A rolling bearing fault diagnosis method, system and diagnosis terminal based on SSCP-CNN

Abstract

The present invention provides a rolling bearing fault diagnosis method, system and diagnosis terminal based on SSCP-CNN, which relates to the technical field of rolling bearings. The method collects data, divides the data into a training set and a test set, and then performs labeling processing; constructs a one-dimensional CNN with two convolutional layers and two pooling layers, and forms a training set; optimizes the hyperparameters in the training set by a sparrow search algorithm; calculates the loss function Loss, and converges the loss function through multiple iterations; uses a small-size convolutional pooling to replace the first maximum pooling layer, and maintains consistent hyperparameters to achieve hyperparameter transmission; uses Adam to optimize the adaptive learning rate for model training; if the loss function convergence requirement is met, the SSCP-CNN training is completed; uses the trained SSCP-CNN to perform fault diagnosis on the test set, outputs the diagnosis result, calculates the performance index and performs result visualization analysis. The present invention maintains the stability of fault diagnosis; in data with different noise intensities, the model can achieve high diagnostic accuracy.

Description

Background Art

[0002] Rolling bearings are important components of mechanical equipment. Their health status has a huge impact on the equipment's efficiency and life expectancy. Therefore, it is of great significance to use the monitoring information of mechanical equipment to diagnose rolling bearing faults. Among them, rolling bearing fault diagnosis methods based on deep learning have been a research hotspot in recent years.

[0003] At present, the CNN neural network is used for rolling bearing fault diagnosis. The CNN neural network consists of convolutional layers, pooling layers and fully connected layers. The pooling layer is mainly used for data dimension reduction and is an important part of the CNN network. Conventional pooling methods include maximum pooling and average pooling. Studies have shown that they have the following disadvantages: [ : Max pooling simply uses the maximum activation value in the pooling area to represent the entire area, which will cause problems such as blurred features and noise diffusion; although average pooling takes into account all activation values ​​in the pooling area, it gives them the same weight, which will lead to the loss of important information. In response to the above problems, the prior art proposes a hybrid pooling method, which has better accuracy than max pooling and average pooling, but the essence of hybrid pooling is to randomly use max pooling or average pooling, which does not fundamentally solve the problems of conventional pooling.

[0004] Some existing technologies have proposed K (K is a preset empirical value) to support spatial pooling, and its output is the average of K larger activation values ​​in the pooling area, ignoring the impact of low activation values; some existing technologies have designed a random maximum pooling strategy, which sorts the activation values ​​in the pooling area and randomly selects one of them as the pooling result, reducing the possibility of noise diffusion, but the accuracy is greatly affected by its randomness. Summary of the invention

[0005] The rolling bearing fault diagnosis method based on SSCP-CNN provided by the present invention solves the problems existing in maximum pooling and average pooling from the perspective of CNN's own structure and further reduces the influence of noise on the fault diagnosis accuracy of one-dimensional time series. The present invention also adopts small-size convolution pooling with adaptive pooling step to replace the conventional pooling SSCP-CNN model. The SSCP-CNN model takes a one-dimensional bearing vibration signal as input and achieves the highest fault diagnosis accuracy in the rolling bearing data set, and the model performance advantage increases continuously with the increase of noise intensity.

[0006] Methods include:

[0007] Step 1: Collect one-dimensional bearing vibration data, divide the one-dimensional bearing vibration data into training set and test set, and use One-hot encoding for labeling;

[0008] Step 2: Based on the SSCP-CNN model, a one-dimensional CNN with two convolutional layers and two pooling layers is constructed, and the data is trained based on the one-dimensional CNN to form a training set;

[0009] Step 3: Optimize the hyperparameters in the training set using the sparrow search algorithm;

[0010] Step 4: Calculate the loss function Loss , and through multiple iterations, the loss function Loss convergence;

[0011] Step 5: Use small-size convolution pooling to replace the first maximum pooling layer, and keep the SSCP-CNN and one-dimensional CNN with consistent hyperparameters to achieve hyperparameter transfer;

[0012] Step 6: Input the training set data into SSCP-CNN and use Adam to optimize the adaptive learning rate for model training;

[0013] Step 7: If the loss function convergence requirement is met, the SSCP-CNN training is completed, otherwise return to step 6;

[0014] Step 8: Use the trained SSCP-CNN to perform fault diagnosis on the test set, output the diagnosis results, calculate the performance indicators and perform visual analysis of the results.

[0015] It should be further explained that the mathematical model of the convolutional layer operation in step 2 is described as:

[0016] (1)

[0017] In the formula, For the l Tier The output of the feature vector, is the input feature vector, For the The weight matrix of the convolution kernel of the layer, is the bias term, is the activation function used to perform nonlinear transformation on the output;

[0018] After the convolutional layer, the rectified linear unit is used to enhance the nonlinear expression ability of the network. The mathematical expression of the rectified linear unit is as follows:

[0019] (2)

[0020] In the formula, is the output value of the convolutional layer, for Activation values ​​after the rectified linear unit.

[0021] It should be further explained that in step 2, the maximum pooling is used to select the largest value in the pooling area to represent the entire pooling area. The expression is as follows:

[0022] (3)

[0023] In the formula, For the l Tier i The feature vector t The value of a neuron, For the t The range of neurons for pooling, k is the length of the pooling region, is the output of the pooling layer.

