Equipment fault diagnosis method and device based on multi-scale 2D dilated convolutional network

Abstract

The present invention discloses a method and device for equipment fault diagnosis based on a multi-scale 2D dilated convolutional network, and relates to the technical field of mechanical equipment fault diagnosis. The method includes: obtaining mechanical equipment data; converting the mechanical equipment data into a two-dimensional image; inputting the two-dimensional image into a trained multi-scale 2D dilated convolutional neural network; and obtaining mechanical equipment fault diagnosis results based on the two-dimensional image and the trained multi-scale 2D dilated convolutional neural network. The present invention introduces dilated convolution into the network, flexibly realizes multi-scale feature extraction, reduces information redundancy, reduces calculation amount, and reduces the number of model parameters; introduces a cascade structure into the network, combines low-resolution information with high-resolution information to improve the fault diagnosis accuracy when the number of samples is insufficient, and improves the model accuracy when the samples are noisy.

Description

Technical Field

[0001] The present invention relates to the technical field of mechanical equipment fault diagnosis, and in particular to an equipment fault diagnosis method and device based on a multi-scale 2D dilated convolutional network. Background Art

[0002] With the development of today's society, industrial equipment used in various fields of people's production and life has become more complex. Due to the complex and changeable operating conditions of industrial equipment systems and the relatively harsh operating environment, the rate of system performance degradation and failure and the probability of failure have increased. For key equipment systems, once an accident caused by failure occurs, it will lead to significant economic losses or even casualties. In industrial production, it is very important to predict equipment failures. Accurately analyzing and predicting equipment safety hazards and providing effective measures can timely control and prevent the loss of productivity caused by equipment in production. Fault diagnosis technology, as a comprehensive discipline integrating multiple technologies, provides effective technical support for the safe and reliable operation of industrial systems. Its main purpose is to explore how to detect, locate and identify various types of faults. The specific content includes locating the location, size and time of the fault, identifying the type of fault, etc.

[0003] Nowadays, there are still many scholars committed to the exploration and research of this field, which has always been one of the research hotspots at home and abroad. Classic fault diagnosis technologies include expert systems, model decomposition, signal processing and other methods, but these methods have disadvantages such as low versatility, complex modeling, weak robustness, and poor representation ability, and their application scope is greatly limited. With the increasing complexity and intelligence of mechanical equipment, coupled with the background of the era of big data development, intelligent fault diagnosis methods with versatility, end-to-end, and powerful representation capabilities have become a hot spot and difficulty in research. Deep learning technology in artificial intelligence is becoming more and more mature, and has the characteristics of versatility and end-to-end. Its powerful feature extraction capability can solve the problem of low precision and accuracy in fault diagnosis due to factors such as complex equipment structure and weak signals.

[0004] CNN (Convolutional Neural Networks) is a feedforward neural network and one of the representative network structures in the field of deep learning. Its essence is a hierarchical model consisting of several convolutional layers, pooling layers, nonlinear activation function layers and fully connected layers. As one of the most effective deep learning, convolutional neural networks have been applied to fault diagnosis. Since the original signal is a one-dimensional vibration signal, 1D-CNN is usually used for mechanical equipment failure. However, 1D convolution does not consider the relationship between the original signal cycles. And the requirements for the original fault signal are high, so 2D-CNN is used in fault diagnosis. 2D-CNN does not need to process the original vibration signal, and directly converts the original 1D vibration signal into a 2D signal that can be recognized by 2D convolution, and considers the relationship between the cycles. Then, CNN is used to extract features from the converted signal, and finally the classification result is output through the fully connected layer. However, as the requirements for classification become higher and higher, in order to obtain a high-precision model, a very deep network layer and a large amount of data are needed for training. However, as the number of network layers deepens, overfitting and gradient disappearance will occur, and the accuracy will no longer increase with the deepening of the network layers. At the same time, deepening the number of network layers will consume a lot of computing costs and the large number of parameters is not conducive to actual industrial deployment. The signals collected in the actual production process will be interfered by noise. When the number of training samples decreases, the existing methods have low accuracy and poor robustness when facing these situations. Summary of the invention

[0005] The present invention is proposed to solve the problems that the prior art has low accuracy and poor robustness in predicting equipment failures.

[0006] In order to solve the above technical problems, the present invention provides the following technical solutions:

[0007] On the one hand, the present invention provides a device fault diagnosis method based on a multi-scale 2D dilated convolutional network, the method is implemented by an electronic device, and the method includes:

[0008] S1. Obtain mechanical equipment data.

[0009] S2. Convert mechanical equipment data into a two-dimensional image.

[0010] S3. Input the two-dimensional image into the trained multi-scale 2D dilated convolutional neural network.

[0011] S4. Based on two-dimensional images and the trained multi-scale 2D dilated convolutional neural network, the mechanical equipment fault diagnosis results are obtained.

[0012] Optionally, converting the mechanical equipment data into a two-dimensional image in S2 includes:

[0013] Select a m*m sequence y(i) from the mechanical equipment data, i=1,2,3,…,m*m; arrange the sequence in order into an m*m matrix p(j,k), j=1,2,3,…,m; k=1,2,3,…,m; the matrix p(j,k) is a two-dimensional image.

[0014] Optionally, the multi-scale 2D dilated convolutional neural network in S3 includes a shallow feature extraction unit and a plurality of residual calculation units.

