A rolling bearing fault diagnosis method and terminal based on temporal memory enhanced network

Abstract

The present invention provides a rolling bearing fault diagnosis method and terminal based on a time-series memory enhancement network, and relates to the technical field of rolling bearing fault detection. The present invention aims at the problem of large diagnostic errors when a CNN-LSTM network processes small sample noisy data. The fault features of the time series data are extracted adaptively by CNN, and then the processing capability of the time series noisy data is enhanced by recursive average filtering of the input data of the LSTM. The forget gate and the input gate of the LSTM unit are coupled again, and the two are respectively connected to the memory unit, thereby improving the memory capacity of the time series data, making it more suitable for the learning of small sample data. Finally, the Softmax function after the fully connected layer is used to realize the recognition of multi-bearing fault states. Experiments based on the rolling bearing data set of Case Western Reserve University show that the model has a high average accuracy rate in standard data and noisy data, and also has high diagnostic accuracy and generalization performance.

Description

Technical Field

[0001] The present invention relates to the technical field of rolling bearing fault detection, and in particular to a rolling bearing fault diagnosis method and terminal based on a sequential memory enhancement network. Background Art

[0002] Rolling bearings are key components for reducing friction losses between shafts, and detecting their health status is an important part of the health management of mechanical equipment throughout its life cycle. Statistics show that the failure rate of bearings is high, accounting for more than half of the failures of rotating mechanical equipment. Therefore, rolling bearing fault diagnosis is one of the research focuses in the field of fault diagnosis.

[0003] Early fault diagnosis is highly dependent on the professional knowledge and troubleshooting experience of experts and scholars. With the rapid development of artificial intelligence, fault diagnosis algorithms based on deep learning have become a research hotspot in recent years. Convolutional Neural Network (CNN) is one of the classic algorithms of deep learning. Early CNN fault diagnosis generally uses two-dimensional convolution to extract two-dimensional feature information. However, since one-dimensional convolutional neural network (1D-CNN) does not need to generate two-dimensional images when processing one-dimensional vibration signals of bearings, it significantly reduces the amount of calculation and is therefore widely used in bearing diagnosis.

[0004] The adoption of 1D-CNN reduces the computational complexity while achieving a bearing fault diagnosis accuracy of 97%. The prior art proposes a one-dimensional large-scale convolutional neural network 1DLCNN, and uses a genetic algorithm to optimize hyperparameters, achieving better diagnostic results than two-dimensional CNN in noisy bearing data. However, the diagnostic effect of CNN depends largely on the number of samples, and the accuracy drops significantly in small sample data. Long Short-Term Memory (LSTM) is a recurrent neural network with long-term memory function, which consists of an input gate, an output gate, a forget gate, and memory cells, and is widely used in time series learning. Studies have shown that enhancing the long-term memory capacity of the network is conducive to processing small sample data. Combining bidirectional LSTM with multi-scale weighted entropy morphological filtering plays a significant role in reconstructing the time-frequency domain features of rolling bearing fault characteristic signals (can be deleted). Another prior art proposes a stacked LSTM fault diagnosis algorithm, which automatically extracts data features by superimposing layers, and achieves good results in rolling bearing fault diagnosis.

[0005] However, when processing small sample noisy data, the error of the LSTM network will increase sharply with the increase in the number of iterations, which limits its application to a certain extent. Summary of the invention

[0006] The present invention provides a rolling bearing fault diagnosis method based on a temporal memory enhancement network, and the method is intended to solve the problem of large diagnosis error when a CNN-LSTM network processes small sample noisy data.

[0007] Methods include:

[0008] Step 1: Obtain the vibration signal of the rolling bearing;

[0009] The length of the constructed sample is The repeated sampling interval is A dataset of fixed segments of a time series;

[0010] The samples and labels of a fixed segment of the time series are ( X 1, X 2,…, X m )and( y 1, y 2,…, y m ), where the sample of the dataset X 1 is represented by ( x 1, x 2,…, x len ),sample X 2 is expressed as ( x 1+s , x 2+s ,…, x len+s ), and so on, sample X m Expressed as ( x 1+(m-1)s , x 2+(m-1)s ,…, x len+(m-1)s );

[0011] Step 2: Normalize the data set;

[0012] Step 3: Divide the dataset into two parts: training set and test set, and use One-hot encoding for labeling;

[0013] Step 4: Build a CNN-LSTM-TSME model. The input of the CNN-LSTM-TSME model is the sample of the time series, and the output is the classification of the time series.