[0024] In step 2, the mean value in the pooling area is selected based on average pooling to represent the entire pooling area. The mathematical expression is as follows:

[0025] (4).

[0026] It should be further explained that the SSCP-CNN model in step 2 also includes: 2 large convolutional kernel convolutional layers, 2 batch normalization layers, 1 small-size convolutional pooling layer, 1 maximum pooling layer, 2 fully connected layers and 1 output layer.

[0027] A BN layer is added after each convolutional layer to keep the input of each layer of the neural network in the same distribution during the training process. The BN layer operates as follows:

[0028] (11).

[0029] It should be further explained that in step 4, the loss function Loss The minimum mean square error is calculated using the following formula:

[0030] (12)

[0031] In the formula, Indicates i The predicted labels of samples, Indicates i The true labels of samples, N Indicates the sample size.

[0032] The present invention also provides a rolling bearing fault diagnosis system based on SSCP-CNN, the system comprising: a data set processing module, a training set construction module, a parameter search module, a loss function calculation module, a parameter transfer module, a model training module, a function state judgment module and a diagnosis test module;

[0033] The data set processing module is used to collect one-dimensional bearing vibration data, divide the one-dimensional bearing vibration data into training set and test set, and use One-hot encoding for labeling processing;

[0034] The training set construction module constructs a one-dimensional CNN with two convolutional layers and two pooling layers based on the SSCP-CNN model, and trains the data based on the one-dimensional CNN to form a training set;

[0035] The parameter search module optimizes the hyperparameters in the training set through the sparrow search algorithm;

[0036] The loss function calculation module is used to calculate the loss function Loss , and through multiple iterations, the loss function Loss convergence;

[0037] The parameter transfer module is used to replace the first maximum pooling layer with a small-size convolution pooling layer, and to keep the SSCP-CNN and one-dimensional CNN with consistent hyperparameters, thus realizing the transfer of hyperparameters.

[0038] The model training module is used to input the training set data into SSCP-CNN and use Adam to optimize the adaptive learning rate for model training;

[0039] The function state judgment module is used to judge whether the loss function meets the convergence requirements. If so, the SSCP-CNN training is completed;

[0040] The diagnostic test module is used to use the trained SSCP-CNN to perform fault diagnosis on the test set, output the diagnosis results, calculate the performance indicators and perform visual analysis of the results.

[0041] The present invention also provides a diagnostic terminal, comprising a memory, a processor and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, the steps of a rolling bearing fault diagnosis method based on SSCP-CNN are implemented.

[0042] It can be seen from the above technical solutions that the present invention has the following advantages:

[0043] The rolling bearing fault diagnosis method of the present invention can adaptively pool the step size of the SSCP and comprehensively consider all the values ​​in the pooling area, and assign different weights according to the convolution kernel, thereby solving the shortcomings of the maximum pooling fuzzy feature and the average pooling equal weight. The minimum level at which the SSCP of the present invention can play a comparative advantage is the network architecture, rather than the simple replacement of the pooling layer. In the experiment of comparing and analyzing SSCP in different network architectures, SSCP-CNN achieved the highest accuracy among the four structures, and at the same time, the highest stability.

[0044] The fault diagnosis experiment shows that the SSCP-CNN of the present invention not only controls the training time within an acceptable range, but also has higher accuracy and stability. The model has more obvious advantages when processing noisy data, and the algorithm performance continues to expand with the increase of noise intensity. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] In order to more clearly illustrate the technical solution of the present invention, the accompanying drawings required for use in the description will be briefly introduced below. Obviously, the accompanying drawings in the following description are only some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative work.

[0046] Figure 1 It is a one-dimensional CNN structure diagram;

[0047] Figure 2 Compute example graphs for small-size convolutional pooling, max pooling, and average pooling;

[0048] Figure 3 Schematic diagram of the SSCP-CNN model;

[0049] Figure 4 This is a flow chart of the rolling bearing fault diagnosis method based on SSCP-CNN;

[0050] Figure 5 There are four different convolutional pooling structure diagrams;

[0051] Figure 6 Schematic diagram of rolling element vibration signal with different signal-to-noise ratios;

[0052] Figure 7 This is a schematic diagram of the accuracy of CNNs with four different convolutional pooling structures under different signal-to-noise ratio data;

[0053] Figure 8 This is a comparison chart of the accuracy of each algorithm under different noise conditions;

[0054] Fig. 9 This is a schematic diagram to visualize the features of each layer of the PooL+PooL structure;

[0055] Fig.10 This is a schematic diagram to visualize the features of each layer of the SSCP +PooL structure; DETAILED DESCRIPTION

[0056] The rolling bearing fault diagnosis method based on SSCP-CNN provided by the present invention involves a convolutional neural network (CNN), which is a typical deep learning network and is widely used in the field of rolling bearing fault diagnosis. The SSCP-CNN rolling bearing fault diagnosis model is used to improve the robustness to environmental noise.