[0015] The multiple residual calculation units are sequentially a first residual calculation unit, a second residual calculation unit and a third residual calculation unit.

[0016] The shallow feature extraction unit is sequentially arranged before the first residual calculation unit.

[0017] The multi-scale 2D dilated convolutional neural network also includes a global average pooling layer and a fully connected layer.

[0018] Optionally, each of the multiple residual computing units includes multiple residual networks connected sequentially, and the multiple residual networks become the first residual network, the second residual network and the third residual network in turn; the multiple residual networks in each residual computing unit all use 2D dilated convolution, and different dilated convolution rates are set.

[0019] Optionally, the training process of the multi-scale 2D dilated convolutional neural network in S3 includes:

[0020] S31, obtaining a sample data set; the sample data set includes a training sample data set and a test sample data set.

[0021] S32. Train the multi-scale 2D dilated convolutional neural network according to the training sample data set. If the preset number of iterations is reached or the training loss function is less than a preset threshold, stop the training to obtain the trained multi-scale 2D dilated convolutional neural network.

[0022] S33. Testing the trained multi-scale 2D dilated convolutional neural network according to the test sample data set to obtain the tested multi-scale 2D dilated convolutional neural network, that is, the trained multi-scale 2D dilated convolutional neural network.

[0023] Optionally, obtaining a sample data set in S31 includes:

[0024] S311. Acquire fault data of mechanical equipment; the fault data is a one-dimensional original vibration signal.

[0025] S312, converting the fault data into a two-dimensional image.

[0026] S313: Mark the two-dimensional image with a fault label to obtain a sample data set.

[0027] Optionally, the training of the multi-scale 2D dilated convolutional neural network according to the training sample data set in S32 includes:

[0028] S321. Input the training sample data set into the shallow feature extraction unit to obtain a shallow feature map.

[0029] S322. Input the shallow feature map into the first residual calculation unit to obtain a feature map of the first resolution.

[0030] S323: Input the training sample data set and the feature map of the first resolution into the second residual calculation unit to obtain a feature map of the second resolution.

[0031] S324: Input the training sample data set and the feature map of the second resolution into a third residual calculation unit to obtain a feature map of the third resolution.

[0032] S325. Input the feature map of the third resolution into the global average pooling layer to obtain the average value after dimensionality reduction.

[0033] S326. The average value after dimension reduction is passed through the softmax activation function of the fully connected layer to obtain the fault diagnosis result of the training sample.

[0034] Optionally, the shallow feature extraction unit includes a convolutional layer, a batch normalization layer, and an activation layer.

[0035] In S321, the training sample data set is input into the shallow feature extraction unit to obtain the shallow feature map including:

[0036] S3211. Input the training sample data set into the convolutional layer to extract feature information to obtain feature information.

[0037] S3212. Input the feature information into the batch standard layer to obtain standardized feature information.

[0038] S3213. Input the standardized feature information into the activation layer to obtain a shallow feature map.

[0039] On the other hand, the present invention provides a device for diagnosing equipment faults based on a multi-scale 2D dilated convolutional network, which is used to implement a method for diagnosing equipment faults based on a multi-scale 2D dilated convolutional network, and the device comprises:

[0040] The data acquisition module is used to acquire mechanical equipment data.

[0041] The data conversion module is used to convert mechanical equipment data into a two-dimensional image.

[0042] The input module is used to input the two-dimensional image into the trained multi-scale 2D dilated convolutional neural network.

[0043] The output module is used to obtain mechanical equipment fault diagnosis results based on two-dimensional images and the trained multi-scale 2D dilated convolutional neural network.

[0044] Optionally, the data conversion module is further used to:

[0045] Select a m*m sequence y(i) from the mechanical equipment data, i=1,2,3,…,m*m; arrange the sequence in order into an m*m matrix p(j,k), j=1,2,3,…,m; k=1,2,3,…,m; the matrix p(j,k) is a two-dimensional image.

[0046] Optionally, the multi-scale 2D dilated convolutional neural network in S3 includes a shallow feature extraction unit and a plurality of residual calculation units.

[0047] The multiple residual calculation units are sequentially a first residual calculation unit, a second residual calculation unit and a third residual calculation unit.

[0048] The feature extraction unit is sequentially arranged before the first residual calculation unit.

[0049] The multi-scale 2D dilated convolutional neural network also includes a global average pooling layer and a fully connected layer.

[0050] Optionally, each of the multiple residual computing units includes multiple residual networks connected sequentially, and the multiple residual networks are respectively a first residual network, a second residual network and a third residual network; the multiple residual networks in each residual computing unit all use 2D dilated convolution and are set with different dilated convolution rates.

[0051] Optionally, the input module is further configured to:

[0052] S31, obtaining a sample data set; the sample data set includes a training sample data set and a test sample data set.

[0053] S32. Train the multi-scale 2D dilated convolutional neural network according to the training sample data set. If the preset number of iterations is reached or the training loss function is less than a preset threshold, stop the training to obtain the trained multi-scale 2D dilated convolutional neural network.

[0054] S33. Testing the trained multi-scale 2D dilated convolutional neural network according to the test sample data set to obtain the tested multi-scale 2D dilated convolutional neural network, that is, the trained multi-scale 2D dilated convolutional neural network.

[0055] Optionally, the input module is further configured to:

[0056] S311. Acquire fault data of mechanical equipment; the fault data is a one-dimensional original vibration signal.

