A bearing fault diagnosis method, device, terminal equipment and storage medium based on neural network

Abstract

The present invention discloses a bearing fault diagnosis method, device, terminal equipment and storage medium based on a neural network. The method comprises: first obtaining a bearing vibration signal of a bearing to be diagnosed in a preset time period; then extracting the bearing vibration signal according to a preset time window length, merging the extracted bearing vibration signals to obtain a bearing time series signal; then inputting the bearing time series signal into a preset first bearing fault diagnosis classifier constructed based on a neural network, so that the preset first bearing fault diagnosis classifier obtains the confidence corresponding to each bearing time series signal according to the bearing time series signal; finally, the result corresponding to the maximum confidence is used as the fault diagnosis result of the bearing to be diagnosed. By implementing the present invention, the specific fault type of the bearing can be identified.

Description

Technical Field

[0001] The present invention relates to the technical field of fault diagnosis, and in particular to a bearing fault diagnosis method, device, terminal equipment and storage medium based on a neural network. Background Art

[0002] Bearing fault diagnosis is to monitor and analyze the vibration, sound, temperature and other parameters of the bearing to determine whether the bearing is faulty, evaluate the severity of the fault and predict the development trend of the fault. Types of bearing faults include: inner ring damage, outer ring damage, ball damage, etc. By regularly monitoring and analyzing the bearing, signs of bearing failure can be discovered in time to avoid further development of the failure, reduce equipment downtime and maintenance costs. In addition, bearing failure may cause equipment downtime and accidents, threatening the normal operation of the production line and the safety of workers. Through timely fault diagnosis, accidents can be avoided and production safety can be guaranteed.

[0003] In the prior art, an autoencoder is usually used to detect bearing faults. However, the fault detection algorithm of this autoencoder can only determine whether the bearing has a fault, but cannot identify the specific fault type. Summary of the invention

[0004] The present invention provides a bearing fault diagnosis method, device, terminal equipment and storage medium based on a neural network, which can identify the specific fault type occurring in the bearing and improve the fault classification capability by deriving fault signals when fault data is scarce.

[0005] An embodiment of the present invention provides a bearing fault diagnosis method based on a neural network, comprising:

[0006] Acquire a bearing vibration signal of the bearing to be diagnosed in a preset time period; wherein the vibration signal includes: a horizontal vibration signal and a vertical vibration signal;

[0007] Extract the bearing vibration signal according to a preset time window length, merge the extracted bearing vibration signals, and obtain a bearing time series signal;

[0008] The bearing time series signal is input into a preset first bearing fault diagnosis classifier constructed based on a neural network, so that the preset first bearing fault diagnosis classifier obtains the confidence corresponding to each bearing time series signal according to the bearing time series signal; wherein the confidence includes: the confidence of no fault and the confidence of each fault type;

[0009] The result corresponding to the maximum confidence degree among the above confidence degrees is used as the fault diagnosis result of the above bearing to be diagnosed.

[0010] Furthermore, the training of the preset first bearing fault diagnosis classifier includes:

[0011] Acquire a first bearing time series signal when the bearing is in a normal state within a first preset time period;

[0012] Randomly collect a number of one-dimensional latent vectors from a normal distribution to obtain a random sampling signal;

[0013] For each random sampling signal, the random sampling signal is input into a corresponding preset decoder to generate a derived fault time series signal, and then a plurality of derived fault time series signals and a fault type corresponding to each derived fault time series signal are obtained;

[0014] Input the above-mentioned several derivative fault time series signals with the first true label and the above-mentioned first bearing time series signal into the first bearing fault diagnosis classifier to be trained, and obtain the confidence of each derivative fault time series signal and the above-mentioned first bearing time series signal being judged as fault-free and the confidence of each fault type judged by the first bearing fault diagnosis classifier to be trained; wherein the above-mentioned first true label is: the fault type label of the above-mentioned derivative fault time series signal and the fault-free label of the above-mentioned first bearing time series signal;

