fitting and improving the precision of a network model; the mapping relation between the input features of the convolutional layer and the convolutional kernel is expressed as:

in the formula, x_iIs the ith channel of the input feature map, y_jIs the jth channel of the output feature map, k is the convolution kernel, b is the offset, M_jIs a channel set for computing the jth channel of the output feature map;

the batch normalization is a method for normalizing input features in the training of a depth residual shrinkage network model, and a formula of the batch normalization is represented as follows:

wherein x is_nInput representing the nth observed value, y_nOutput representing nth observed value, N_batchRepresenting the size of the minimum lot, ∈ is a common sense close to zero, γ is the parameter value for the scaled distribution, β is the parameter value for the shifted distribution; μ is the average of the observed values measured,

standard deviation of observed values; sigma²Is the variance of the observed value;

the activation function is a nonlinear transformation part of a depth residual shrinkage network model and is used for maintaining the stability of fault characteristics, and the activation function is a Relu activation function and is expressed as follows:

the Global Average Pooling (GAP) is a method of calculating an average value from each channel of a feature map; GAP may reduce the number of weights used by the FC output layer, thereby reducing the likelihood that the deep neural network encounters overfitting. The GAP can also solve the displacement variation problem, so that the characteristics learned by the deep neural network are not influenced by the position variation of the fault pulse. In multi-class recognition tasks, cross entropy errors are often used as the objective function for minimization. Cross entropy errors generally result in higher training efficiency than traditional squared average errors, because the gradient of cross entropy errors is less likely to cancel zero relative to weights. Applying a cross entropy error as an objective function of minimization in the global mean pooling, and applying a softmax function to strengthen the range of the fault feature to (0,1), wherein the softmax function is expressed as follows:

wherein x is_j、y_jInput and output, x, of the jth neuron, respectively, in the function_iIs the input of the ith neuronN is_classIs the number of neurons;

the cross entropy error for each observation is expressed as:

where t is the target output, t_jIs the actual probability of the j-th observation;

after the cross entropy error is calculated, optimizing the parameters by adopting a gradient descent algorithm, and fully training a depth residual shrinkage network model after multiple iterations;

the depth residual shrinkage network model adopts a soft threshold segmentation method to reduce noise, the soft threshold segmentation method can remove the features of which the absolute value is smaller than the threshold and reduce the features of which the absolute value is larger than the threshold to 0, and the soft threshold function in the soft threshold segmentation method is as follows:

where x and y are input and output, respectively, and τ represents a threshold;

the threshold value setting meets two conditions, namely, the threshold value is positive, and the threshold value cannot be larger than the input maximum value; in addition, it is preferable to set respective independent thresholds according to the input noise.

The derivative of the soft threshold function is

From the above equation, it can be seen that the function can only be 1 or 0, which has the same properties as Relu. Therefore, the soft threshold can not only reduce noise interference, but also avoid the problem of model gradient disappearance.

The depth residual shrinkage network model has a plurality of residual shrinkage building units, as shown in fig. 4, C is the size of the channel, W is the size of the width, and number 1 is the size of the height; the residual shrinkage construction unit comprises two BN layers, two Relu layers, two convolution layers, a threshold module, a soft threshold module and an identity shortcut;

the threshold module consists of an absolute layer, a GAP layer, a BN layer, a ReLU layer and two output FC layers; the threshold module is used for automatically determining a threshold value in a training process of the deep residual shrinkage network model, the process is that the characteristic mapping is transmitted to a absolute layer, a GAP layer, a BN layer, a ReLU and two output FC layers, the number of channels of the second layer FC layer is the same as the characteristic mapping, finally, the output of the FC network is scaled to the range of (0,1), the scaling parameters are obtained by a formula, and the scaling parameters are obtained by the formula:

wherein z and α are characteristic and scale parameters of the neuron, respectively, and c is an index;

the threshold calculation formula is as follows:

τ_c＝α_c·average|X_a，b，c|；

where τ is a threshold of the feature map X; a, b and c are the width, height and channel index of the feature map x, respectively;

and forming a multi-mode depth residual shrinkage network model by adopting a parallel structure through the plurality of depth residual shrinkage network models, as shown in fig. 3. Signals preprocessed by 3 sensors are input into a multi-modal depth residual shrinkage network model, in a single channel, after passing through a plurality of residual shrinkage building units, features extracted by the 3 sensors are fused in a feature fusion layer, the fused features are input into a Flatten layer and then transferred to a Dense layer, and finally classification is carried out by softmax logistic regression, wherein fault types are shown in a table 1.

