A bearing fault detection system and method based on acoustic emission signals

Abstract

The present invention relates to a bearing fault detection system and method based on acoustic emission signals, which is characterized by comprising the following steps: an acoustic emission signal acquisition device collects the acoustic emission signals of the equipment bearing at a fixed point and sends them to a data server; the data server performs wavelet noise reduction processing on the collected acoustic emission signals, then performs standardized preprocessing operations on the data, and establishes a sliding rectangular window function, and performs data increment processing on the acoustic emission signals, and then makes the incremental data into a data set; a deep neural network model is established in an algorithm server, and the model is trained; the trained deep neural network model is deployed in a model deployment machine to perform bearing fault detection and early warning. The advantages of the present invention are: it can realize intelligent diagnosis and analysis of bearing faults of ball mills and magnetic separators, timely diagnose the level of bearing faults, achieve early prediction, flexibility and accuracy, timely early warning and notify operation and maintenance personnel to quickly deal with it, and improve equipment safety performance.

Description

Technical Field

[0001] The present invention belongs to the technical field of equipment fault detection, and in particular relates to a bearing fault detection system and method based on acoustic emission signals. Background Art

[0002] Ball mills and magnetic separators are key equipment in the production process of magnetite beneficiation plants. These two types of equipment have complex structures. Once a failure occurs, it will not only affect production, but in serious cases may cause equipment damage and shutdown, and even threaten the personal safety of employees. Production practice shows that the bearings of ball mills and magnetic separators are one of the main sources of equipment failure. More than 30% of rotating machinery failures are caused by bearing failures. When the equipment fails, the temperature rise, noise, and vibration of the equipment may be abnormal. At present, intelligent devices mainly include non-contact temperature measurement and contact vibration detection, etc., and lack the means to collect and analyze the acoustic emission signals of the equipment.

[0003] Acoustic emission signal detection technology is a dynamic non-destructive monitoring technology. The acoustic emission signal stimulated in the early stage of defect development of the object will be received. Compared with traditional detection and analysis methods, such as vibration analysis, oil sample analysis and magnetic leakage, acoustic emission signal detection technology has the characteristics of short response time, wide frequency band, easy sensor installation and high signal-to-noise ratio. Therefore, in the field of mechanical fault diagnosis research, acoustic emission signal detection technology has very broad application and development prospects.

[0004] With the rapid development of the field of artificial intelligence, more and more researchers have combined acoustic emission signal detection technology with machine learning or deep learning algorithms to propose intelligent fault diagnosis methods. In view of the existing contact acoustic emission signal detection methods and systems, the present invention adopts a non-contact detection method and system based on acoustic emission signals, which enhances the flexibility and safety of detection and is more sensitive to weak fault information that occurs in the early stage of bearings. Compared with existing LSTM and other model methods, the present invention simplifies the complexity and parameter amount of the model structure and reduces computational consumption while making full use of the sample's timing information to ensure an extremely high fault detection rate. Summary of the invention

[0005] In view of the problem that online monitoring of equipment bearings lacks multi-dimensional fault diagnosis, the purpose of the present invention is to provide a bearing fault detection system and method based on acoustic emission signals. By studying the acoustic emission signal characteristics of bearings during equipment operation, the state characteristics of the equipment under various operating conditions are explored, and a bearing fault detection method matching the acoustic emission signal is established to provide effective status warning.

[0006] The objective of the present invention is achieved through the following technical solutions:

[0007] The present invention provides a bearing fault detection system based on acoustic emission signals, characterized in that the fault detection system comprises an acoustic emission signal acquisition device fixedly arranged on the side of the ball mill and the magnetic separator bearing, a signal data transmission cable, a data server, a data transmission cable, an algorithm server, a model training machine, a model deployment machine and an early warning device,

[0008] The input end of the data server is electrically connected to the acoustic emission signal acquisition device through a signal data transmission line, and the output end is electrically connected to the input end of the algorithm server through a data transmission line. The output end of the algorithm server is electrically connected to the input end of the model training machine and the input end of the model deployment machine respectively, and the output end of the model deployment machine is electrically connected to the early warning device;

[0009] A bearing fault detection method based on acoustic emission signals of the present invention adopts the detection system according to claim 1, and is characterized in that it comprises the following steps:

[0010] A. The acoustic emission signal acquisition device collects the acoustic emission signal of the bearing to be tested and transmits it to the data server for storage;