[0015] For each confidence level, calculating a first cross entropy loss function between the confidence level and the first true label, and determining whether the first cross entropy loss function converges;

[0016] If the above-mentioned first cross entropy loss function converges, it is determined that the training of the above-mentioned first bearing fault diagnosis classifier is completed, and the above-mentioned preset first bearing fault diagnosis classifier is obtained; if the above-mentioned first cross entropy loss function does not converge, adjust the weights of each neural network in the above-mentioned first bearing fault diagnosis classifier, and continue to train the above-mentioned first bearing fault diagnosis classifier.

[0017] Furthermore, the training of the above preset decoder includes:

[0018] Acquire a second bearing time series signal of the bearing under each fault type within a first preset time period;

[0019] For each fault type, each second bearing time series signal with a second true label and the first bearing time series signal are respectively input into the corresponding encoder after initial training for encoding, so as to obtain the second feature signal corresponding to each of the second bearing time series signals and the first feature signal corresponding to the first bearing time series signal; wherein the second true label is: the fault type label corresponding to each of the second bearing time series signals and the fault-free label corresponding to the first bearing time series signal;

[0020] Inputting the first characteristic signal and the second characteristic signal into the initially trained decoder corresponding to each initially trained encoder for signal reconstruction, thereby obtaining a first reconstructed signal corresponding to the first bearing time series signal and a second reconstructed signal corresponding to each second bearing time series signal;

[0021] Inputting the first reconstructed signal and the second reconstructed signal into the corresponding second bearing fault diagnosis classifier to be trained for classification, and obtaining the fault recognition result of the second bearing fault diagnosis classifier to be trained for the first reconstructed signal and the second reconstructed signal;

[0022] A joint loss function value is calculated based on the fault identification result, the second reconstructed signal, the second bearing time series signal, the second feature signal and the second true label;

[0023] Determine whether the above-mentioned joint loss function value converges; if so, determine that the training of each decoder after initial training is completed to obtain the above-mentioned preset decoder; if not, adjust the encoder after initial training, each decoder after initial training and each neural network weight in the second bearing fault diagnosis classifier, and continue to train the above-mentioned decoders after initial training.

[0024] Furthermore, the acquisition of the first bearing time series signal and the second bearing time series signal includes:

[0025] Acquire a first bearing vibration signal when the bearing is in a normal state and a second bearing vibration signal when the bearing is in each fault type state within a first preset time period;

[0026] The first bearing vibration signal and the second bearing vibration signal are extracted respectively according to the preset time window length to obtain a plurality of first bearing vibration signals and a plurality of second bearing vibration signals under each fault type;

[0027] Combining the above-mentioned first bearing vibration signals to obtain a first bearing time series signal;

[0028] A plurality of second bearing vibration signals under each fault type are combined to obtain a second bearing time series signal corresponding to each fault type.

[0029] Furthermore, the above-mentioned calculation of the joint loss function value based on the above-mentioned fault identification result, the second reconstructed signal, the second bearing time series signal, the second feature signal and the second true label includes:

[0030] For each decoder after initial training, each time a second reconstructed signal is obtained, a first mean square error loss function value between the second reconstructed signal and the second bearing time series signal is calculated;

[0031] For each encoder after initial training, each time a second characteristic signal is obtained, a relative entropy divergence value between the second characteristic signal and a normal distribution is calculated;

[0032] After each fault identification result is obtained, a second cross entropy loss function value between the fault identification result and the second true label is calculated;

[0033] The sum of the first mean square error loss function value, the relative entropy divergence value and the second cross entropy loss function value is calculated to obtain a joint loss function value.