[0014] Step 5: Use the grid search algorithm to determine the optimal initial model parameters for batch size and number of hidden units of the LSTM network;

[0015] Step 6: Use BPTT back propagation to train the CNN-LSTM-TSME model, extract feature information, and learn fault features;

[0016] Step 7: When the mean square error loss rate of the CNN-LSTM-TSME model tends to be stable, stop training and save the training weights of the CNN-LSTM-TSME model;

[0017] Step 8: Use the test set to test the performance of the trained CNN-LSTM-TSME model, calculate the evaluation index, output the calculation results, and end the calculation.

[0018] It should be further explained that Step 1 also includes: configuring the input sequence to ,in N is the data length;

[0019] The convolution kernel performs convolution operations from the first sample point of the data set to the last sample point, where j The feature map of a convolution kernel is:

[0020] (1)

[0021] In the formula, L is the size of the convolution kernel;

[0022] After the convolution operation is completed, the result Enter the Relu activation function to activate the neurons in the network.

[0023] It should be further explained that the Relu activation function expression is:

[0024] (2)

[0025] The output of the convolutional layer is:

[0026] (3)

[0027] In the formula, For the The input of the layer is also the The output of the layer; is the activation function; is the set of input mapping layers; For the The weight matrix of the layer convolution kernel; is the bias term.

[0028] It should be further explained that the fully connected layer connects all the output results of the pooling layer and outputs them to the classifier; the forward propagation output of the fully connected layer is

[0029] (4)

[0030] In the formula, For the +1 level Output neurons logits value.

[0031] It should be further explained that Step 1 also includes:

[0032] The training samples for overlapping sampling are N data points, and the length of a single sample after sampling is The repetition rate is γ , then the number of sampled training samples is m for:

[0033] (12)

[0034] in The floor symbol.

[0035] It should be further explained that the normalization processing method in step 2 uses the following formula of the maximum and minimum value normalization method to normalize all samples in the data set:

[0036] .

[0037] It should be further noted that the method also includes: improving the LSTM network by:

[0038] (1) The forget gate and the input gate are coupled. After coupling, the activation value of the input gate = 1-the activation value of the forget gate, so that the input value at the new moment directly affects the forgotten information state;

[0039] (2) Use recursive averaging filtering as The filter is set to have a length of The queue in which The arithmetic mean of the vibration data is taken as Time input, recorded as ; The queue data at the next moment will discard the head data and synchronously update the collected data as the tail data; the recursive average calculation formula is as follows:

[0040] (5)

[0041] (3) Add two new hole connections to enable the forget gate to accept -1 moment memory unit As input, the output gate can accept Moment Memory Unit As input;

[0042] The specific calculation formula for improving the LSTM network is as follows:

[0043] (6)

[0044] In the formula, is the activation value of the forget gate, is the activation function, is the weight vector of the forget gate, for -1 The state of the memory unit at time for -1 moment memory unit output, is the bias of the forget gate;

[0045] (7)

[0046] In the formula, is the activation value of the input gate, is the weight vector of the input gate, Bias for input gate input control;

[0047] (8)

[0048] In the formula, is the candidate state of the memory cell, is the weight vector of the candidate state of the input gate, is the bias of the candidate state of the input gate;

[0049] (9)

[0050] In the formula, for The state of the memory unit at each moment;

[0051] (10)

[0052] In the formula, is the activation value of the output gate, is the weight vector of the output gate, is the bias of the output gate;

[0053] (11)

[0054] In the formula, is the output of the memory unit, is the activation value of the output gate.

[0055] It should be further explained that in step 6, the CNN-LSTM-TSME model is trained. A batch normalization layer is added after each convolution layer of CNN-LSTM-TSME to split the training data into small batches, and the batch-normalized data enters the pooling layer.

[0056] Batch normalization is shown in the following formula:

[0057] (13)

[0058] In the formula, is the mean, is the standard deviation.