[0011] B. Perform wavelet threshold noise reduction processing on the collected acoustic emission signals in the data server;

[0012] C. The data server performs standardization processing on the noise-reduced acoustic emission signal data;

[0013] D. Establishing a sliding rectangular window function in the data server to perform sliding data increment processing on the acoustic emission signal after the standardization processing, and obtaining the acoustic emission signal data sample after the increment;

[0014] E. In the data server, the fault type is used as a label to calibrate the acoustic emission signal data samples corresponding to the type and construct a data set;

[0015] F. In the data server, the data set is divided into training set, validation set and test set in proportion;

[0016] G. Establish a deep neural network model for bearing fault detection in the algorithm server;

[0017] H. Model training The deep neural network model in the machine control algorithm server is trained. The training set is used to train the model, the validation set is used to verify the model effect and guide the training of the model, and the test set is used to test the generalization ability of the model;

[0018] I. Deploy the trained deep neural network model in the model deployment machine;

[0019] When detecting bearing faults, the standardized acoustic emission signals obtained through the above steps A, B, and C are input into the neural network model in the model deployment machine for inference and prediction. When the predicted fault probability value corresponding to a certain fault state is greater than the warning threshold, an immediate warning is issued;

[0020] Preferably, in step B, the process of performing wavelet threshold denoising on the collected acoustic emission signals is as follows: First, perform wavelet decomposition on the original acoustic emission signal f(t) to obtain wavelet coefficients ω at each scale j,k ; By performing threshold processing on the wavelet coefficients ω j,k to obtain the processed wavelet coefficients Then, perform wavelet reconstruction on to obtain the denoised signal The principle for selecting the wavelet threshold is the fixed threshold principle: where L is the signal length and T is the fixed threshold;

[0021] Preferably, in step D, the method of performing sliding data increment processing on the acoustic emission signals is as follows: It is realized by establishing a sliding rectangular window function to slide and intercept the acoustic emission signals after wavelet denoising and normalization processing; Set a sliding rectangular window with a width of l, which slides and intercepts a sample sequence with the width of the sliding rectangular window from the zero moment of the acoustic emission signal at a certain step size s, (s < l), and each subsequent segment is used as a new sample sequence; Discard the last segment of data with a length less than the width of the sliding window, that is, round down the sample size; Considering the information coherence between adjacent intercepted signal segments, the sliding step size of the sliding rectangular window should be less than the width of the sliding rectangular window itself; At the same time, for the convenience of model construction and training and considering the characteristics of the computing hardware environment, the width of the sliding rectangular window is generally selected as l = 2 a The total number of samples after each acoustic emission signal sequence undergoes sliding data increment processing is:

[0022] Preferably, in step E, in the data server, the acoustic emission signal data after increment processing is used as a sample. The samples obtained by sequentially processing the acoustic emission signals corresponding to each bearing fault through steps A, B, C, and D are labeled one-to-one with the bearing fault type as the label of the sample, thereby establishing a data set.

[0023] Preferably, in step F, the data server divides the data set into a training set, a validation set, and a test set according to a ratio. The data set is divided into a training set, a validation set, and a test set according to the total number of samples in a ratio of 6:2:2.

[0024] Preferably, in step G, the main structure of the deep neural network model is composed of a one-dimensional convolutional network, a gated recurrent network and a multi-layer perceptron network. The output dimension of the deep neural network model is the number of fault categories n. Finally, after passing through the Softmax layer, the probability value of each fault category is obtained, and then the corresponding fault category is determined;

[0025] In the one-dimensional convolutional network of the deep neural network, the activation function of each layer is selected from the sigmoid function, the Tanh function, the ReLU function and the LeakyReLU function, the convolution kernel size of each convolutional layer is 2, and the step size is 1; to prevent gradient explosion or diffusion and increase the training speed, a BatchNormalization layer is introduced after each one-dimensional convolutional layer; at the same time, Dropout is introduced after each one-dimensional convolutional layer to prevent the occurrence of overfitting problems;

[0026] In the deep neural network, except for the last gated recurrent network layer, the parameter return_sequences of each gated recurrent network layer is set to True.