[0034] Furthermore, the training of the encoder after the initial training and the decoder after the initial training includes:

[0035] Inputting the first bearing time series signal into the encoder to be trained for encoding to obtain a third characteristic signal;

[0036] Inputting the third characteristic signal into a decoder to be trained to perform signal reconstruction, thereby obtaining a reconstructed third reconstructed time series signal;

[0037] When each third reconstructed time series signal is obtained, a second mean square error loss function value between the third reconstructed time series signal and the first bearing time series signal is calculated, and a relative entropy divergence value between the third characteristic signal and a normal distribution function is calculated;

[0038] Calculate the sum of the second mean square error loss function value and the relative entropy divergence value to obtain the first loss function value;

[0039] Every time a first loss function value is obtained, it is determined whether the above-mentioned first loss function value converges; if so, it is determined that the training of the above-mentioned initial training encoder and the initial training decoder is completed; if not, the weights of each neural network in the above-mentioned encoder and decoder are adjusted, and the above-mentioned encoder and decoder are continued to be trained.

[0040] Furthermore, after obtaining the above fault diagnosis result, the method further includes:

[0041] Extracting the confidence level corresponding to the fault diagnosis result, and sending the diagnosis result and the confidence level to a human-computer interaction interface for display;

[0042] If the fault diagnosis result is not fault-free, the confidence level corresponding to the fault diagnosis result is compared with a preset confidence threshold;

[0043] When the above confidence level is higher than the above preset confidence threshold, an early warning is issued.

[0044] Based on the above method embodiment, the present invention provides a corresponding device embodiment;

[0045] The present invention provides a bearing fault diagnosis device based on a neural network, comprising:

[0046] Signal acquisition module, signal merging module, fault diagnosis module and fault diagnosis result determination module;

[0047] The signal acquisition module is used to acquire the bearing vibration signal of the bearing to be diagnosed in a preset time period; wherein the vibration signal includes: a horizontal vibration signal and a vertical vibration signal;

[0048] The signal merging module is used to extract the bearing vibration signal according to a preset time window length, and merge the extracted bearing vibration signals to obtain a bearing time series signal;

[0049] The above-mentioned fault diagnosis module is used to input the above-mentioned bearing time series signal into a preset first bearing fault diagnosis classifier constructed based on a neural network, so that the above-mentioned preset first bearing fault diagnosis classifier obtains the confidence corresponding to each bearing time series signal according to the above-mentioned bearing time series signal; wherein the above-mentioned confidence includes: the confidence of no fault and the confidence of each fault type;

[0050] The fault diagnosis result determination module is used to use the result corresponding to the maximum confidence level among the above confidence levels as the fault diagnosis result of the bearing to be diagnosed.

[0051] Based on the above method embodiment, the present invention provides a terminal device embodiment;

[0052] The present invention provides a terminal device, including a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor. When the processor executes the computer program, it implements a neural network-based bearing fault diagnosis method according to any embodiment of the present invention.

[0053] Based on the above method embodiment, the present invention provides a storage medium embodiment;

[0054] The present invention provides a storage medium, including a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor. When the processor executes the computer program, a neural network-based bearing fault diagnosis method according to any embodiment of the present invention is implemented.

[0055] The embodiments of the present invention have the following beneficial effects:

[0056] The present invention provides a bearing fault diagnosis method, device, terminal equipment and storage medium based on neural network. The above method first obtains the bearing vibration signal of the bearing to be diagnosed in a preset time period; wherein the above vibration signal includes: a horizontal vibration signal and a vertical vibration signal; then the above bearing vibration signal is extracted according to the preset time window length, and the extracted bearing vibration signals are merged to obtain a bearing time series signal; then the above bearing time series signal is input into a preset first bearing fault diagnosis classifier constructed based on a neural network, so that the above preset first bearing fault diagnosis classifier obtains the confidence corresponding to each bearing time series signal according to the above bearing time series signal; wherein the above confidence includes: a confidence of no fault and a confidence of each fault type; finally, the result corresponding to the maximum confidence among the above confidences is used as the fault diagnosis result of the above bearing to be diagnosed. Therefore, the present invention performs fault diagnosis on the bearing based on the bearing time series signal under bearing vibration through the trained preset first bearing fault diagnosis classifier, and determines the fault type of the bearing without fault or fault, thereby realizing the identification of the fault type of the faulty bearing. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] Figure 1 It is a flowchart of a bearing fault diagnosis method based on a neural network provided in one embodiment of the invention.