[0057] S312, converting the fault data into a two-dimensional image.

[0058] S313: Mark the two-dimensional image with a fault label to obtain a sample data set.

[0059] Optionally, the input module is further configured to:

[0060] S321. Input the training sample data set into the shallow feature extraction unit to obtain a shallow feature map.

[0061] S322. Input the shallow feature map into the first residual calculation unit to obtain a feature map of the first resolution.

[0062] S323: Input the training sample data set and the feature map of the first resolution into the second residual calculation unit to obtain a feature map of the second resolution.

[0063] S324: Input the training sample data set and the feature map of the second resolution into a third residual calculation unit to obtain a feature map of the third resolution.

[0064] S325. Input the feature map of the third resolution into the global average pooling layer to obtain the average value after dimensionality reduction.

[0065] S326. The average value after dimension reduction is passed through the softmax activation function of the fully connected layer to obtain the fault diagnosis result of the training sample.

[0066] Optionally, the shallow feature extraction unit includes a convolutional layer, a batch normalization layer, and an activation layer.

[0067] Optionally, the input module is further configured to:

[0068] S3211. Input the training sample data set into the convolutional layer to extract feature information to obtain feature information.

[0069] S3212. Input the feature information into the batch standard layer to obtain standardized feature information.

[0070] S3213. Input the standardized feature information into the activation layer to obtain a shallow feature map.

[0071] On the one hand, an electronic device is provided, comprising a processor and a memory, wherein the memory stores at least one instruction, and the at least one instruction is loaded and executed by the processor to implement the above-mentioned device fault diagnosis method based on a multi-scale 2D dilated convolutional network.

[0072] On the one hand, a computer-readable storage medium is provided, wherein at least one instruction is stored in the storage medium, and the at least one instruction is loaded and executed by a processor to implement the above-mentioned device fault diagnosis method based on a multi-scale 2D dilated convolutional network.

[0073] The beneficial effects brought about by the technical solution provided by the embodiment of the present invention include at least:

[0074] In the above scheme, a method for exploring the two-dimensional features of the original signal is provided, taking into account the connection between cycles. In addition, this data preprocessing method can reduce the experience of experts as much as possible without pre-calculating the refinement parameters. The present application effectively avoids computational waste by implementing multi-scale through dilated convolution, and at the same time extracts features of different receptive fields, avoiding information redundancy, thereby reducing the risk of overfitting.

[0075] This application greatly reduces the amount of network model calculation and parameter through the improved residual network. By reducing the number of samples, the improvement of model diagnosis accuracy is verified in the bearing data set. By introducing noise into the samples, the robustness of the model is verified.

[0076] This application introduces dilated convolution into the network, flexibly realizes multi-scale feature extraction, reduces the redundancy of information, reduces the amount of calculation, and reduces the number of model parameters; introduces a cascade structure into the network, combines low-resolution information with high-resolution information to improve the fault diagnosis accuracy when the number of samples is insufficient, and improves the model accuracy when the samples are noisy. BRIEF DESCRIPTION OF THE DRAWINGS

[0077] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0078] Figure 1 It is a schematic flow chart of a device fault diagnosis method based on a multi-scale 2D dilated convolutional network provided by an embodiment of the present invention;

[0079] Figure 2 Schematic diagram of a multi-scale 2D dilated convolutional neural network provided by an embodiment of the present invention;

[0080] Figure 3 2 is a schematic diagram comparing ordinary convolution and dilated convolution provided by an embodiment of the present invention;

[0081] Figure 4 This is a schematic diagram of a noise-free result of a first data set provided by an embodiment of the present invention;

[0082] Figure 5 This is a schematic diagram of the first data set with noise results provided by an embodiment of the present invention;

[0083] Figure 6 is a schematic diagram of a noise-free result of a second data set provided by an embodiment of the present invention;

[0084] Figure 7 is a schematic diagram of a noisy result of a second data set provided by an embodiment of the present invention;

[0085] Figure 8 It is a schematic diagram of the process flow of a multi-scale 2D dilated convolutional neural network training method provided by an embodiment of the present invention;

[0086] Fig. 9 is a schematic diagram of an improved residual network structure provided by an embodiment of the present invention;

[0087] Fig.10 It is a block diagram of a device fault diagnosis apparatus based on a multi-scale 2D dilated convolutional network provided by an embodiment of the present invention;

[0088] Fig.11 It is a structural schematic diagram of an electronic device provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0089] In order to make the technical problems, technical solutions and advantages to be solved by the present invention more clear, a detailed description will be given below with reference to the accompanying drawings and specific embodiments.

[0090] like Figure 1 As shown, an embodiment of the present invention provides a device fault diagnosis method based on a multi-scale 2D dilated convolutional network, which can be implemented by an electronic device. Figure 1 The flow chart of the device fault diagnosis method based on the multi-scale 2D dilated convolutional network is shown. The processing flow of the method may include the following steps:

[0091] S11. Obtain mechanical equipment data.

[0092] S12. Convert the mechanical equipment data into a two-dimensional image.

[0093] Optionally, select a sequence y(i) of length m*m from the mechanical equipment data, i=1,2,3,…,m*m; arrange the sequence in order into an m*m matrix p(j,k), j=1,2,3,…,m; k=1,2,3,…,m; the matrix p(j,k) is a two-dimensional image.