[0027] In step G, the main network structure of the deep neural network model is composed of a one-dimensional convolutional network (Conv1D), a gated recurrent network (GRU) and a multi-layer perceptron network. The output dimension of the deep neural network is the number of fault categories n. Finally, after passing through the Softmax layer, the probability value of each fault category is obtained, and then the corresponding fault category is determined;

[0028] In the one-dimensional convolutional network of the deep neural network, the activation function of each layer is selected from the sigmoid function, the Tanh function, the ReLU function and the LeakyReLU function, and the convolution kernel size of each convolutional layer is 2 and the step size is 1; in order to prevent gradient explosion or diffusion and increase the training speed, a batch normalization layer is introduced after each dimensional convolutional layer; at the same time, a random dropout is introduced after each one-dimensional convolutional layer to prevent the occurrence of overfitting problems;

[0029] In the deep neural network, except for the last gated recurrent network GRU layer, the parameter return_sequences of each GRU layer is set to True;

[0030] Preferably, in step H, when the deep neural network model is used for training, the network loss function can be selected from a cross entropy loss function and a mean square error loss function, and the optimizer can be selected from Adam, RMSProp, and AdaGrad.

[0031] In step H, in the process of using the training set to train the model, using the verification set to verify the model effect, and guiding the model training, if the model accuracy is difficult to meet the target requirements, the collection of bearing acoustic emission signals will be increased, and training samples will be made and added to the original training set, and the model will be retrained until the model effect reaches the ideal effect on the verification set.

[0032] Preferably, in step J, the deep neural network model deployed in the model deployment machine uses the acoustic emission signals of the ball mill and magnetic separator collected on-site in the grinding and magnetic workshop as the original signals as input data, and the data is obtained after wavelet threshold denoising and data standardization processing by the algorithm server; its output is the failure probability of eight types of ball mill and magnetic separator bearing failure types, and the probability value results are sorted from large to small, and the failure type with the largest failure probability is selected as the final output; if the failure probability corresponding to the output failure type is greater than the threshold, an early warning device is immediately used for early warning. Compared with the prior art, the advantages of the present invention are:

[0033] 1) The present invention adopts a non-contact detection method and system based on acoustic emission signals, which enhances the detection flexibility and safety and is more sensitive to weak fault information that occurs in the early stage of the bearing. In addition, the present invention makes full use of the timing information of the samples, while ensuring an extremely high fault detection rate, simplifies the complexity and parameter amount of the deep neural network model structure based on the characteristics of the acoustic emission signals of the bearing, and reduces the computational overhead.

[0034] 2) The present invention comprehensively uses advanced technologies such as artificial intelligence, Internet of Things sensing, acoustic vibration recognition, etc. to realize information perception, noise data analysis, and intelligent judgment of fault defects of ball mills and magnetic separators.

[0035] 3) The present invention can realize intelligent diagnosis and analysis of ball mill and magnetic separator bearing faults, timely warn without power outages and promptly notify operation and maintenance personnel to take quick measures, thereby improving equipment safety performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 It is a structural block diagram of the bearing fault detection system of the present invention;

[0037] Figure 2 It is a flowchart of the bearing fault detection method of the present invention;

[0038] Figure 3 A schematic diagram of the process of wavelet threshold denoising of the present invention;

[0039] Figure 4 It is the algorithm structure diagram of the present invention;

[0040] Figure 5 This is a block diagram of a deep neural network structure according to a specific embodiment of the present invention.

[0041] Figure 1 Among them, 1. Magnetic grinding workshop site, 2. Ball mill and magnetic separator, 3. Acoustic emission signal acquisition device, 4. Signal data transmission line, 5. Data server, 6. Data transmission line, 7. Algorithm server, 8. Model training machine, 9. Model deployment machine, 10. Early warning device, 11. Backend service system computer room. DETAILED DESCRIPTION

[0042] The present invention will be further described below in conjunction with the accompanying drawings and specific implementations.

[0043] Example

[0044] like Figure 1 As shown, a bearing fault detection system based on acoustic emission signals of the present invention is characterized in that the fault detection system includes an acoustic emission signal acquisition device 3, a signal data transmission line 4, a data server 5, a data transmission line 6, an algorithm server 7, a model training machine 8, a model deployment machine 9 and an early warning device 10, which are fixedly arranged on the bearing side of the ball mill and the magnetic separator 2 at the grinding and magnetic workshop site 1, wherein the data server 5, the data transmission line 6, the algorithm server 7, the model training machine 8, the model deployment machine 9 and the early warning device 10 are all arranged in the background service system room 11.