[0058] Figure 2 It is a schematic diagram of the training process of a preset first bearing fault diagnosis classifier provided in one embodiment of the present invention.

[0059] Figure 3 It is a schematic diagram of fine-tuning and training an encoder and a decoder corresponding to a single fault type provided by an embodiment of the present invention.

[0060] Figure 4 It is a structural diagram of a second bearing fault diagnosis classifier provided by an embodiment of the present invention.

[0061] Figure 5 It is a schematic diagram of a first bearing time series signal provided by an embodiment of the present invention.

[0062] Figure 6 It is a schematic diagram of a Mask corresponding to fine-tuning training of an encoder and a decoder corresponding to a single fault type provided by an embodiment of the present invention.

[0063] Figure 7 It is a schematic diagram of the training process of an encoder after initial training and a decoder after initial training provided by an embodiment of the present invention.

[0064] Figure 8 It is a schematic diagram of the encoding process of the encoder provided by an embodiment of the present invention.

[0065] Fig. 9 It is a schematic diagram of the decoding process of a decoder provided by an embodiment of the present invention.

[0066] Fig.10 It is a structural schematic diagram of a bearing fault diagnosis device based on a neural network provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0067] The technical solution of the present invention will be described clearly and completely below in conjunction with the accompanying drawings 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.

[0068] like Figure 1 As shown, an embodiment of the present invention provides a bearing fault diagnosis method based on a neural network, comprising:

[0069] Step S101: obtaining a bearing vibration signal of a bearing to be diagnosed in a preset time period; wherein the vibration signal includes: a horizontal vibration signal and a vertical vibration signal;

[0070] Specifically, two vibration acceleration sensors are installed on the bearing cap of the bearing to collect the vertical vibration signal and the horizontal vibration signal of the bearing respectively.

[0071] Specifically, the vibration acceleration sensor consists of two miniature accelerometers at 90° to each other; the first is placed on the vertical axis and the second is placed on the horizontal axis. The two accelerometers are placed radially on the outer ring of the bearing. Acceleration measurements are sampled at a fixed frequency, and data is continuously collected at a preset time interval at fixed intervals.

[0072] Step S102: extracting the bearing vibration signal according to a preset time window length, and merging the extracted bearing vibration signals to obtain a bearing time series signal;

[0073] Preferably, the collected bearing vibration signal can be segmented into time windows at fixed time intervals (usually a few minutes) to obtain a signal of fixed length. When segmenting the bearing vibration signal, the bearing horizontal vibration signal and the bearing vertical vibration signal are segmented separately, and then the segmented results are merged to obtain a bearing time series signal.

[0074] Step S103: input the above-mentioned bearing time series signal into a preset first bearing fault diagnosis classifier constructed based on a neural network, so that the above-mentioned preset first bearing fault diagnosis classifier obtains the confidence corresponding to each bearing time series signal according to the above-mentioned bearing time series signal; wherein the above-mentioned confidence includes: the confidence of no fault and the confidence of each fault type;

[0075] Specifically, the fault types include inner ring damage, outer ring damage, and ball damage.

[0076] Specifically, the first bearing fault diagnosis classifier is preset and deployed on an industrial computer or an enterprise server and loaded into a GPU or a memory of a computer. According to the received fixed-length bearing time series signal, the bearing fault condition is diagnosed and determined in real time.