[0059] It should be further explained that in the method, a grid search strategy is used to optimize the hyperparameters of CNN-LSTM-TSME. The description of the grid search strategy is as follows:

[0060] (14)

[0061] In the formula, X 1 means X Optimal lower bounds for parameters; X M express X Optimizing upper bounds of parameters; Y 1 means Y Optimal lower bounds for parameters; Y N express Y Optimal upper bound for parameters.

[0062] The present invention also provides a 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 a temporal memory enhanced network are implemented.

[0063] It can be seen from the above technical solutions that the present invention has the following advantages:

[0064] The rolling bearing fault diagnosis method based on the time series memory enhanced network provided by the present invention aims at the problem that the diagnosis error is large when the CNN-LSTM network processes small sample noisy data. The fault characteristics of the time series data are adaptively extracted by CNN, and the processing capability of the time series noisy data is enhanced by recursive averaging filtering on the input data of the LSTM. The forget gate and the input gate of the LSTM unit are coupled again and the two are connected to the memory unit respectively, thereby improving the memory capacity of the time series data and being more suitable for the learning of small sample data. Finally, the Softmax function after the fully connected layer is used to realize the recognition of the fault status of multiple bearings.

[0065] The CNN-LSTM-TSME of the present invention recursively averages and filters the input data, which reduces the influence of noise data on the recognition accuracy, improves the accuracy of fault diagnosis of noisy data, and can still achieve a recognition accuracy higher than 95% under the condition of a signal-to-noise ratio of 2. CNN-LSTM-TSME couples the forget gate and the input gate, shortens the calculation time, and reduces the calculation cost; the forget gate and the input gate are respectively connected to the memory unit, which improves the memory capacity of long time series data, and enables the sample length to realize the time series memory enhancement function. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] 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.

[0067] Figure 1 The present invention is a flow chart of an embodiment of a rolling bearing fault diagnosis method based on a sequential memory enhancement network;

[0068] Figure 2 This is a schematic diagram of a standard LSTM unit;

[0069] Figure 3 To improve the schematic diagram of LSTM unit;

[0070] Figure 4 This is the CNN-LSTM-TSME network architecture diagram;

[0071] Figure 5 The data distribution diagram for different bearing conditions;

[0072] Figure 6 It is a schematic diagram of overlapping sampling of one-dimensional vibration signal;

[0073] Figure 7 This is a schematic diagram of the grid search strategy;

[0074] Figure 8 It is a schematic diagram of data with different signal-to-noise ratios; DETAILED DESCRIPTION

[0075] Compared with the prior art, NN-LSTM and other existing models, the rolling bearing fault diagnosis method based on the time series memory enhancement network provided by the present invention has achieved higher fault diagnosis accuracy and better generalization performance. Among them, the rolling bearing fault diagnosis method based on the time series memory enhancement network can acquire and process related data based on artificial intelligence technology. Among them, the method utilizes the CNN-LSTM-TSME model, CNN adaptation, LSTM network and Softmax function, and combines digital computers or digital computer-controlled machines to simulate, extend and expand human intelligence, so as to realize the identification of multi-bearing fault states.

[0076] The rolling bearing fault diagnosis method of the present invention also has a machine learning function, wherein the machine learning and deep learning in the method of the present invention generally include artificial neural networks, belief networks, reinforcement learning, transfer learning, inductive learning, and formula-based learning technologies.

[0077] The rolling bearing fault diagnosis method is applied to one or more terminals, wherein the 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.

[0078] The 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.

[0079] The terminal may also include network equipment and / or user equipment, wherein the network equipment includes, but is not limited to, a single network server, a server group consisting of multiple network servers, or a cloud consisting of a large number of hosts or network servers based on cloud computing.

[0080] The network where the 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.

[0081] 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.

[0082] See also Figure 1 The figure is a flow chart of a rolling bearing fault diagnosis method based on a temporal memory enhancement network in a specific embodiment, the method comprising:

[0083] Step 1: Obtain the vibration signal of the rolling bearing;

[0084] The length of the constructed sample is The repeated sampling interval is A dataset of fixed segments of a time series;

[0085] The samples and labels of a fixed segment of the time series are ( X 1, X 2,…, X m )and( y 1, y 2,…, y m ), where the sample of the dataset X 1 is represented by ( x 1, x 2,…, x len ),sample X 2 is expressed as ( x 1+s , x 2+s ,…, x len+s ), and so on, sample X m Expressed as ( x 1+(m-1)s , x 2+(m-1)s ,…, x len+(m-1)s );

[0086] Step 2: Normalize the data set;

[0087] Step 3: Divide the dataset into two parts: training set and test set, and use One-hot encoding for labeling;

[0088] Step 4: Build a CNN-LSTM-TSME model. The input of the CNN-LSTM-TSME model is the sample of the time series, and the output is the classification of the time series.