[0094] In a feasible implementation, the fault data of mechanical equipment is usually a one-dimensional original vibration signal. CNN does not require the form of input data when performing classification, that is, the original data can be used for fault diagnosis without any processing. However, in order to improve the diagnostic accuracy, the data is generally pre-processed such as random cropping, horizontal flipping, etc. to enhance the data or perform Fourier transform, S transform and other methods to convert between time domain, frequency domain, and time-frequency domain, but these methods are all based on the original one-dimensional basis and do not consider the connection between data cycles. Therefore, in order to consider the connection between cycles, this application converts the one-dimensional signal into a two-dimensional image.

[0095] Specifically, the matrix p(j,k) is shown in the following formula (1):

[0096]

[0097] The advantage of this data processing method is that it provides a way to explore the two-dimensional characteristics of the original signal, taking into account the connection between cycles. In addition, this data preprocessing method can reduce the expert experience as much as possible without pre-calculating the refinement parameters.

[0098] S13. Input the two-dimensional image into the trained multi-scale 2D dilated convolutional neural network.

[0099] Among them, Figure 2 As shown, the multi-scale 2D dilated convolutional neural network includes a shallow feature extraction unit and multiple residual calculation units.

[0100] The multiple residual calculation units may be at least three residual calculation units, which may be a first residual calculation unit, a second residual calculation unit, and a third residual calculation unit, etc. in sequence.

[0101] The shallow feature extraction unit is sequentially arranged before the first residual calculation unit.

[0102] The output of the previous residual calculation unit together with the original two-dimensional image serves as the input of the next residual calculation unit.

[0103] The multi-scale 2D dilated convolutional neural network also includes a global average pooling layer and a fully connected layer.

[0104] Optionally, each of the multiple residual calculation units is connected in sequence by multiple residual networks, and the multiple residual networks can be at least three residual networks; they can be the first residual network, the second residual network, and the third residual network, etc.; the three residual networks in each residual calculation unit are configured to use 2D dilated convolution, and are respectively configured with different dilated convolution rates.

[0105] In a feasible implementation, multiple residual computing units have the same structure, but because different expansion convolution rates are used, the convolution kernel size of each residual network is different. In order to ensure the consistency of input and output sizes, convolution uses zero padding operation.

[0106] Generally, after the size of the convolution kernel is given by an ordinary CNN, the size of the convolution kernel will not change during the entire convolution process, thereby reducing the risk of overfitting. However, the flexibility of the convolution kernel is also reduced. Therefore, the determination of the convolution and size should be cautious and appropriate, which may waste computing resources and time. Fixed convolution kernels also limit their application in multi-scale data analysis. In order to solve this problem, this application applies dilated convolution. The dilated convolution has two parameters: one is the convolution kernel size, and the other is the dilated convolution rate r, which represents the size of the expansion. The comparison between ordinary convolution and dilated convolution is as follows: Figure 3 As shown in the figure, the common convolution (dilated convolution = 1) and dilated convolution (i.e. dilated convolution = 2) are similar in that the size of the convolution kernel is the same, that is, the number of parameters in the neural network is the same, and the difference is that the dilated convolution has a larger receptive field, which is the size of the convolution kernel seen on the image. For the current layer input x, for the output y at each position i and the convolution kernel (or weight) ω, the corresponding relationship of the common convolution is shown in the following formula (2):

[0107] y(i)=∑ k xω(k) (2)

[0108] The corresponding relationship of the dilated convolution is shown in the following formula (3):

[0109] y(i)=∑ k x(i+r*k)ω(k) (3)

[0110] Usually, to ensure that there is no difference between the input size and the output size, the size of the convolution kernel is selected as an odd number. Therefore, when selecting the expanded convolution rate, in order to ensure that the size of the convolution kernel after expansion is an odd number, a multiple of 2 is selected as the expanded convolution rate. For example, the expanded convolution rate of the first residual network can be 2; the expanded convolution rate of the second residual network can be 4; and the expanded convolution rate of the third residual network can be 8.

[0111] By convolving the input data with a multi-scale kernel, feature maps at different scales can be obtained. If different-sized convolution kernels are directly applied to two-dimensional convolution layers to achieve multi-scale, it will lead to computational waste. Therefore, this application inserts r-1 0 elements between the convolution kernels of each dimension instead of directly changing the convolution kernel size. Therefore, achieving multi-scale through dilated convolution effectively avoids computational waste, while extracting features of different receptive fields, avoiding information redundancy, and thus reducing the risk of overfitting.

[0112] Optionally, the training process of the multi-scale 2D dilated convolutional neural network in S13 includes S131-S133:

[0113] S131. Obtain a sample data set.

[0114] The sample data set includes a training sample data set and a test sample data set.

[0115] In a feasible implementation manner, the above step S131 may include the following steps S1311-S1313:

[0116] S1311. Obtain fault data of mechanical equipment; the fault data is a one-dimensional original vibration signal.

[0117] S1312, converting the fault data into a two-dimensional image.

[0118] S1313: Mark the two-dimensional image with a fault label to obtain a sample data set.

[0119] S132. Train the multi-scale 2D dilated convolutional neural network according to the training sample data set. If a preset number of iterations is reached or the training loss function is less than a preset threshold, stop the training to obtain a trained multi-scale 2D dilated convolutional neural network.

[0120] Optionally, training the multi-scale 2D dilated convolutional neural network according to the training sample data set may include S1321-S1326:

[0121] S1321. Input the training sample data set into the shallow feature extraction unit to obtain a shallow feature map.