TABLE 1 failure types

TABLE 1

Further, the applicant performs fault diagnosis on the hydraulic water-proof valve by using the trained multi-modal deep residual shrinkage network model, and the result is shown in fig. 5. FIG. 5 is a graph of diagnostic confusion matrix results from a fusion of multi-sensor data from test stations. From the results of the confusion matrix, it can be seen that the diagnostic accuracy at this stage is very good.

Further, the applicant compares the trained multi-modal deep residual shrinkage network model with various existing deep learning models for fault diagnosis, and the results are shown in fig. 6, where the comparison is based on a multi-sensor data + cnn (mmcnn) diagnosis method, a single-sensor data + cnn (scnn) fault diagnosis method, and a single-sensor data + drsn (sdrsn) fault diagnosis method. As can be seen from fig. 6, the result of the multi-modal deep residual shrinkage network model (MMDDRSN) diagnosis adopted by the present invention reaches 100%, and the present invention has very good fault diagnosis capability.

In conclusion, the invention adopts the multi-sensor data fusion method to overcome the problem of uncertain fault description in the single-sensor method, namely, overcome the uncertainty of the single sensor. In the aspect of robustness of fault diagnosis, the multi-sensor data fusion method can be more stable than a single-sensor method. According to the method, on one hand, noise components in the signals are automatically reduced through an attention mechanism and soft thresholding, manual extraction of characteristics of a plurality of sensors is not needed, manpower is not liberated, on the other hand, uncertainty and inaccuracy of data of a single sensor are avoided, and fault characteristics can be better mined by combining information of a plurality of sensors. Compared with other multi-sensor data methods, the method is more efficient, can obtain higher diagnosis precision, automatically reduces noise, extracts features, and is suitable for engineering machinery.

Claims (5)

1. The hydraulic water-proof valve fault diagnosis method based on the multi-mode deep residual shrinkage network is characterized by comprising the following steps of: the method comprises the following steps:

step 1, acquiring original fault signals of a hydraulic waterproof valve by using a plurality of sensors, and preprocessing the original fault signals;

step 2, constructing a multi-modal depth residual shrinkage network model, taking the preprocessed signals as input, extracting fault characteristics corresponding to different sensors through the multi-modal depth residual shrinkage network model, and fusing the extracted fault characteristics to obtain a trained multi-modal depth residual shrinkage network model;

and 3, carrying out fault diagnosis on the hydraulic waterproof valve by using the trained multi-mode deep residual shrinkage network model.

2. The hydraulic water-proof valve fault diagnosis method based on the multi-modal depth residual shrinkage network according to claim 1, characterized in that: the raw fault signals collected by the sensor include vibration signals, pressure signals and flow signals.

3. The hydraulic water-proof valve fault diagnosis method based on the multi-modal depth residual shrinkage network according to claim 1, characterized in that: the preprocessing comprises signal segmentation and data normalization;

the signal segmentation is to intercept an original fault signal into samples by a slip method, and obtain a new sample by offsetting or sliding a certain distance of a previous sample, wherein the slip length is set according to a formula:

wherein l_pIs the length of the original fault signal data point; l_sIs the sample length; n is_sRepresenting the number of samples;

the data standardization is to map all data samples to a closed interval [0,1] through a normalization strategy, and convert each original fault signal into a standardized format:

wherein the content of the first and second substances,

in order to be in a standardized format,

the ith sample point of the kth sample in the original fault signal; max (x)^k) And min (x)^k) Respectively representing the maximum and minimum traversed in the kth sample.