[0077] In a preferred embodiment, the training of the preset first bearing fault diagnosis classifier includes:

[0078] Acquire a first bearing time series signal when the bearing is in a normal state within a first preset time period;

[0079] Randomly collect a number of one-dimensional latent vectors from a normal distribution to obtain a random sampling signal;

[0080] For each random sampling signal, the random sampling signal is input into a corresponding preset decoder to generate a derived fault time series signal, and then a plurality of derived fault time series signals and a fault type corresponding to each derived fault time series signal are obtained;

[0081] Input the above-mentioned several derivative fault time series signals with the first true label and the above-mentioned first bearing time series signal into the first bearing fault diagnosis classifier to be trained, and obtain the confidence of each derivative fault time series signal and the above-mentioned first bearing time series signal being judged as fault-free and the confidence of each fault type judged by the first bearing fault diagnosis classifier to be trained; wherein the above-mentioned first true label is: the fault type label of the above-mentioned derivative fault time series signal and the fault-free label of the above-mentioned first bearing time series signal;

[0082] For each confidence level, calculating a first cross entropy loss function between the confidence level and the first true label, and determining whether the first cross entropy loss function converges;

[0083] If the above-mentioned first cross entropy loss function converges, it is determined that the training of the above-mentioned first bearing fault diagnosis classifier is completed, and the above-mentioned preset first bearing fault diagnosis classifier is obtained; if the above-mentioned first cross entropy loss function does not converge, adjust the weights of each neural network in the above-mentioned first bearing fault diagnosis classifier, and continue to train the above-mentioned first bearing fault diagnosis classifier.

[0084] Schematically, the training process diagram of the preset first bearing fault diagnosis classifier is as follows: Figure 2 As shown, a number of one-dimensional latent vectors are randomly sampled from the standard normal distribution to obtain a random sampling signal TS_sample_for_fualt, size = (arbitrary number, 1, 512), which is equivalent to containing any number of one-dimensional latent vectors z of length 512. The multi-convolutional neural network ConvNet_2 is used as a decoder to deconvolve the above random sampling signal to generate a new derivative fault time series signal TS_gen under abnormal state, size = (arbitrary number, 1, 512), wherein the decoder corresponding to each fault type is deconvoluted to generate a derivative fault time series signal corresponding to the fault type, which is a derivative signal of the time series signal of the bearing under fault state. By collecting any number of z vectors, any number of TS_gen can be obtained.

[0085] Specifically, a large number of derivative fault time series signals TS_gen and a large number of first bearing time series signals TS when the bearings are in normal state, size = (arbitrary number, 1,512) are used to train the above-mentioned first bearing fault diagnosis classifier. In the process of multiple iterations, the gradient descent algorithm is used to calculate the gradient, and the neural network weights of the first bearing fault diagnosis classifier ConvNet_big are optimized according to the gradient until the first cross entropy loss function Loss_big_cls converges.

[0086] Preferably, a number of derived fault time series signals are obtained by random sampling, so that when fault data is scarce, the fault classification capability of the first bearing fault diagnosis classifier can be improved by the derived fault time series signals.

[0087] In this preferred embodiment, the first bearing fault diagnosis classifier is trained using the first bearing time series signal and the derived fault time series signal until the first cross entropy loss function converges, thereby obtaining the above-mentioned preset first bearing fault diagnosis classifier.

[0088] In another preferred embodiment, the training of the preset decoder includes:

[0089] Acquire a second bearing time series signal of the bearing under each fault type within a first preset time period;

[0090] For each fault type, each second bearing time series signal with a second true label and the first bearing time series signal are respectively input into the corresponding encoder after initial training for encoding, so as to obtain the second feature signal corresponding to each of the second bearing time series signals and the first feature signal corresponding to the first bearing time series signal; wherein the second true label is: the fault type label corresponding to each of the second bearing time series signals and the fault-free label corresponding to the first bearing time series signal;

[0091] Inputting the first characteristic signal and the second characteristic signal into the initially trained decoder corresponding to each initially trained encoder for signal reconstruction, thereby obtaining a first reconstructed signal corresponding to the first bearing time series signal and a second reconstructed signal corresponding to each second bearing time series signal;