[0122] Among them, the shallow feature extraction unit includes a convolutional layer, a batch standard layer and an activation layer.

[0123] In a feasible implementation manner, the above step S1321 may include the following steps S13211-S13213:

[0124] S13211. Input the training sample data set into the convolutional layer to extract feature information to obtain feature information.

[0125] S13212. Input the feature information into the batch standard layer to obtain standardized feature information.

[0126] S13213. Input the standardized feature information into the activation layer to obtain a shallow feature map.

[0127] S1322. Input the shallow feature map into the first residual calculation unit to obtain a feature map of the first resolution.

[0128] S1323: Input the training sample data set and the feature map of the first resolution into the second residual calculation unit to obtain a feature map of the second resolution.

[0129] S1324: Input the training sample data set and the feature map of the second resolution into a third residual calculation unit to obtain a feature map of the third resolution.

[0130] S1325. Input the feature map of the third resolution into the global average pooling layer to obtain the average value after dimensionality reduction.

[0131] S1326. The average value after dimension reduction is activated by the softmax function of the fully connected layer to obtain the fault diagnosis result of the training sample.

[0132] S133. Testing the trained multi-scale 2D dilated convolutional neural network according to the test sample data set to obtain the tested multi-scale 2D dilated convolutional neural network, that is, the trained multi-scale 2D dilated convolutional neural network.

[0133] S14. Based on the two-dimensional image and the trained multi-scale 2D dilated convolutional neural network, the mechanical equipment fault diagnosis results are obtained.

[0134] In a feasible implementation, the embodiment of the present application uses two public bearing data sets for simulation. The first data set includes bearing fault types, namely inner ring fault, outer ring fault, and ball fault. Each fault corresponds to three damage sizes, a total of 9 fault types plus one normal type, a total of 10 types. The second data set includes inner ring fault and outer ring fault and four corresponding loads, plus two normal data, a total of 10 types. The two data sets are tested in the presence and absence of noise respectively. The results are as follows Figure 4-7 shown.

[0135] In an embodiment of the present invention, a method for exploring the two-dimensional features of the original signal is provided, taking into account the connection between cycles. In addition, this data preprocessing method can reduce the experience of experts as much as possible without pre-calculating the refinement parameters. The present application effectively avoids computational waste by implementing multi-scale through dilated convolution, and at the same time extracts features of different receptive fields, avoiding information redundancy, thereby reducing the risk of overfitting.

[0136] This application greatly reduces the amount of network model calculation and parameter through the improved residual network. By reducing the number of samples, the improvement of model diagnosis accuracy is verified in the bearing data set. By introducing noise into the samples, the robustness of the model is verified.

[0137] This application introduces dilated convolution into the network, flexibly realizes multi-scale feature extraction, reduces the redundancy of information, reduces the amount of calculation, and reduces the number of model parameters; introduces a cascade structure into the network, combines low-resolution information with high-resolution information to improve the fault diagnosis accuracy when the number of samples is insufficient, and improves the model accuracy when the samples are noisy.

[0138] like Figure 8 As shown, an embodiment of the present invention provides a multi-scale 2D dilated convolutional neural network training method, which can be implemented by a district electronic device. Figure 8 The flowchart of the multi-scale 2D dilated convolutional neural network training method shown in the figure, the processing flow of the method may include the following steps:

[0139] S21. Obtain a sample data set.

[0140] The sample data set includes a training sample data set and a test sample data set.

[0141] In a feasible implementation manner, the above step S21 may include the following steps S211-S213:

[0142] S211. Acquire fault data of mechanical equipment; the fault data is a one-dimensional original vibration signal.

[0143] S212: Convert the fault data into a two-dimensional image.

[0144] S213: Mark the two-dimensional image with a fault label to obtain a sample data set.

[0145] In a feasible implementation, the converted images are labeled with corresponding fault labels and then divided into training set samples and test set samples in a certain ratio; the training set is used to minimize the error between the predicted label and the actual label to train the model. The test set is used to test the diagnostic ability of the model.

[0146] S22. Input the training sample data set into the shallow feature extraction unit to obtain a shallow feature map.

[0147] Among them, the shallow feature extraction unit includes a convolutional layer, a batch standard layer and an activation layer.

[0148] In a feasible implementation manner, the above step S22 may include the following steps S221-S223:

[0149] S221. Input the training sample data set into the convolutional layer to extract feature information to obtain feature information.

[0150] S222: Input the feature information into the batch standard layer to obtain standardized feature information.

[0151] In one possible implementation, the batch norm layer gives a minimum batch {x1,x2,...,x m} and parameters γ and β are transformed as shown in equations (4)-(7):

[0152]

[0153] Among them, μ is the mean; σ 2 is the variance; is the output after standardization; is the batch norm layer output.

[0154] S223. Input the standardized feature information into the activation layer to obtain a shallow feature map.

[0155] In a feasible implementation, the activation layer uses ReLU (Rectified Linear Unit, linear activation function), which is defined as shown in the following formula (8):

[0156]

[0157] in, is the output of the batch norm layer.

[0158] S23. Input the shallow feature map into the first residual calculation unit to obtain a feature map of the first resolution.

[0159] In a feasible implementation, the first residual network extracts features of different sizes, and outputs a feature map with a higher resolution after passing through the first residual calculation unit. The first residual calculation unit includes three residual networks. The expansion convolution rate of the first residual network can be 2, the expansion convolution rate of the second residual network can be 4, and the expansion convolution rate of the third residual network can be 8. The training process and diagnosis process of each residual network are as follows: Fig. 9 shown.