[0057] Among them, the rolling bearing fault diagnosis method based on SSCP-CNN provided by the present invention has both hardware-level technology and software-level technology. Basic technologies generally include technologies such as sensors, special artificial intelligence chips, cloud computing, distributed storage, big data processing technology, operation / interaction system, mechatronics, etc. The software technology of the rolling bearing fault diagnosis method mainly includes machine learning / deep learning and programming language, and the programming language includes but is not limited to object-oriented programming language-such as Java, Smalltalk, C++, and also includes conventional procedural programming language-such as "C" language or similar programming language. Moreover, machine learning and deep learning usually include artificial neural networks, belief networks, reinforcement learning, transfer learning, inductive learning, and teaching learning. A one-dimensional CNN with two convolutional layers and two pooling layers is constructed through the SSCP-CNN model, and the sensor monitoring, data transmission and other technologies are used. Adaptive pooling step size is adopted, and different weights are given to each activation value in the pooling area, which can not only complete the downsampling function, but also comprehensively utilize all activation values ​​in the pooling area. It further effectively solves the problem that the accuracy of the random maximum pooling strategy is greatly affected by its randomness, which sorts the activation values ​​in the pooling area and randomly selects one of them as the pooling result.

[0058] The rolling bearing fault diagnosis method based on SSCP-CNN is applied to one or more diagnostic terminals, wherein the diagnostic terminal is a device that can automatically perform numerical calculations and / or information processing according to pre-set or stored instructions, and its hardware includes but is not limited to microprocessors, application specific integrated circuits (ASIC), programmable gate arrays (FPGA), digital signal processors (DSP), embedded devices, etc.

[0059] The diagnostic terminal can be any electronic product that can interact with the user, such as a personal computer, a tablet computer, a smart phone, a personal digital assistant (PDA), an interactive network television (IPTV), a smart wearable device, etc.

[0060] The network where the diagnostic terminal is located includes but is not limited to the Internet, wide area network, metropolitan area network, local area network, virtual private network (Virtual Private Network, VPN), etc.

[0061] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0062] See also Figure 4 The figure is a flow chart of a rolling bearing fault diagnosis method based on SSCP-CNN in a specific embodiment, the method comprising:

[0063] Step 1: Collect one-dimensional bearing vibration data, divide the one-dimensional bearing vibration data into training set and test set, and use One-hot encoding for labeling;

[0064] In an exemplary embodiment, CNN has the characteristics of sparse connection, parameter sharing and equivariant representation. It is a typical feedforward neural network, usually composed of multiple convolutional layers and pooling layers stacked together. The one-dimensional CNN structure is as follows: Figure 1 shown.

[0065] The mathematical model of convolutional layer operation can be described as:

[0066] (1)

[0067] In the formula, For the l Tier The output of the feature vector, is the input feature vector, For the The weight matrix of the convolution kernel of the layer, is the bias term, is the activation function used to perform nonlinear transformation on the output.

[0068] After the convolutional layer, the Rectified Linear Unit (ReLU) is used to enhance the nonlinear expression ability of the network and avoid overfitting problems. The mathematical expression of ReLU is as follows:

[0069] (2)

[0070] In the formula, is the output value of the convolutional layer, for Activation value after ReLU.

[0071] The pooling layer of this embodiment can effectively reduce the feature dimension, thereby reducing the amount of learnable parameters and avoiding overfitting in the CNN training process. The maximum pooling and average pooling are widely used. The maximum pooling selects the largest value in the pooling area to represent the entire pooling area. The mathematical expression is as follows:

[0072] (3)

[0073] In the formula, For the l Tier i The feature vector t The value of a neuron, For the t The range of neurons for pooling ( k is the length of the pooling region), is the output of the pooling layer.

[0074] Average pooling selects the mean value in the pooling area to represent the entire pooling area. The mathematical expression is as follows:

[0075] (4)

[0076] Step 2: Based on the SSCP-CNN model, a one-dimensional CNN with two convolutional layers and two pooling layers is constructed, and the data is trained based on the one-dimensional CNN to form a training set;

[0077] In this embodiment, although the pooling layer can reduce the amount of training parameters, reduce the risk of overfitting, and improve generalization ability, the conventional pooling layer has obvious defects: the output result of the maximum pooling is determined by the maximum value in the pooling area, and other values ​​in the pooling area are discarded while downsampling, which will affect the accuracy of the model.

[0078] Especially when the input data contains high-frequency noise, the maximum value is also affected by the high-frequency noise. The results after maximum pooling are all affected by noise, causing noise diffusion, which in turn aggravates the impact of the noise signal. Average pooling gives the same weight to all values, which will cause blurring or even loss of important feature information, resulting in a decrease in accuracy.

[0079] In order to solve the problems existing in maximum pooling and average pooling, this embodiment proposes to use a small-size convolution layer with an adaptive pooling step size to replace the SSCP-CNN of the pooling layer. The basic idea of ​​small-size convolution pooling is to comprehensively consider the values ​​of each neuron in the pooling area to avoid the problem that the maximum pooling only considers the maximum value and ignores the remaining activation values, and adjusts the weights of each neuron in the pooling area through the convolution kernel, which makes up for the equal weight problem of average pooling.

[0080] The calculation formula for small-size convolution pooling is as follows:

[0081] (5)

[0082] In the formula, The first one is the small size convolution pooling i outputs, is the Relu activation function, n is the sample length, k is the length of the pooling region, r is the pooling step size, is the weight of the small-size convolution pooling, For the i A bias.