[0045] The input end of the data server 5 is electrically connected to the acoustic emission signal acquisition device 3 through a signal data transmission line 4, and the output end is electrically connected to the input end of the algorithm server 7 through a data transmission line 6. The output end of the algorithm server 7 is electrically connected to the input end of the model training machine 8 and the input end of the model deployment machine 9 respectively, and the output end of the model deployment machine 9 is electrically connected to the early warning device 10;

[0046] like Figure 2-Figure 5 As shown, a bearing fault detection method based on acoustic emission signals of the present invention adopts the above-mentioned bearing fault detection system and includes the following steps:

[0047] Step 1: Select an acoustic emission signal collection point and place an acoustic signal collection device 3 next to the ball mill or magnetic separator bearing at the collection point;

[0048] Step 2: The acoustic emission signal acquisition device 3 respectively acquires the acoustic emission signals when the equipment is working normally and when the bearing fails. A total of 8 types of bearing acoustic emission signals are acquired, of which 1 type is the acoustic emission signal when the equipment is working normally. The state type of the equipment when working normally is included in the 8 types of faults as a special fault type - no fault, and the remaining 7 types are acoustic emission signals when the 7 types of bearings fail; each type of acoustic emission signal acquires data at about 860,000 time points at a frequency of 48KHz; the acquired acoustic emission signal data is transmitted to the data server 5 and saved;

[0049] Step 3: In the data server, the acoustic emission signal of the bearing in step 2 is subjected to wavelet denoising. According to the difference in the wavelet coefficients of the acoustic signal emitted by the bearing vibration and the noise signal in the wavelet domain, their coefficients are processed by the threshold function to achieve the purpose of denoising. First, the original signal f(t) is subjected to wavelet decomposition, and the formula is as follows:

[0050]

[0051] in satisfy It is called the basic wavelet, s is the scale parameter, and when processing discretized data, s takes the value of 2 j , j∈Z; after wavelet decomposition, the wavelet coefficients ω of each scale are obtained j,k , by the wavelet coefficient ω j,k Perform threshold processing to obtain the processed wavelet coefficients And then By performing wavelet reconstruction, we can get the denoised signal The reconstruction formula is as follows:

[0052]

[0053] in The selection principle of wavelet threshold is fixed threshold principle:

[0054]

[0055] Where L is the signal length and T is the fixed threshold;

[0056] Step 4: Standardize the denoised acoustic emission signal, i.e., normalize it, to facilitate the subsequent training of the deep neural network model;

[0057] In this embodiment, since there are fewer acoustic emission signals for each type of fault, it is necessary to perform sliding data increment processing on the data of the acoustic emission signal; in the data server, a sliding rectangular window function with a window width of 1024 is established to perform data increment processing on the acoustic emission signal after the standardized processing, and the sliding step size is set to 1024×80%. According to the sample size calculation formula, each fault type has 1000 fault samples, and the feature dimension of each sample is 1024;

[0058] Step 5: In the data server, use the eight fault types as labels to calibrate the corresponding 1,000 fault samples; perform this operation on each fault type, calibrate the samples corresponding to the fault type, and thus construct a data set containing the eight fault types;

[0059] Step 6: In the data server, divide the data set into training set, validation set, and test set in a ratio of 6:2:2;

[0060] Step 7: In the algorithm server 7, a deep neural network model for bearing fault detection is established. The deep neural network model constructed in this embodiment is as follows: Figure 4 As shown in the figure; the deep neural network model consists of a one-dimensional convolution module (Conv1D) and a gated recurrent network module (GRU).

[0061] The data first passes through a one-dimensional convolution layer with 32 convolution kernels, a convolution kernel size of 2 and a stride of 1. To prevent gradient explosion or diffusion and speed up training, a batch normalization layer is introduced after the convolution layer; to prevent over-training, random dropout (Dropout) is introduced after the BN layer, and the dropout rate is set to 0.5; the activation function selects the linear rectification (ReLU) function; after the convolution layer, the data feature dimension changes from 1024×1 to 1024×32;

[0062] Immediately following the activation function are two layers of gated recurrent network modules. The first layer of gated recurrent network modules has parameters units set to 32 and return_sequences set to True, which means that the output of each gated recurrent network module is used as the input of the next layer of gated recurrent network modules. The second layer of gated recurrent network modules has parameters units set to 8 and return_sequences set to False.