[0089] Step 5: Use the grid search algorithm to determine the optimal initial model parameters for batch size and number of hidden units of the LSTM network;

[0090] Step 6: Use BPTT back propagation to train the CNN-LSTM-TSME model, extract feature information, and learn fault features;

[0091] Step 7: When the mean square error loss rate of the CNN-LSTM-TSME model tends to be stable, stop training and save the training weights of the CNN-LSTM-TSME model;

[0092] Step 8: Use the test set to test the performance of the trained CNN-LSTM-TSME model, calculate the evaluation index, output the calculation results, and end the calculation.

[0093] In an exemplary embodiment, CNN is a multi-layer perceptron inspired by visual neural mechanisms and used to recognize images, etc., and mainly includes an input layer, a convolution layer, a pooling layer, and an output layer. 1D-CNN uses one-dimensional convolution layers and pooling layers and is suitable for processing time series, etc.

[0094] Assume the input sequence is ,in N is the data length. The convolution kernel performs convolution operations from the first sample point of the data set to the last sample point, where j The feature map of a convolution kernel is:

[0095] (1)

[0096] In the formula, L is the size of the convolution kernel.

[0097] After the convolution operation is completed, the result Enter the Relu activation function to activate the neurons in the network.

[0098] The Relu activation function expression is:

[0099] (2)

[0100] The output of the convolutional layer is:

[0101] (3)

[0102] In the formula, For the The input of the layer is also the The output of the layer; is the activation function; is the set of input mapping layers; For the The weight matrix of the layer convolution kernel; is the bias term.

[0103] The pooling layer reduces the dimension of the output of the convolutional layer to extract features. Common methods include maximum pooling and average pooling.

[0104] The fully connected layer connects all the output results of the pooling layer and outputs them to the classifier. The forward propagation output of the fully connected layer is:

[0105] (4)

[0106] In the formula, For the +1 level Output neurons logits value.

[0107] The present invention also relates to the improvement of the temporal memory enhancement of LSTM, such as Figure 2 and Figure 3 As shown, the LSTM unit uses input gate, output gate and forget gate to control the selective flow of information.

[0108] First, the input data is partially deleted after passing through the forget gate, and the input at the current moment is read. , the output at the previous moment , weight matrix and bias vector, and control the output to [0,1] through the activation function, where "0" means "completely discarded" and "1" means "completely retained".

[0109] Then, the data is transferred to the input gate to receive the current time enter And the output of the previous moment , thereby determining the new information in the cell state. Finally, the output gate outputs based on the cell state and updates the timing of the cell state.

[0110] The present invention improves the LSTM unit, such as Figure 2 As shown, the changes relative to the standard LSTM are as follows:

[0111] (1) Couple the forget gate and the input gate. Considering that the forgotten information and the retained information in the information flow are relative, the forget gate and the input gate are coupled.

[0112] After coupling, the activation value of the input gate = 1-the activation value of the forget gate, so that the input value at the new moment directly affects the forgotten information state, reducing the amount of calculation parameters.

[0113] (2) Recursive average filtering .

[0114] Considering that recursive filtering has a good effect of suppressing and weakening periodic interference in noisy data and has high smoothness, recursive averaging filtering is selected as filter.

[0115] Set the length to The queue in which The arithmetic mean of the vibration data is taken as Time input, recorded as At the next moment, the queue data will discard the head data and synchronously update the collected data as the tail data to ensure that the queue length is fixed. The recursive average calculation formula is as follows:

[0116] (5)

[0117] (3) Add two new hole connections.

[0118] In order to make the model better maintain long-term memory and improve its long-term memory ability, two new hole connections are added so that the forget gate can accept -1 moment memory unit As input, the output gate can accept Moment Memory Unit As input, temporal memory enhancement is achieved, which plays a key role in subsequent data processing.