[0083] Small-size convolution pooling, maximum pooling, and average pooling are shown below. Figure 2 As shown, Figure 2 The “*” in the figure represents the convolution operation.

[0084] In this embodiment, in order to verify the feasibility of replacing the pooling layer with a small-size convolution pooling, standard pooling and convolution operation formulas are configured.

[0085] The standard pooling formula is:

[0086] (6)

[0087] In the formula, Indicates the width is w , the height is h , the number of channels is u The pooled input, for s arrive f The mapping of , and the calculation formula is , represents the output of the convolutional layer, p forp norm, when p When it approaches infinity, the maximum pooling operation model is obtained:

[0088] (7)

[0089] The standard convolution formula is:

[0090] (8)

[0091] In the formula, is the weight of the convolutional layer.

[0092] From formula (6) and formula (8), we can see that, on the one hand, both pooling and convolution operations depend on the activation value of the previous feature layer, and the pooling layer is a special convolution operation in terms of operation; on the other hand, convolution relies on the activation function to control the output interval within [0,1], while pooling relies on p The norm controls the output interval within [0,1]. The two methods of compressing the output interval are different, but the data output interval remains the same. Therefore, it is feasible to replace the original pooling layer with a small-size convolution pooling. Considering that the role of conventional pooling is downsampling, the small-size convolution layer also has the function of downsampling.

[0093] In this embodiment, the standard pooling input and output data length relationship formula is:

[0094] (9)

[0095] In the formula, is the input data length, The length of the output data.

[0096] The formula for the relationship between the input and output data lengths of small-size convolution pooling is:

[0097] (10)

[0098] In the formula, is the filling length.

[0099] By comparing formulas (9) and (10), it can be found that the input and output data length relationship formulas of the standard pooling layer and the small-size convolution pooling are similar. When the input data length, pooling length and stride are kept the same, and zero padding is not set, the small-size convolution pooling and the standard pooling have the same output data dimension, that is, the pooling layer and the small-size convolution pooling have the same downsampling function.

[0100] For the construction of the SSCP-CNN model in this embodiment, a solution is constructed by replacing the pooling layer with a small-size convolutional pooling based on the one-dimensional CNN of the SSCP-CNN model, as shown in the following example. Figure 3 shown.

[0101] According to the method of this embodiment, the SSCP-CNN model includes 2 large convolution kernel convolution layers, 2 batch normalization layers (Batch Normalization, BN), 1 small size convolution pooling layer, 1 maximum pooling layer, 2 fully connected layers and 1 output layer. Compared with the one-dimensional CNN, the model is improved as follows:

[0102] (1) The pooling layer uses a combination of a small-size pooling layer + a maximum pooling layer instead of a combination of a maximum pooling layer + a maximum pooling layer. This comprehensively considers the values ​​of each neuron in the pooling area and assigns different weights to them, solving the problems of maximum pooling and average pooling and improving the accuracy.

[0103] (2) Two large convolution kernel convolution layers are used. First, the receptive field is expanded to extract more useful features and better capture the detailed information of the one-dimensional signal. Second, the amount of calculation is significantly reduced and the running time is shortened. Third, high-frequency noise can be suppressed and the accuracy of the model in noisy fault data can be improved.

[0104] (3) The structure adopts two convolutional layers and two pooling layers, which reduces the amount of parameter calculation and improves the model operation efficiency. At the same time, it avoids overfitting caused by too deep a network and improves the generalization ability of the model.

[0105] (4) SSCP-CNN directly uses the original one-dimensional vibration signal as input without the need for preprocessing such as time-frequency domain conversion or conversion of one-dimensional vibration data into two-dimensional images. In addition, to better fit the structural characteristics of one-dimensional vibration signals, the convolution kernel and pooling kernel of SSCP-CNN are both one-dimensional structures.

[0106] (5) Adding a BN layer after each convolutional layer keeps the input of each layer of the neural network identically distributed during the training process, solving the gradient vanishing and gradient exploding problems in the neural network and improving the generalization ability and accuracy of the model. The BN layer operation method is as follows:

[0107] (11)

[0108] In the formula, is the mean, is the standard deviation.

[0109] Step 3: Optimize the hyperparameters in the training set using the sparrow search algorithm;

[0110] Step 4: Calculate the loss function Loss and make the loss function Loss converge through multiple iterations;

[0111] Step 5: Use small-size convolution pooling to replace the first maximum pooling layer, and keep the SSCP-CNN and one-dimensional CNN with consistent hyperparameters to achieve hyperparameter transfer;

[0112] Step 6: Input the training set data into SSCP-CNN and use Adam to optimize the adaptive learning rate for model training;

[0113] Step 7: If the loss function convergence requirement is met, the SSCP-CNN training is completed, otherwise return to step 6;

[0114] Step 8: Use the trained SSCP-CNN to perform fault diagnosis on the test set, output the diagnosis results, calculate the performance indicators and perform visual analysis of the results.

[0115] It should be understood that the order of execution of the steps in the above embodiment does not necessarily mean the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiment of the present invention.