[0063] Finally, the data is normalized (Softmax) and outputs a one-dimensional array. The array has 8 values, corresponding to the probability values ​​of 8 types of faults. The array values ​​are sorted from large to small, and the sorting of the 8 types of faults is recorded. Then the first value in the processed array is the probability value corresponding to the category with the highest probability of fault, and the corresponding fault category is the final result of the model prediction.

[0064] Step 8: The model training machine 8 controls the algorithm server 7 to train the deep neural network model for bearing fault detection using the training set. In this embodiment, the batch size of the training is set to 128, the loss function uses the cross entropy loss function, the optimizer uses the Adam optimizer, and the learning rate is set to 10 -3 ;

[0065] Step 9: After the model is trained, the model training machine 8 verifies the effect of the deep neural network model through the validation set. If the effect is not ideal, the deep neural network model is retrained; if the effect is still not ideal, a new data set is constructed again through steps 2-7, the new data set is merged with the original data set, the merged data set is re-divided through step 6, and the model is retrained, and this step is repeated until the verification effect of the deep neural network model in the validation set reaches the expected result;

[0066] Step 10: The model training machine 8 tests the deep neural network model obtained in step 8 through the test set to test the generalization ability of the deep neural network model. The deep neural network model with good generalization ability is more sensitive to the slight changes in the bearing fault and can better realize the function of equipment fault detection;

[0067] Step 11: deploying the deep neural network model in the model deployment machine 9;

[0068] pass Figure 1 As shown, the bearing acoustic emission signal is collected in real time and transmitted to the data server 5 for processing in steps 3 and 4, and then the processed acoustic emission signal is inferred and predicted using the deep neural network model in the model deployment machine 9 to obtain the current operating status of the ball mill and magnetic separator bearings; if a fault probability value corresponding to a certain fault state is detected to be greater than the warning threshold, a warning is immediately issued through the warning device 10.

[0069] In short, the present invention provides a bearing fault detection method based on acoustic emission signals, and its general steps are as follows: first, the acquisition device collects the acoustic emission signals of the ball mill and the magnetic separator at a fixed point, and sends them to the data server; the data server performs wavelet noise reduction on the collected acoustic emission signals; the data is subjected to standardized preprocessing operations and a sliding rectangular window function is established, and the acoustic emission time series signal is subjected to data increment processing; thereafter, the incremental data is used as a sample, and the corresponding bearing fault type is used as a label of the sample to produce an acoustic emission signal data set, and then the acoustic emission signal data set is divided into a training set, a validation set, and a test set in a ratio of 6:2:2; then, the algorithm server establishes a deep neural network model, and uses the training set to train the model, the validation set to verify the model effect, and the test set to test the model generalization ability; finally, the deep neural network model trained by the algorithm server is deployed on the model deployment machine to perform bearing fault detection and early warning.