[0119] The specific calculation formula of the improved LSTM network is as follows:

[0120] (6)

[0121] In the formula, is the activation value of the forget gate, is the activation function, is the weight vector of the forget gate, for -1 The state of the memory unit at time for -1 moment memory unit output, is the bias of the forget gate.

[0122] (7)

[0123] In the formula, is the activation value of the input gate, is the weight vector of the input gate, Bias for the input gate input control.

[0124] (8)

[0125] In the formula, is the candidate state of the memory cell, is the weight vector of the candidate state of the input gate, is the bias of the candidate state of the input gate.

[0126] (9)

[0127] In the formula, for The state of the memory unit at all times.

[0128] (10)

[0129] In the formula, is the activation value of the output gate, is the weight vector of the output gate, is the bias of the output gate.

[0130] (11)

[0131] In the formula, is the output of the memory unit, is the activation value of the output gate.

[0132] The CNN-LSTM-TSME network structure is as follows Figure 4 As shown, the CNN-LSTM-TSME network of the present invention involves a one-dimensional large convolution kernel CNN, which can effectively suppress high-frequency noise in the data, but the classification error is large when processing small sample data, and the LSTM network enhanced by temporal memory improves the classification accuracy of small sample data. Therefore, the combination of the two networks can process both data interfered by high-frequency noise and data of small samples.

[0133] The vibration signal first enters the feature extraction layer composed of convolutional layers and pooling layers to extract fault features and reduce dimensionality, thereby generating several feature maps. The signal is then transmitted to the LSTM layer to be divided according to the time axis. Finally, after the signal passes through two fully connected layers, the output layer uses the Softmax activation function to classify and output multiple bearing states.

[0134] The present invention also conducts experiments based on the rolling bearing fault diagnosis method to verify the beneficial effects of the present invention.

[0135] The experimental data set of the present invention comes from the Rolling Bearing Data Center of Case Western Reserve University (CWRU). The data acquisition system consists of an induction motor, a torque sensor, a dynamometer and related control electronics. The motor load of the system is 1hp, the speed is 1772r / min, and the data sampling frequency is 12kHz. The fault setting is completed by introducing single-point defects on the inner ring, outer ring and rolling element of the bearing using electrospark machining technology. The defect sizes are 7 inches, 14 inches and 21 inches. The data distribution of different bearing states is different. Examples of sampling data for normal state and three fault states of inner ring, outer ring and rolling element (defect size is 7 inches) are shown in Figure 5 .

[0136] Overlapping sampling in the experimental process is a key step in data preprocessing and an important method to reduce irrelevant features in small sample data and thus improve the accuracy of fault identification.

[0137] Overlap sampling can overcome the problem of insufficient data volume, which not only reduces the complexity of sample data processing, but also retains the key information in the original data, making the fault sample data points as fully utilized as possible. Assume that the training samples of overlapping sampling are N data points, and the length of a single sample after sampling is The repetition rate is γ , then the number of training samples that can be sampled is m for

[0138] (12)

[0139] in The floor symbol.

[0140] The overlapping sampling process of the one-dimensional vibration signal of the rolling bearing is as follows: Figure 6 As shown in Table 1, the parameters of the sampled dataset are shown in Table 1.

[0141] Table 1 Rolling bearing data set parameters

[0142]

[0143] During the model training process, each iterative training will update parameters such as weights, causing the data distribution to change. In order to adapt to this change, relevant parameters need to be continuously adjusted, which greatly increases the training overhead.

[0144] Especially when the number of model layers increases, the resulting gradient explosion problem becomes more prominent. In order to reduce the impact of data distribution changes on network training and keep the input of each layer of the network identically distributed, Ioffe and Szgedy proposed batch normalization (Batch Normalization, BN). This method divides the data into small batches and normalizes each layer during its forward pass, achieving adaptive reparameterization. The batch normalization method is shown in the following formula:

[0145] (13)

[0146] In the formula, is the mean, is the standard deviation.

[0147] Therefore, batch normalization can unify the distribution of data, thereby improving the overall performance of the model and making the network robust to poor initialization of weights. CNN-LSTM-TSME adds a batch normalization layer after each convolutional layer to split the data into small batches, and the batch-normalized data enters the pooling layer.