[0092] Inputting the first reconstructed signal and the second reconstructed signal into the corresponding second bearing fault diagnosis classifier to be trained for classification, and obtaining the fault recognition result of the second bearing fault diagnosis classifier to be trained for the first reconstructed signal and the second reconstructed signal;

[0093] A joint loss function value is calculated based on the fault identification result, the second reconstructed signal, the second bearing time series signal, the second feature signal and the second true label;

[0094] Determine whether the above-mentioned joint loss function value converges; if so, determine that the training of each decoder after initial training is completed to obtain the above-mentioned preset decoder; if not, adjust the encoder after initial training, each decoder after initial training and each neural network weight in the second bearing fault diagnosis classifier, and continue to train the above-mentioned decoders after initial training.

[0095] Specifically, the gradient value is calculated using a gradient descent algorithm, and the neural network weights in the encoder after initial training, each decoder after initial training, and the second bearing fault diagnosis classifier are optimized according to the gradient value.

[0096] Specifically, in the entire process of training the preset decoder, it is actually a process of fine-tuning the encoder and decoder uniformly. In this process, each fault type corresponds to an encoder, a decoder and a second bearing fault diagnosis classifier. Schematically, assuming that there are four fault types A, B, C and D, and there are four groups of encoders, decoders and second bearing fault diagnosis classifiers a, b, c and d, and each combination has an encoder, a decoder and a second bearing fault diagnosis classifier. Then, during training, the signal data corresponding to the A fault type (normal signal data + signal data of the A fault type) is input into the a group for training, and the joint loss function value corresponding to the A fault type is calculated; the signal data corresponding to the B fault type (normal signal data + signal data of the B fault type) is input into the b group for training, and the joint loss function value corresponding to the B fault type is calculated; the signal data corresponding to the C fault type (normal signal data + signal data of the C fault type) is input into the c group for training, and the joint loss function value corresponding to the C fault type is calculated; the signal data corresponding to the D fault type (normal signal data + signal data of the D fault type) is input into the d group for training, and the joint loss function value corresponding to the D fault type is calculated.

[0097] Schematically, the diagram of fine-tuning the encoder and decoder corresponding to a single fault type is as follows: Figure 3As shown, during fine-tuning training, the second bearing time series signal TS_fualt corresponding to the fault type, input size = (10, N, 2), where N represents the signal length, and the first bearing time series signal TS_norm, the first bearing time series signal is the same as the TS in the previous embodiment, and the input size = (Batchsize-10, N, 2) are merged into a batch to obtain a three-dimensional matrix of shape = (Batchsize, N, 2), where the value of Batchsize is the sum of the number of second bearing time series signals and first bearing time series signals. After the obtained three-dimensional matrix is ​​input into the encoder ConvNet_1, the second characteristic signal TS_fualt_enc, size = (10, 1, 512) and the first characteristic signal TS_norm_enc, size = (Batchsize-10, 1, 512) are obtained respectively. Then the first characteristic signal and the second characteristic signal are input into the decoder for signal reconstruction, and the first reconstructed signal TS_norm_dec corresponding to the first bearing time series signal is obtained, the output size = (Batchsize-10, N, 2), and the second reconstructed signal TS_fualt_dec corresponding to the second bearing time series signal is obtained, the output size = (10, N, 2). Finally, the first reconstructed signal and the second reconstructed signal are input into the second bearing fault diagnosis classifier ConvNet_cls for fault identification, and the fault identification result is obtained, that is, the identification result of the corresponding fault type and the identification result of no fault, wherein the judgment of the identification result needs to be determined according to the corresponding confidence value, and the corresponding identification result with the highest confidence is used as the identification result of the second bearing fault diagnosis classifier. Then, based on the fault identification result, the second bearing time series signal, the second reconstructed signal obtained after the decoder, the second feature signal and the second true label, the joint loss corresponding to this fault type is calculated.