[0160] S24: Input the training sample data set and the feature map of the first resolution into a second residual calculation unit to obtain a feature map of the second resolution.

[0161] In a feasible implementation, the higher-resolution feature map output after passing through the first residual calculation unit is combined with the original image through cascading and sent to the second residual calculation unit to extract the missing information. The second residual calculation unit includes the same three residual networks as the first residual calculation unit.

[0162] S25. Input the training sample data set and the feature map of the second resolution into a third residual calculation unit to obtain a feature map of the third resolution.

[0163] In a feasible implementation, the higher-resolution feature map output after passing through the second residual calculation unit is cascaded and combined with the original image and sent to the third residual calculation unit to extract the missing information. The third residual calculation unit includes three residual networks that are the same as the first and second residual calculation units.

[0164] S26. Input the feature map of the third resolution into the global average pooling layer to obtain the average value after dimensionality reduction.

[0165] S27. The average value after dimension reduction is passed through the softmax activation function of the fully connected layer to obtain the fault diagnosis result of the training sample.

[0166] In a feasible implementation, the network can better express the complex nonlinear relationship between the bearing vibration signal and the bearing state. The network saves the network parameter weights for each round of training, and the training ends when the network training reaches a satisfactory effect.

[0167] If the preset number of iterations is reached or the training loss function is less than a preset threshold, the training is stopped in S28 to obtain a trained multi-scale 2D dilated convolutional neural network.

[0168] S29. Test the trained multi-scale 2D dilated convolutional neural network according to the test sample data set to obtain the tested multi-scale 2D dilated convolutional neural network, that is, the trained multi-scale 2D dilated convolutional neural network.

[0169] In a feasible implementation, a test set is used to test the quality of network performance. The quality of network performance is determined by comparing the fault labels of real data with the fault labels predicted by the network. When the test results are satisfactory, the network no longer needs to rely on the experience of the operator and end-to-end fault diagnosis can be achieved.

[0170] In an embodiment of the present invention, a method for exploring the two-dimensional features of the original signal is provided, taking into account the connection between cycles. In addition, this data preprocessing method can reduce the experience of experts as much as possible without pre-calculating the refinement parameters. The present application effectively avoids computational waste by implementing multi-scale through dilated convolution, and at the same time extracts features of different receptive fields, avoiding information redundancy, thereby reducing the risk of overfitting.

[0171] This application greatly reduces the amount of network model calculation and parameter through the improved residual network. By reducing the number of samples, the improvement of model diagnosis accuracy is verified in the bearing data set. By introducing noise into the samples, the robustness of the model is verified.

[0172] This application introduces dilated convolution into the network, flexibly realizes multi-scale feature extraction, reduces the redundancy of information, reduces the amount of calculation, and reduces the number of model parameters; introduces a cascade structure into the network, combines low-resolution information with high-resolution information to improve the fault diagnosis accuracy when the number of samples is insufficient, and improves the model accuracy when the samples are noisy.

[0173] like Fig.10 As shown, an embodiment of the present invention provides a device fault diagnosis apparatus 1000 based on a multi-scale 2D dilated convolutional network. The device 1000 is applied to implement a device fault diagnosis method based on a multi-scale 2D dilated convolutional network. The device 1000 includes:

[0174] The data acquisition module 1001 is used to acquire mechanical equipment data.

[0175] The data conversion module 1002 is used to convert the mechanical equipment data into a two-dimensional image.

[0176] The input module 1003 is used to input the two-dimensional image into the trained multi-scale 2D dilated convolutional neural network.

[0177] The output module 1004 is used to obtain mechanical equipment fault diagnosis results based on the two-dimensional image and the trained multi-scale 2D dilated convolutional neural network.

[0178] Optionally, the data conversion module 1002 is further configured to:

[0179] Select a m*m sequence y(i) from the mechanical equipment data, i=1,2,3,…,m*m; arrange the sequence in order into an m*m matrix p(j,k), j=1,2,3,…,m; k=1,2,3,…,m; the matrix p(j,k) is a two-dimensional image.

[0180] Optionally, the multi-scale 2D dilated convolutional neural network in S3 includes a shallow feature extraction unit and a plurality of residual calculation units.

[0181] The multiple residual calculation units are sequentially a first residual calculation unit, a second residual calculation unit and a third residual calculation unit.

[0182] The feature extraction unit is sequentially arranged before the first residual calculation unit.

[0183] The multi-scale 2D dilated convolutional neural network also includes a global average pooling layer and a fully connected layer.

[0184] Optionally, each of the multiple residual computing units includes multiple residual networks connected sequentially, and the multiple residual networks are respectively a first residual network, a second residual network and a third residual network; the multiple residual networks in each residual computing unit all use 2D dilated convolution and are set with different dilated convolution rates.

[0185] Optionally, the input module 1003 is further used to:

[0186] S31, obtaining a sample data set; the sample data set includes a training sample data set and a test sample data set.

[0187] S32. Train the multi-scale 2D dilated convolutional neural network according to the training sample data set. If the preset number of iterations is reached or the training loss function is less than a preset threshold, stop the training to obtain the trained multi-scale 2D dilated convolutional neural network.