[0148] The CNN-LSTM-TSME of the present invention has many hyperparameters, and the selection of hyperparameters will directly affect the accuracy and robustness of the model. In order to better select the batch size and the number of LSTM hidden units to improve the fault diagnosis accuracy of the model, a grid search strategy is used to optimize the hyperparameters. The specific strategy is as follows: Figure 7 , some hyper parameters are shown in Table 2.

[0149] The grid search strategy can be described as:

[0150] (14)

[0151] In the formula, X 1 means X Optimal lower bounds for parameters; X M express X Optimizing upper bounds of parameters; Y 1 means Y Optimal lower bounds for parameters; Y N express Y Optimal upper bound for parameters.

[0152] Table 2 CNN-LSTM-TSME model parameters

[0153]

[0154] The number of network training iterations was 450, and the Adam algorithm was used to adaptively optimize the learning rate, with an initial learning rate of 0.1.

[0155] The experiments were conducted on Matlab 2021a, with an Intel(R) Core(TM) i7-7700HQ processor, a processing frequency of 2.80GHz, and a GPU of GTX1060.

[0156] The value set of batch size and LSTM hidden unit number is {32, 64, 128, 256, 512}. Divide the grid, calculate the accuracy on all grid points, and use the grid points under the optimal solution as the model parameter value. The average training time is shown in Table 3.

[0157] It can be seen that when the batch size of the model is 256 and the number of hidden units is 128, the average accuracy reaches the highest 99.65%, and the average training time is 23.44s. Relatively speaking, the average accuracy of the batch size of 64, the number of hidden units is 128, and the batch size of 128, the number of hidden units is 128, and the average accuracy is 99.60% and 99.42% respectively, but their training time is obviously longer. Analysis of the reasons, since the batch size determines the scale of a single training, the appropriate batch size can improve the accuracy while ensuring the convergence speed. When it is too small, it will not only lead to a long training time, but also prone to over-learning problems, resulting in a decrease in the accuracy of the model. When it is too large, although the training speed is accelerated, the model may not be able to fully utilize the training data, and under-learning problems will occur, resulting in poor generalization ability of the model; within a certain range, the more hidden units there are, the more complete the learning of the time series data features, and the higher the accuracy, but too high a number of hidden units will generate a training burden, resulting in over-learning and increasing model recognition errors.

[0158] Table 3 Average training time under different batch sizes and numbers of hidden units

[0159]

[0160] Based on the above experimental method, the following experimental results can be obtained. Specifically, according to the parameter setting of the present invention, 1D-CNN, LSTM, 1D-CNN-LSTM, and CNN-LSTM-TSME are used for comparative tests, and the results are shown in Table 4 (to reduce the randomness of the experimental results, each test is repeated 10 times).

[0161] It can be seen that the accuracy of CNN-LSTM-TSME is as high as 99.65%, and it has better performance than other models on small sample data.

[0162] Specifically, the networks of 1D-CNN and LSTM are shallower than 1D-CNN-LSTM, and their data feature extraction and classification capabilities are weaker; Spark-IRFA random forest has randomness, and the interference is more obvious in small sample data, and its accuracy is not as good as that of large sample data; the use of artificial extraction of data time and frequency domain features does not give full play to CNN's powerful nonlinear learning ability for data, so its accuracy is lower than that of CNN-LSTM-TSME.

[0163] Table 4 Comparison of diagnostic results of different algorithms

[0164]

[0165] The present invention visualizes fault classification. The T-SNE (T-Distributed StochasticNeighbor Embedding, T-SNE) algorithm is used to visualize high-dimensional data. After the first convolution layer, a two-dimensional scatter plot is formed, and after the second convolution layer, a two-dimensional scatter plot is formed. It can be seen that after the sample data passes through the first convolution layer, the same category tends to gather together, but the interval between heterogeneous samples is too small, and complete classification is not achieved.

[0166] After the second layer of convolution, samples of the same type are further clustered and shrunk, and the intervals between samples of different types are obvious, and the classification effect is good, which indirectly verifies the effectiveness of the model.

[0167] The present invention also analyzes the generalization performance in a noisy environment. The main sources of bearing data noise are: sensor error, mechanical and electrical equipment vibration, etc. In order to verify the generalization performance of the model in a noisy environment, Gaussian white noise is added to the data set and a fault diagnosis experiment is conducted. For the convenience of comparative analysis, the CNN-LSTM-TSME model keeps the original parameter settings unchanged.