[0116] In one embodiment of the present invention, a rolling bearing fault diagnosis method based on SSCP-CNN is verified, and a possible embodiment is given below to illustrate its specific implementation scheme in a non-limiting manner.

[0117] The data set involved in this embodiment uses the open bearing database of Case Western Reserve University to verify the performance of SSCP-CNN. The acquisition system is driven by a three-phase induction motor, and the vibration signal of the faulty bearing is recorded by two accelerometers. The motor load of the system is 1hp, the speed is 1772r / min, and the data is collected at a sampling frequency of 12 KHz. The formation of the fault is to introduce single-point defects on the inner ring, outer ring and rolling element of the bearing through electrospark machining technology. The damage size of each place is 7 inches, 14 inches and 21 inches respectively. Together with the normal state, a total of 10 data samples of different states are generated, as shown in Table 1.

[0118] Table 1 Rolling bearing data set parameters

[0119]

[0120] For the SSCP-CNN parameter selection, the selection of hyperparameters will seriously affect the model performance. This embodiment adopts the sparrow search algorithm, which is inspired by the foraging behavior of sparrows and has the characteristics of strong optimization ability and fast convergence speed. The sparrow search algorithm is selected to optimize the hyperparameters of SSCP-CNN, and the optimization objects are the size and step size of the two convolution kernels and the size and step size of the two pooling layers.

[0121] Assume that the number of potential solutions in the search space is n, the variable dimension of the problem to be optimized is d , then the potential solution space is expressed as:

[0122] (13)

[0123] The corresponding fitness matrix F X for:

[0124] (14)

[0125] According to the SSA rule, the finder's position update formula is as follows:

[0126] (15)

[0127] In the formula, t is the current iteration number, is the maximum number of iterations, Indicates t The first iteration i A sparrow in j The location of the dimension, is a random number, Indicates the alarm value. represents the safety threshold, is a random number that follows a normal distribution, L Represents a 1×d matrix where all activation values ​​are 1.

[0128] The position update formula of the joiner is as follows:

[0129] (16)

[0130] In the formula, For the t +1 iteration finder occupies the best position, For the t The global worst position at iteration , Indicates t The first iteration i A sparrow in j dimensional position, A represents a 1×d matrix and each activation value in the matrix is ​​randomly assigned 1 or -1. .

[0131] When SSA optimizes the one-dimensional CNN, the initial population is 5, the warning value is 0.6, the proportion of discoverers is 0.7, the proportion of joiners is 0.3, and the proportion of sparrows aware of danger is 0.2. The optimization parameters and results are shown in Table 2. The size and step size of the first layer pooling can be converted to the size and step size of the small-size convolution pooling. In addition to the parameters optimized by SSA, the dropout is 0.1, the optimizer is Adam, the initial learning rate is set to 0.5, and it decreases to half of the original every 5 rounds. The total number of training rounds is 50, and the minimum training batch is 128.

[0132] Table 2 Optimization objects and results of the sparrow optimization algorithm

[0133]

[0134] The number of learnable parameters of the convolution and pooling layers of SSCP-CNN is shown in Table 3. The pooling layer has no learnable parameters, and using SSCP instead of the pooling layer increases the number of learnable parameters by 4128.

[0135] Table 3. Parameters of SSCP-CNN convolution and pooling layers

[0136]

[0137] This example verifies the effectiveness of the SSCP-CNN structure. Figure 5 Four CNN models shown. (a) Maximum pooling layer + maximum pooling layer combination, i.e., one-dimensional CNN; (b) Maximum pooling layer + small-size convolution pooling combination; (c) Small-size convolution pooling + maximum pooling layer combination, i.e., SSCP-CNN, (d) Small-size convolution pooling + small-size convolution pooling combination. The experiments were conducted in Matlab2021a, with an Intel(R) Core(TM) i7-7700HQ CPU, a processing frequency of 2.80GHz, and a GPU of GTX1060. Each set of experiments was performed 10 times to avoid random errors.

[0138] The test results of the standard data test are shown in Table 4 (the data in the table are the mean and variance after removing the maximum and minimum values ​​of the test results). It can be seen that SSCP+PooL achieved the highest accuracy of 98.47%, which is significantly better than PooL+SSCP and SSCP+SSCP, slightly better than one-dimensional CNN, and the variance is smaller than one-dimensional CNN, indicating that SSCP-CNN has more stable performance while maintaining a high accuracy rate and is more suitable for rolling bearing fault diagnosis.

[0139] Table 4 Accuracy and training time of four different convolutional pooling structures

[0140]

[0141] Although the combination of PooL+SSCP and SSCP+PooL is the same, its average accuracy is 7.40% lower than that of SSCP+PooL. The reason is that the early network contains more data features, and the accuracy is higher when SSCP is used as the first layer of the network pooling. In addition, SSCP+SSCP has the lowest average accuracy and the largest variance, indicating that SSCP+SSCP is relatively unstable. In the four-layer network architecture of two convolution layers and two pooling layers, it is an over-stacked SSCP, resulting in serious data overfitting and seriously low accuracy. From the above analysis, it can be seen that SSCP replacing maximum pooling can improve the average accuracy based on the fact that SSCP is in a suitable network architecture. The level at which SSCP exerts its comparative advantage is the network architecture, not the simple replacement of the pooling layer. If only SSCP is considered to replace the pooling layer without considering the position and role of SSCP in the network architecture, the accuracy may be quite different.