Claims

1. A bearing fault detection method based on acoustic emission signals, The following steps are involved: A. The acoustic emission signal acquisition device collects the acoustic emission signal of the bearing to be tested and transmits it to the data server for storage; B. The data server performs wavelet threshold denoising on the collected acoustic emission signals; the process of performing wavelet threshold denoising on the collected acoustic emission signals is as follows: firstly, the original signal f(t) is decomposed by wavelet to obtain the wavelet coefficients ω of each scale. j,k , by the wavelet coefficient ω j,k Perform threshold processing to obtain the processed wavelet coefficients Then the wavelet coefficients By performing wavelet reconstruction, we can get the denoised signal The selection principle of wavelet threshold is fixed threshold principle: Where L is the signal length and T is the fixed threshold; C. The data server performs standardization processing on the noise-reduced acoustic emission signal; D. A sliding rectangular window function is established in the data server to perform sliding data increment processing on the standardized acoustic emission signal, obtaining the incremented acoustic emission signal data sample; the method for performing sliding data increment processing on the standardized acoustic emission signal is as follows: it is achieved by establishing a sliding rectangular window function to perform sliding interception on the vibration time series signal. A sliding rectangular window with a width of l is set, which slides and intercepts a sample sequence of the width of the sliding rectangular window starting from the zero moment of the acoustic emission signal at a certain step size s, where S < l. Each next segment is used as a new sample sequence, and the last segment with a length less than the width of the sliding window is discarded, that is, the sample quantity is rounded down. To ensure the information coherence between adjacent intercepted signal segments, the sliding step size of the sliding rectangular window should be less than the width of the sliding rectangular window itself. At the same time, for the convenience of model construction and training and considering the characteristics of the computing hardware environment, the width of the sliding rectangular window is selected as l = 2 a , the total number of samples after the sliding data increment processing for each acoustic emission time series signal sequence is: E. The data server uses the fault type as a label to calibrate the acoustic emission signal data sample corresponding to the type, and constructs a data set; the data server uses the acoustic emission signal data after incremental processing as a sample, which is a sample obtained by sequentially processing the acoustic emission signal corresponding to each bearing fault through steps A, B, C, and D, and calibrates it one by one with the bearing fault type as the label of the sample, thereby establishing a data set; F. The data server divides the data set into training set, validation set and test set in proportion; The data server divides the data set into training set, validation set and test set in proportion, that is, the data set is divided into training set, validation set and test set in proportion of 6:2:2 according to the total number of samples; G. Establish a deep neural network model for bearing fault detection in the algorithm server; the main structure of the deep neural network model is composed of a one-dimensional convolutional network, a gated recurrent network and a multi-layer perceptron network. The output dimension of the deep neural network model is the number of fault categories n. Finally, after passing through the Softmax layer, the probability value of each fault category is obtained, and then the corresponding fault category is determined; In the one-dimensional convolutional network of the deep neural network, the activation function of each layer is selected from the sigmoid function, the Tanh function, the ReLU function and the LeakyReLU function, the convolution kernel size of each convolutional layer is 2, and the step size is 1; in order to prevent gradient explosion or diffusion and increase the training speed, a BatchNormalization layer is introduced after each one-dimensional convolutional layer; at the same time, Dropout is introduced after each one-dimensional convolutional layer to prevent the occurrence of overfitting problems; In the deep neural network, except for the last gated recurrent network layer, the parameter return_sequences of each gated recurrent network layer is set to True; H. The model training machine controls the deep neural network model in the algorithm server for training, respectively uses the training set to train the model, uses the validation set to verify the model effect, and guides the model training, and uses the test set to test the model generalization ability; in the process of respectively using the training set to train the model, using the validation set to verify the model effect, and guiding the model training, if the model accuracy is difficult to meet the target requirements, the collection of bearing acoustic emission signals will be added, and the training samples will be made into training samples and added to the original training set, and the model will be retrained until the model effect reaches the ideal effect on the validation set; I. Deploy the trained deep neural network model in the model deployment machine; J. During bearing fault detection, the standardized acoustic emission signal obtained through the above steps A, B, and C is input into the neural network model in the model deployment machine for inference and prediction. When the fault probability value corresponding to a certain fault state is predicted to be greater than the warning threshold, an early warning is immediately issued; the deep neural network model deployed in the model deployment machine uses the acoustic emission signals of the ball mill and magnetic separator collected on-site in the grinding and magnetic workshop as the input data as the original signal, and the data is obtained after wavelet threshold denoising and data standardization processing by the algorithm server; the output is the fault probability of eight types of ball mill and magnetic separator bearing fault types, and the probability value results are sorted from large to small, and the fault type with the largest fault probability is selected as the final output; if the fault probability corresponding to the output fault type is greater than the threshold, an early warning is immediately issued through the early warning device.

2. A bearing fault detection method based on acoustic emission signals according to claim 1, It is characterized in that In step H, when using the deep neural network model for training, the network loss function selects one from the cross entropy loss function and the mean square error loss function, and the optimizer selects one from Adam, RMSProp, and AdaGrad.

3. A bearing fault detection system based on acoustic emission signals, according to the bearing fault detection method based on acoustic emission signals according to claim 1 or 2, wherein the fault detection system comprises an acoustic emission signal acquisition device fixedly arranged on the side of the ball mill and the magnetic separator bearing, a signal data transmission cable, a data server and a data transmission cable, It is characterized in that It also includes algorithm servers, model training machines, model deployment machines and early warning devices. The input end of the data server is electrically connected to the acoustic emission signal acquisition device through a signal data transmission line, the output end of the data server is electrically connected to the input end of the algorithm server through a data transmission line, the output end of the algorithm server is electrically connected to the input end of the model training machine and the input end of the model deployment machine respectively, and the output end of the model deployment machine is electrically connected to the early warning device.

Images