[0168] The signal-to-noise ratio is defined as:

[0169] (15)

[0170] in is the power of the signal, is the power of the noise, lg Represents the base 10 logarithm.

[0171] Assuming the digital signal The sequence length is N , its signal power for:

[0172] (16)

[0173] After calculation, the signal powers of 10 bearing states after overlapping sampling are shown in Table 5.

[0174] Table 5 Signal power of 10 bearing statuses

[0175]

[0176] The signal-to-noise ratios are set to 10, 8, 6, 4, and 2, respectively. Figure 8 The figure is a schematic diagram of rolling element failure data under 6 different noise conditions. The average accuracy of each model under different signal-to-noise ratios is shown in Table 6.

[0177] It can be seen that the CNN-LSTM-TSME model has a high recognition accuracy for data with medium and weak signal-to-noise ratios. In the three types of fault data with signal-to-noise ratios of 10, 8, and 6, it achieved the highest average accuracy among the six comparison models, which were 99.31%, 99.04%, and 98.26%, respectively. In the fault data with a signal-to-noise ratio of 4, an average accuracy of 97.30% was achieved, which is 0.16 percentage points better than the CNN-LSTM-TSME model.

[0178] However, in the fault data with a signal-to-noise ratio of 2, the CNN-LSTM-TSME model achieved the highest average accuracy of 95.93%. The overall analysis shows that the CNN-LSTM-TSME model has good generalization performance in noisy data.

[0179] Table 6 Identification accuracy of bearing fault locations under Gaussian white noise with different signal-to-noise ratios

[0180]

[0181] The rolling bearing fault diagnosis method based on the above-mentioned temporal memory enhanced network can solve the problem that the CNN-LSTM network is not suitable for small sample noise data. The CNN-LSTM-TSME fault diagnosis model has the ability to extract features and process temporal noisy data. It can directly and adaptively extract sample features from one-dimensional vibration signals to realize an end-to-end classification mode.

[0182] The following effects are obtained by the method of the present invention:

[0183] (1) CNN-LSTM-TSME recursively averages the input data to reduce the impact of noise data on recognition accuracy and improve the accuracy of fault diagnosis of noisy data. Under the condition of a signal-to-noise ratio of 2, it can still achieve a recognition accuracy of more than 95%.

[0184] (2) CNN-LSTM-TSME couples the forget gate and the input gate, shortening the calculation time and reducing the calculation overhead; the forget gate and the input gate are connected to the memory unit respectively, which improves the memory capacity of long time series data and enables the sample length to achieve the time series memory enhancement function.

[0185] The units and algorithm steps of each example described in the embodiment disclosed in the rolling bearing fault diagnosis method based on the time-series memory enhanced network 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.

[0186] In embodiments of the present invention, computer program code for performing operations of the present disclosure may be written in one or more programming languages ​​or combinations thereof, including, but not limited to, object-oriented programming languages ​​such as Java, Smalltalk, C++, and conventional procedural programming languages ​​such as "C" or similar programming languages. The program code may be executed entirely on the user's computer, partially on the user's computer, as a stand-alone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer.