[0188] S33. Testing the trained multi-scale 2D dilated convolutional neural network according to the test sample data set to obtain the tested multi-scale 2D dilated convolutional neural network, that is, the trained multi-scale 2D dilated convolutional neural network.

[0189] Optionally, the input module 1003 is further used to:

[0190] S311. Acquire fault data of mechanical equipment; the fault data is a one-dimensional original vibration signal.

[0191] S312, converting the fault data into a two-dimensional image.

[0192] S313: Mark the two-dimensional image with a fault label to obtain a sample data set.

[0193] Optionally, the input module 1003 is further used to:

[0194] S321. Input the training sample data set into the shallow feature extraction unit to obtain a shallow feature map.

[0195] S322. Input the shallow feature map into the first residual calculation unit to obtain a feature map of the first resolution.

[0196] S323: Input the training sample data set and the feature map of the first resolution into the second residual calculation unit to obtain a feature map of the second resolution.

[0197] S324: Input the training sample data set and the feature map of the second resolution into a third residual calculation unit to obtain a feature map of the third resolution.

[0198] S325. Input the feature map of the third resolution into the global average pooling layer to obtain the average value after dimensionality reduction.

[0199] S326. The average value after dimension reduction is passed through the softmax activation function of the fully connected layer to obtain the fault diagnosis result of the training sample.

[0200] Optionally, the shallow feature extraction unit includes a convolutional layer, a batch normalization layer, and an activation layer.

[0201] Optionally, the input module 1003 is further used to:

[0202] S3211. Input the training sample data set into the convolutional layer to extract feature information to obtain feature information.

[0203] S3212. Input the feature information into the batch standard layer to obtain standardized feature information.

[0204] S3213. Input the standardized feature information into the activation layer to obtain a shallow feature map.

[0205] In an embodiment of the present invention, a method for exploring the two-dimensional features of the original signal is provided, taking into account the connection between cycles. In addition, this data preprocessing method can reduce the experience of experts as much as possible without pre-calculating the refinement parameters. The present application effectively avoids computational waste by implementing multi-scale through dilated convolution, and at the same time extracts features of different receptive fields, avoiding information redundancy, thereby reducing the risk of overfitting.

[0206] This application greatly reduces the amount of network model calculation and parameter through the improved residual network. By reducing the number of samples, the improvement of model diagnosis accuracy is verified in the bearing data set. By introducing noise into the samples, the robustness of the model is verified.

[0207] This application introduces dilated convolution into the network, flexibly realizes multi-scale feature extraction, reduces the redundancy of information, reduces the amount of calculation, and reduces the number of model parameters; introduces a cascade structure into the network, combines low-resolution information with high-resolution information to improve the fault diagnosis accuracy when the number of samples is insufficient, and improves the model accuracy when the samples are noisy.

[0208] Fig.11 1 is a schematic diagram of the structure of an electronic device 1100 provided by an embodiment of the present invention. The electronic device 1100 may have relatively large differences due to different configurations or performances, and may include one or more processors (CPU) 1101 and one or more memories 1102, wherein the memory 1102 stores at least one instruction, and the at least one instruction is loaded and executed by the processor 1101 to implement the following device fault diagnosis method based on a multi-scale 2D dilated convolutional network:

[0209] S1. Obtain mechanical equipment data.

[0210] S2. Convert mechanical equipment data into a two-dimensional image.

[0211] S3. Input the two-dimensional image into the trained multi-scale 2D dilated convolutional neural network.

[0212] S4. Based on two-dimensional images and the trained multi-scale 2D dilated convolutional neural network, the mechanical equipment fault diagnosis results are obtained.

[0213] In an exemplary embodiment, a computer-readable storage medium is also provided, such as a memory including instructions, which can be executed by a processor in a terminal to complete the above-mentioned device fault diagnosis method based on a multi-scale 2D dilated convolutional network. For example, the computer-readable storage medium can be a ROM, a random access memory (RAM), a CD-ROM, a tape, a floppy disk, an optical data storage device, etc.

[0214] A person skilled in the art will understand that all or part of the steps to implement the above embodiments may be accomplished by hardware or by instructing related hardware through a program, and the program may be stored in a computer-readable storage medium, and the above-mentioned storage medium may be a read-only memory, a disk or an optical disk, etc.