[0142] In order to compare the performance differences between SSCP-CNN and the current mainstream algorithms, five algorithms, including BPNN, CNN-1D (max pooling), CNN-1D (mixed pooling), denoising autoencoder (Denosising Autoencoder, DAE), and CNN-SVM, were selected for experiments. The results are shown in Table 5.

[0143] The BPNN network structure is relatively shallow, with insufficient data mining capabilities and limited accuracy. Based on CNN-1D, SSCP-CNN uses small-size convolution pooling with adaptive pooling steps to replace the original maximum pooling, which slightly increases the number of calculation parameters, and slightly increases the training time and accuracy. DAE and CNN-SVM use deep learning to achieve higher accuracy, but are slightly lower than SSCP-CNN in the fault diagnosis classification task. In addition, CNN-SVM is a hybrid model, and the time overhead of training the network is relatively large.

[0144] Table 5 Comparison of results of different algorithms

[0145]

[0146] In order to verify the generalization performance of the model in a noisy environment, Gaussian white noise is added to the standard dataset so that the data of the 10 label categories in Table 1 generate datasets with signal-to-noise ratios (SNR) of 10, 8, 6, 4, 2, 0, -2, -4, and -6 respectively.

[0147] The signal-to-noise ratio is defined as:

[0148] (17)

[0149] In the formula, is the power of the signal, is the power of the noise, and lg represents the logarithm with base 10.

[0150] Assuming the digital signal The sequence length is N , its signal power for:

[0151] (18)

[0152] An example of a noisy signal is Figure 6 As shown in the figure, it can be seen that as the noise intensity increases, the original data features are blurred by the noise to a greater extent, which will have an adverse effect on the accuracy of fault diagnosis.

[0153] The test results of the four CNN models are as follows: Figure 7 As shown. In the data with signal-to-noise ratios of 10 and 8, SSCP+PooL has obvious advantages over PooL+PooL of one-dimensional CNN, with the median increased by about 5%-8%; in the data with signal-to-noise ratios of 4 and 2, the median accuracy of the two is basically the same, but the lower bound of the accuracy of SSCP+PooL is still better than PooL+PooL; in the data with a signal-to-noise ratio of 6, the median and lower bound of the accuracy of the two are basically the same, but SSCP+PooL has multiple accuracies close to 100%, while PooL+PooL has abnormal values ​​with an accuracy of less than 80%. Especially under the high noise interference conditions of signal-to-noise ratios of -2, -4 and -6, the advantages of SSCP+PooL are more prominent, among which the median accuracy of experiments with signal-to-noise ratios of -4 and -6 is increased by about 10% compared with PooL+PooL. Therefore, SSCP+PooL has obvious advantages in processing noisy data, and its advantages are further expanded when the data is interfered by strong noise.

[0154] In this embodiment, the performance comparison of SSCP-CNN and other algorithms is as follows: Figure 8 As shown. It can be seen that with the increase of noise intensity, the accuracy of each algorithm shows a downward trend, but the accuracy of SSCP-CNN decreases relatively least, and it can still maintain an accuracy of more than 90% even when the SNR is -6. The accuracy of the comparison algorithm decreases seriously when the noise intensity is high. When the SNR is -6, the accuracy of BPNN is less than 50%, and the accuracy of CNN-1D (maximum pooling), CNN-1D (mixed pooling), DAE, CNN-SVM and other algorithms is between 70% and 80%, which is not suitable for fault diagnosis of noisy data. Therefore, SSCP-CNN has obvious advantages in fault diagnosis of noisy data.

[0155] From the experiment of this embodiment, it can be found that the PooL+PooL structure of the one-dimensional CNN and the SSCP+PooL structure of the SSCP-CNN are the structures with research value and application prospects among the four convolution pooling. Data with SNR of -6 were selected for the experiment, and the scatter distribution of the first-dimensional eigenvalues ​​and the second-dimensional eigenvalues ​​of the convolution pooling layers of the two structures were compared and analyzed by T-SNE (T-distributed Stochastic Neighbor Embedding) dimensionality reduction visualization.

[0156] Fig. 9 This is a visualization of the features of each layer of the PooL+PooL structure. It can be seen from the figure that after the first maximum pooling layer, the 2nd, 5th, 6th, and 8th categories overlap seriously, and it is still difficult to achieve complete classification. The remaining categories have clear intervals; after the second convolution layer, the good adaptive noise reduction of the convolution makes the categories that are difficult to divide in the first maximum pooling layer gather in the same category in space. However, there are still outliers in individual sample categories. For example, there are 5th and 6th category sample points distributed in the vicinity of the 8th category sample point (marked with red arrows in the figure). The fault diagnosis accuracy after the second convolution layer has been significantly improved compared with the first maximum pooling layer. However, the second maximum pooling layer disperses the spatial distribution of samples, and samples of the same category tend to be dispersed, resulting in a decrease in fault diagnosis accuracy.