4. The hydraulic water-proof valve fault diagnosis method based on the multi-modal depth residual shrinkage network according to claim 1, characterized in that: the method comprises the steps of constructing a multi-mode depth residual shrinkage network model, namely, firstly establishing a depth residual shrinkage network model; the depth residual shrinkage network model comprises a convolution layer, batch normalization, an activation function, a global average pooling function and a loss function;

the convolutional layer is used for extracting different input characteristics, reducing parameters in network training, avoiding over-fitting and improving the precision of a network model; the mapping relation between the input features of the convolutional layer and the convolutional kernel is expressed as:

in the formula, x_iIs the ith channel of the input feature map, y_jIs the jth channel of the output feature map, k is the convolution kernel, b is the offset, M_jIs a channel set for computing the jth channel of the output feature map;

the batch normalization is a method for normalizing input features in the training of a depth residual shrinkage network model, and a formula of the batch normalization is represented as follows:

wherein x is_nInput representing the nth observed value, y_nOutput representing nth observed value, N_batchRepresenting the size of the minimum lot, ∈ is a common sense close to zero, γ is the parameter value for the scaled distribution, β is the parameter value for the shifted distribution; μ is the average of the observed values measured,

standard deviation of observed values; sigma²Is the variance of the observed value;

the activation function is a nonlinear transformation part of a depth residual shrinkage network model and is used for maintaining the stability of fault characteristics, and the activation function is a Relu activation function and is expressed as follows:

the global average pooling is a method of calculating an average value from each channel of the feature map; applying a cross entropy error as an objective function of minimization in the global mean pooling, and applying a softmax function to strengthen the range of the fault feature to (0,1), wherein the softmax function is expressed as follows:

wherein x is_j、y_jInput and output, x, of the jth neuron, respectively, in the function_iIs the input of the ith neuron, N_classIs the number of neurons;

the cross entropy error for each observation is expressed as:

where t is the target output, t_jIs the actual probability of the j-th observation;

after the cross entropy error is calculated, optimizing the parameters by adopting a gradient descent algorithm, and fully training a depth residual shrinkage network model after multiple iterations;

the depth residual shrinkage network model adopts a soft threshold segmentation method to reduce noise, and a soft threshold function in the soft threshold segmentation method is as follows:

where x and y are input and output, respectively, and τ represents a threshold;

the threshold value setting meets two conditions, namely, the threshold value is positive, and the threshold value cannot be larger than the input maximum value;

after the depth residual error shrinkage network model is constructed, the multiple depth residual error shrinkage network models adopt a parallel structure to form a multi-mode depth residual error shrinkage network model.

5. The hydraulic water-proof valve fault diagnosis method based on the multi-modal depth residual shrinkage network according to claim 4, characterized in that: the depth residual error shrinkage network model is provided with a plurality of residual error shrinkage building units, and the residual error shrinkage building units comprise two BN layers, two Relu layers, two convolution layers, a threshold module, a soft threshold module and an identity shortcut;

the threshold module consists of an absolute layer, a GAP layer, a BN layer, a ReLU layer and two output FC layers; the threshold module is used for automatically determining a threshold value in a training process of the deep residual shrinkage network model, the process is that the feature mapping is transmitted to a absolute layer, a GAP layer, a BN layer, a ReLU and two output FC layers, the number of channels of the second layer FC layer is the same as that of the feature mapping, finally, the output of the FC network is scaled to the range of (0,1), and the scaling parameters are obtained by a formula:

wherein z and α are characteristic and scale parameters of the neuron, respectively, and c is an index;

the threshold calculation formula is as follows:

τ_c＝α_c·average|x_a，b，c|；

where τ is the threshold of the feature map x; a, b and c are the width, height and channel index of the feature map x, respectively;

after passing through a plurality of residual shrinkage building units, fusing the features extracted by the multiple sensors in a feature fusion layer, inputting the fused features into a Flatten layer, transmitting the fused features to a Dense layer, and finally classifying the fused features by adopting softmax logistic regression.

Images