[0215] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A device fault diagnosis method based on a multi-scale 2D dilated convolutional network, characterized in that: The method comprises: S1. Obtain mechanical equipment data; S2, converting the mechanical equipment data into a two-dimensional image; S3, inputting the two-dimensional image into a trained multi-scale 2D dilated convolutional neural network; S4, obtaining a mechanical equipment fault diagnosis result based on the two-dimensional image and the trained multi-scale 2D dilated convolutional neural network; The multi-scale 2D dilated convolutional neural network in S3 includes a shallow feature extraction unit and a plurality of residual calculation units; The plurality of residual calculation units are sequentially a first residual calculation unit, a second residual calculation unit and a third residual calculation unit; The shallow feature extraction unit is sequentially arranged before the first residual calculation unit; The multi-scale 2D dilated convolutional neural network also includes a global average pooling layer and a fully connected layer; Each of the plurality of residual computing units comprises a plurality of residual networks connected in sequence, wherein the plurality of residual networks are a first residual network, a second residual network, and a third residual network in sequence; the plurality of residual networks in each residual computing unit all adopt 2D dilated convolution, and are provided with different dilated convolution rates, and the dilated convolution rate is a multiple of 2; Among them, the 2D dilated convolution inserts r-1 0 elements between the convolution kernels of each dimension, and r represents the dilated convolution rate; The training process of the multi-scale 2D dilated convolutional neural network in S3 includes: S31, obtaining a sample data set; the sample data set includes a training sample data set and a test sample data set; S32, training the multi-scale 2D dilated convolutional neural network according to the training sample data set, and stopping the training if a preset number of iterations is reached or the training loss function is less than a preset threshold, to obtain a trained multi-scale 2D dilated convolutional neural network; S33, testing the trained multi-scale 2D dilated convolutional neural network according to the test sample data set to obtain a tested multi-scale 2D dilated convolutional neural network, that is, a trained multi-scale 2D dilated convolutional neural network; The step of training the multi-scale 2D dilated convolutional neural network according to the training sample data set in S32 includes: S321, inputting the training sample data set into the shallow feature extraction unit to obtain a shallow feature map; S322, inputting the shallow feature map into a first residual calculation unit to obtain a feature map of a first resolution; S323, inputting the training sample data set and the feature map of the first resolution into a second residual calculation unit to obtain a feature map of the second resolution; S324, inputting the training sample data set and the feature map of the second resolution into a third residual calculation unit to obtain a feature map of the third resolution; S325, inputting the feature map of the third resolution into a global average pooling layer to obtain an average value after dimensionality reduction; S326. The average value after dimension reduction is subjected to the softmax activation function of the fully connected layer to obtain the fault diagnosis result of the training sample.

2. The method according to claim 1, characterized in that: The step of converting the mechanical equipment data into a two-dimensional image in S2 includes: A sequence y(i) of length m*m is selected from the mechanical equipment data, i=1,2,3,…,m*m; the sequence is arranged in order into an m*m matrix p(j,k), j=1,2,3,…,m; k=1,2,3,…,m; the matrix p(j,k) is a two-dimensional image.

3. The method according to claim 1, characterized in that The obtaining of the sample data set in S31 includes: S311, acquiring fault data of mechanical equipment; the fault data is a one-dimensional original vibration signal; S312, converting the fault data into a two-dimensional image; S313: Mark the two-dimensional image with a fault label to obtain a sample data set.

4. The method according to claim 1, characterized in that: The shallow feature extraction unit includes a convolution layer, a batch standard layer and an activation layer; The step of inputting the training sample data set into the shallow feature extraction unit in step S321 to obtain a shallow feature map includes: S3211, inputting the training sample data set into the convolution layer to extract feature information to obtain feature information; S3212, inputting the feature information into the batch standard layer to obtain standardized feature information; S3213. Input the standardized feature information into the activation layer to obtain a shallow feature map.

5. A device fault diagnosis device based on a multi-scale 2D dilated convolutional network, characterized in that: The device comprises: A data acquisition module, used to acquire mechanical equipment data; A data conversion module, used for converting the mechanical equipment data into a two-dimensional image; An input module, inputting the two-dimensional image into a trained multi-scale 2D dilated convolutional neural network; An output module, used to obtain a mechanical equipment fault diagnosis result based on the two-dimensional image and the trained multi-scale 2D dilated convolutional neural network; The multi-scale 2D dilated convolutional neural network includes a shallow feature extraction unit and a plurality of residual calculation units; The plurality of residual calculation units are sequentially a first residual calculation unit, a second residual calculation unit and a third residual calculation unit; The shallow feature extraction unit is sequentially arranged before the first residual calculation unit; The multi-scale 2D dilated convolutional neural network also includes a global average pooling layer and a fully connected layer; Each of the plurality of residual computing units comprises a plurality of residual networks connected in sequence, wherein the plurality of residual networks are a first residual network, a second residual network, and a third residual network in sequence; the plurality of residual networks in each residual computing unit all adopt 2D dilated convolution, and are provided with different dilated convolution rates, and the dilated convolution rate is a multiple of 2; Among them, the 2D dilated convolution inserts r-1 0 elements between the convolution kernels of each dimension, and r represents the dilated convolution rate; The training process of the multi-scale 2D dilated convolutional neural network includes: S31, obtaining a sample data set; the sample data set includes a training sample data set and a test sample data set; S32, training the multi-scale 2D dilated convolutional neural network according to the training sample data set, and stopping the training if a preset number of iterations is reached or the training loss function is less than a preset threshold, to obtain a trained multi-scale 2D dilated convolutional neural network; S33, testing the trained multi-scale 2D dilated convolutional neural network according to the test sample data set to obtain a tested multi-scale 2D dilated convolutional neural network, that is, a trained multi-scale 2D dilated convolutional neural network; The step of training the multi-scale 2D dilated convolutional neural network according to the training sample data set in S32 includes: S321, inputting the training sample data set into the shallow feature extraction unit to obtain a shallow feature map; S322, inputting the shallow feature map into a first residual calculation unit to obtain a feature map of a first resolution; S323, inputting the training sample data set and the feature map of the first resolution into a second residual calculation unit to obtain a feature map of the second resolution; S324, inputting the training sample data set and the feature map of the second resolution into a third residual calculation unit to obtain a feature map of the third resolution; S325, inputting the feature map of the third resolution into a global average pooling layer to obtain an average value after dimensionality reduction; S326. The average value after dimension reduction is subjected to the softmax activation function of the fully connected layer to obtain the fault diagnosis result of the training sample.

Images