[0157] Fig.10 Visualization of the features of each layer of the SSCP+PooL structure. It can be found from the figure that after small-size convolution pooling, the 2nd, 5th, 6th, and 8th categories have changed from serious overlap to a certain degree of similar gathering compared to the PooL+PooL structure, and no longer present a chaotic distribution in space, but the intervals between the 2nd and 8th categories are not clear and still need to be further divided. The present invention is based on the good adaptive denoising of noisy data by small-size convolution pooling. The various sample points in the second convolution layer and the maximum pooling layer are clearly separable, and there are no discrete points, which significantly improves the accuracy of fault diagnosis.

[0158] Therefore, it can be concluded from the feature visibility graph that in a four-layer CNN, when SSCP replaces the first maximum pooling layer of the PooL+PooL structure of a one-dimensional CNN, the diagnostic accuracy of noisy fault data can be improved and the spatial distribution of samples with clear intervals can be maintained.

[0159] In combination with the above methods, the small-scale convolution pooling in the method adopts an adaptive pooling step size and assigns different weights to each activation value in the pooling area, which can not only complete the downsampling function, but also comprehensively utilize all activation values ​​in the pooling area. The fault diagnosis experimental results of the bearing data set of Western Reserve University show that SSCP-CNN achieved an accuracy of 98.47% in standard data and maintained a high stability; in the noisy data with different noise intensities, the model can achieve the highest diagnostic accuracy, and when the noise intensity increases, the model advantage is further expanded.

[0160] The following is an embodiment of a rolling bearing fault diagnosis system based on SSCP-CNN provided in an embodiment of the present disclosure. The system and the rolling bearing fault diagnosis method based on SSCP-CNN in the above-mentioned embodiments belong to the same inventive concept. For details not described in detail in the embodiment of the rolling bearing fault diagnosis system based on SSCP-CNN, reference can be made to the embodiment of the rolling bearing fault diagnosis method based on SSCP-CNN.

[0161] The system includes: a data set processing module, a training set building module, a parameter search module, a loss function calculation module, a parameter transfer module, a model training module, a function state judgment module and a diagnostic test module;

[0162] The data set processing module is used to collect one-dimensional bearing vibration data, divide the one-dimensional bearing vibration data into training set and test set, and use One-hot encoding for labeling processing;

[0163] The training set construction module constructs a one-dimensional CNN with two convolutional layers and two pooling layers based on the SSCP-CNN model, and trains the data based on the one-dimensional CNN to form a training set;

[0164] The parameter search module optimizes the hyperparameters in the training set through the sparrow search algorithm;

[0165] The loss function calculation module is used to calculate the loss function Loss , and through multiple iterations, the loss function Loss convergence;

[0166] The parameter transfer module is used to replace the first maximum pooling layer with a small-size convolution pooling layer, and to keep the SSCP-CNN and one-dimensional CNN with consistent hyperparameters, thus realizing the transfer of hyperparameters.

[0167] The model training module is used to input the training set data into SSCP-CNN and use Adam to optimize the adaptive learning rate for model training;

[0168] The function state judgment module is used to judge whether the loss function meets the convergence requirements. If so, the SSCP-CNN training is completed;

[0169] The diagnostic test module is used to use the trained SSCP-CNN to perform fault diagnosis on the test set, output the diagnosis results, calculate the performance indicators and perform visual analysis of the results.

[0170] The rolling bearing fault diagnosis system of the present invention can adaptively pool the step size of the SSCP, which comprehensively considers all the values ​​in the pooling area and assigns different weights according to the convolution kernel, thereby solving the shortcomings of the maximum pooling fuzzy features and the average pooling equal weight. The minimum level at which the SSCP of the present invention can play a comparative advantage is the network architecture, rather than the simple replacement of the pooling layer. In the experiment of comparing and analyzing SSCP in different network architectures, SSCP-CNN achieved the highest accuracy among the four structures, and at the same time, the highest stability.

[0171] The fault diagnosis experiment shows that the SSCP-CNN of the present invention not only controls the training time within an acceptable range, but also has higher accuracy and stability. The model has more obvious advantages when processing noisy data, and the algorithm performance continues to expand with the increase of noise intensity.

[0172] The units and algorithm steps of each example described in the embodiments disclosed in the rolling bearing fault diagnosis system of the present invention can be implemented by electronic hardware, computer software or a combination of the two. In order to clearly illustrate the interchangeability of hardware and software, the composition and steps of each example have been generally described in the above description according to the function. Whether these functions are executed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of the present invention.

[0173] The rolling bearing fault diagnosis system of the present invention is a combination of the units and algorithm steps of each example described in the embodiments disclosed herein, and can be implemented by electronic hardware, computer software or a combination of the two. In order to clearly illustrate the interchangeability of hardware and software, the composition and steps of each example have been generally described in the above description according to function. Whether these functions are executed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of the present invention.

[0174] Those skilled in the art will appreciate that various aspects of the rolling bearing fault diagnosis method of the present invention can be implemented as a system, method or program product. Therefore, various aspects of the present disclosure can be specifically implemented in the following forms, namely: a complete hardware implementation, a complete software implementation (including firmware, microcode, etc.), or an implementation combining hardware and software aspects, which can be collectively referred to as "circuit", "module" or "system" here.