[0187] 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 a temporal memory enhancement network, characterized in that the method include: Step 1: Obtain the vibration signal of the rolling bearing; Construct a data set of fixed segments of time series with sample length len and repeated sampling interval s; The samples and labels of a fixed segment of the time series are (X1, X2, …, Xm) and (y1, y2, …, ym) respectively, where the sample X1 of the data set is represented by (x1, x2, …, xlen), the sample X2 is represented by (x1+s, x2+s, …, xlen+s), and so on, the sample Xm is represented by (x1+(m-1)s, x2+(m-1)s, …, xlen+(m-1)s); Step 2: Normalize the data set; Step 3: Divide the dataset into two parts: training set and test set, and use One-hot encoding for labeling; Step 4: Build a CNN-LSTM-TSME model. The input of the CNN-LSTM-TSME model is the sample of the time series, and the output is the classification of the time series. Step 5: Use the grid search algorithm to determine the optimal initial model parameters for batch size and number of hidden units of the LSTM network; Step 6: Use BPTT back propagation to train the CNN-LSTM-TSME model, extract feature information, and learn fault features; Step 7: When the mean square error loss rate of the CNN-LSTM-TSME model tends to be stable, stop training and save the training weights of the CNN-LSTM-TSME model; Step 8: Use the test set to test the performance of the trained CNN-LSTM-TSME model, calculate the evaluation index, output the calculation results, and end the calculation; Step 1 also includes: configuring the input sequence to be X = [x1, x2, ..., x N ], where N is the data length; The convolution kernel performs convolution operation from the first sample point of the data set to the last sample point, where the feature map of the jth convolution kernel is: Where L is the size of the convolution kernel; After the convolution operation is completed, the result Input the Relu activation function to activate the neurons in the network; The Relu activation function expression is: The output of the convolutional layer is: In the formula, is the input of the lth layer and also the output of the l-1th layer; σ is the activation function; M j is the set of input mapping layers; is the weight matrix of the convolution kernel of the lth layer; is the bias term; The fully connected layer connects all the output results of the pooling layer and outputs them to the classifier; the forward propagation output of the fully connected layer is In the formula, x l+1(j) is the logits value of the j-th output neuron in the l+1th layer; Step 1 also includes: Assume that the training samples of overlapping sampling have N data points, the length of a single sample after sampling is len, and the repetition rate is γ, then the number of sampled training samples m is: in is the floor rounding symbol; The method also includes: improving the LSTM network by: (1) The forget gate and the input gate are coupled. After coupling, the activation value of the input gate = 1-the activation value of the forget gate, so that the input value at the new moment directly affects the forgotten information state; (2) Select recursive averaging filtering as x t The filter is set to a queue of length len, and the arithmetic mean of the len vibration data in the queue is used as the input at time t, denoted as Filter(x t ); the next moment, the queue data will discard the head data and synchronously update the collected data as the tail data; the recursive average calculation formula is as follows: (3) Add two new hole connections so that the forget gate can accept the memory unit C at time t-1 t-1 As input, the output gate can accept the memory unit C at time t t As input; The specific calculation formula for improving the LSTM network is as follows: f t =σ(W f *[c t-1 ,h t-1 ,Filter(x t )]+b f ) (6) In the formula, f t is the activation value of the forget gate, σ is the activation function, W f is the weight vector of the forget gate, C t-1 is the state of the memory unit at time t-1, h t-1 is the memory unit output at time t-1, b f is the bias of the forget gate; i t =σ(W i *[h t-1 ,Filter(x t )]+b i ) (7) In the formula, i t is the activation value of the input gate, W i is the weight vector of the input gate, b i Bias for input gate input control; g t =tanh(W c *[h t-1 ,Filter(x t )]+b c ) (8) In the formula, g t is the candidate state of the memory unit, W c is the weight vector of the candidate state of the input gate, b c is the bias of the candidate state of the input gate; C t =f t *C t-1 +(1-f t )*g t (9) In the formula, C t is the state of the memory unit at time t; o t =σ(W o *[c t ,h t-1 ,Filter(x t )]+b o ) (10) In the formula, o t is the activation value of the output gate, W o is the weight vector of the output gate, b o is the bias of the output gate; h t =o t *tanh(C t ) (11) In the formula, h t is the output of the memory unit, o t is the activation value of the output gate; In step 6, the CNN-LSTM-TSME model is trained. A batch normalization layer is added after each convolutional layer of CNN-LSTM-TSME to split the training data into small batches, and the batch-normalized data enters the pooling layer. Batch normalization is shown in the following formula: In the formula, is the mean, x var is the standard deviation.

2. The rolling bearing fault diagnosis method based on the temporal memory enhancement network according to claim 1 is characterized in that: The normalization method in step 2 uses the following formula of the maximum and minimum normalization method to normalize all samples in the data set: X'=(x-x min ) / (x max -x min )。 3. The rolling bearing fault diagnosis method based on the temporal memory enhancement network according to claim 1 is characterized in that: In the method, the grid search strategy is used to optimize the hyperparameters of CNN-LSTM-TSME. The grid search strategy is described as follows: minf(x)X1≤X≤X M ,Y1≤Y≤Y N (14) Where X1 represents the optimization lower bound of the X parameter; XM represents the optimization upper bound of the X parameter; Y1 represents the optimization lower bound of the Y parameter; YN represents the optimization upper bound of the Y parameter.

4. A 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 the temporal memory enhancement network as described in any one of claims 1 to 3 are implemented.

Images