[0175] The above description of the disclosed embodiments enables one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A rolling bearing fault diagnosis method based on SSCP-CNN, characterized in that: Methods include: Step 1: Collect one-dimensional bearing vibration data, divide the one-dimensional bearing vibration data into training set and test set, and use One-hot encoding for labeling; Step 2: Based on the SSCP-CNN model, a one-dimensional CNN with two convolutional layers and two pooling layers is constructed, and the data is trained based on the one-dimensional CNN to form a training set; Step 3: Optimize the hyperparameters in the training set using the sparrow search algorithm; Step 4: Calculate the loss function Loss , and through multiple iterations, the loss function Loss convergence; Step 5: Use small-size convolution pooling to replace the first maximum pooling layer, and keep the SSCP-CNN and one-dimensional CNN with consistent hyperparameters to achieve hyperparameter transfer; The calculation formula for small-size convolution pooling is as follows: (5) In the formula, The first one is the small size convolution pooling i outputs, is the Relu activation function, n is the sample length, k is the length of the pooling region, r is the pooling step size, is the weight of the small-size convolution pooling, For the i A bias; Step 6: Input the training set data into SSCP-CNN and use Adam to optimize the adaptive learning rate for model training; Step 7: If the loss function convergence requirement is met, the SSCP-CNN training is completed, otherwise return to step 6; Step 8: Use the trained SSCP-CNN to perform fault diagnosis on the test set, output the diagnosis results, calculate the performance indicators and perform visual analysis of the results.

2. The rolling bearing fault diagnosis method based on SSCP-CNN according to claim 1 is characterized in that: The mathematical model of the convolutional layer operation in step 2 is described as: (1) In the formula, For the l Tier The output of the feature vector, is the input feature vector, For the The weight matrix of the convolution kernel of the layer, is the bias term, is the activation function used to perform nonlinear transformation on the output; After the convolutional layer, the rectified linear unit is used to enhance the nonlinear expression ability of the network. The mathematical expression of the rectified linear unit is as follows: (2) In the formula, is the output value of the convolutional layer, for Activation values ​​after the rectified linear unit.

3. The rolling bearing fault diagnosis method based on SSCP-CNN according to claim 2 is characterized in that: In step 2, the maximum pooling is used to select the largest value in the pooling area to represent the entire pooling area. The expression is as follows: (3) In the formula, For the l Tier i The feature vector t The value of a neuron, For the t The range of neurons for pooling, k is the length of the pooling region, is the output of the pooling layer.

4. The rolling bearing fault diagnosis method based on SSCP-CNN according to claim 1 or 2, characterized in that: In step 2, the mean value in the pooling area is selected based on average pooling to represent the entire pooling area. The mathematical expression is as follows: (4)。 5. The rolling bearing fault diagnosis method based on SSCP-CNN according to claim 1 or 2, characterized in that: The SSCP-CNN model in step 2 also contains: 2 large convolutional kernel convolutional layers, 2 batch normalization layers, 1 small-size convolutional pooling layer, 1 maximum pooling layer, 2 fully connected layers and 1 output layer.

6. The rolling bearing fault diagnosis method based on SSCP-CNN according to claim 5 is characterized in that: A BN layer is added after each convolutional layer to keep the input of each layer of the neural network in the same distribution during the training process. The BN layer operates as follows: (11) In the formula, is the mean, is the standard deviation, and the number of potential solutions in the search space is n.

7. The rolling bearing fault diagnosis method based on SSCP-CNN according to claim 1 or 2, characterized in that: In step 4, the loss function Loss The minimum mean square error is calculated using the following formula: (12) In the formula, Indicates i The predicted labels of samples, Indicates i The true labels of samples, N Indicates the sample size.

8. A rolling bearing fault diagnosis system based on SSCP-CNN, characterized in that: The system adopts the rolling bearing fault diagnosis method based on SSCP-CNN as described in any one of claims 1 to 6; The system includes: a data set processing module, a training set building module, a parameter search module, a loss function calculation module, a parameter transfer module, a model training module, a function state judgment module and a diagnostic test module; The data set processing module is used to collect one-dimensional bearing vibration data, divide the one-dimensional bearing vibration data into training set and test set, and use One-hot encoding for labeling processing; The training set construction module constructs a one-dimensional CNN with two convolutional layers and two pooling layers based on the SSCP-CNN model, and trains the data based on the one-dimensional CNN to form a training set; The parameter search module optimizes the hyperparameters in the training set through the sparrow search algorithm; The loss function calculation module is used to calculate the loss function Loss , and through multiple iterations, the loss function Loss convergence; The parameter transfer module is used to replace the first maximum pooling layer with a small-size convolution pooling layer, and to keep the SSCP-CNN and one-dimensional CNN with consistent hyperparameters, thus realizing the transfer of hyperparameters. The model training module is used to input the training set data into SSCP-CNN and use Adam to optimize the adaptive learning rate for model training; The function state judgment module is used to judge whether the loss function meets the convergence requirements. If so, the SSCP-CNN training is completed; The diagnostic test module is used to use the trained SSCP-CNN to perform fault diagnosis on the test set, output the diagnosis results, calculate the performance indicators and perform visual analysis of the results.

9. A diagnostic terminal, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the program, the steps of the rolling bearing fault diagnosis method based on SSCP-CNN as described in any one of claims 1 to 7 are implemented.

Images