[0098] Schematically, the structure of the second bearing fault diagnosis classifier is as follows: Figure 4 As shown, Softmax is used to map the output of the fully connected layer to the probability distribution of the prediction result, which is expressed as follows:

[0099]

[0100] In the formula, x represents the probability distribution predicted by the second bearing fault diagnosis classifier for a single bearing time series signal. The probability distribution is represented by a vector, and x i represents the probability that the currently predicted bearing time series signal is predicted as the i-th fault type or predicted as no fault, x j represents the probability that the currently predicted bearing time series signal is predicted as the jth fault type or predicted as no fault, and k represents the total number of fault types.

[0101] It is worth noting that the structure of the above-mentioned second bearing fault diagnosis classifier is similar to that of the preset first bearing fault diagnosis classifier. The preset first bearing fault diagnosis classifier is only expanded in dimension, and the expansion size is set according to the amount of data.

[0102] In this preferred embodiment, the encoder and decoder are fine-tuned and trained based on the second bearing time series signal and the first bearing time series signal of the bearing under various fault types.

[0103] In another preferred embodiment, the acquisition of the first bearing time series signal and the second bearing time series signal includes:

[0104] Acquire a first bearing vibration signal when the bearing is in a normal state and a second bearing vibration signal when the bearing is in each fault type state within a first preset time period;

[0105] The first bearing vibration signal and the second bearing vibration signal are extracted respectively according to the preset time window length to obtain a plurality of first bearing vibration signals and a plurality of second bearing vibration signals under each fault type;

[0106] Combining the above-mentioned first bearing vibration signals to obtain a first bearing time series signal;

[0107] A plurality of second bearing vibration signals under each fault type are combined to obtain a second bearing time series signal corresponding to each fault type.

[0108] Specifically, two vibration acceleration sensors are installed on the bearing cover of a rotating mechanical device (such as a steam turbine, compressor, fan, and pump, etc.), one of which is used to collect the vertical vibration signal of the bearing; the other is used to collect the horizontal vibration signal of the bearing. The monitoring data is stored in the database or file of the industrial computer through the gateway for use in this method. The vibration acceleration sensor consists of two miniature accelerometers at 90° to each other. The first is placed on the vertical axis and the second is placed on the horizontal axis. The two accelerometers are placed radially on the outer ring of the bearing. The acceleration measurement is sampled at 48kHz, and 1 second of data is collected every 1 minute, and 60 minutes of data are collected continuously. First, the monitoring data of multiple bearings in normal state is collected. If the bearing has multiple load conditions when working, it is necessary to collect the horizontal vibration signal and vertical vibration signal of the bearing under various load conditions. Data collection in normal state requires monitoring data of no less than H1 moments, where H1 = 48000 × 60 × number of load conditions. Data collection under fault conditions requires monitoring data for no less than H2 moments, where H2 = 48000 × number of load conditions, i.e., vibration signals within at least 1 second must be collected for each fault type.

[0109] Specifically, for a single bearing in a normal state, the vibration signal is collected continuously for 60 minutes, and N=128 is set as the length of the time window, and a sliding window is performed on the 60-minute continuous vibration signal. The time window is taken once at a fixed interval M (the value of M depends on the situation, generally M=N / 2), thereby obtaining a time series signal with a signal length of N. The horizontal vibration signal and the vertical vibration signal are both extracted in this way to obtain signals A and B of the horizontal vibration signal and the vertical vibration signal, and A and B are merged into a time series signal, namely the first bearing time series signal TS. The schematic diagram of the first bearing time series signal is shown in the figure. Figure 5 As shown. According to this method, data of multiple bearings in normal state are collected, and the time series signal of the first bearing is extracted. The collected bearings can be under different load conditions. The more the number of collected bearings, the better, at least 10.

[0110] Specifically, the vibration signal of the bearing in the fault state is obtained in the same way as the time series signal of the first bearing. The length of the time series signal and the sliding window interval used are consistent with the signal processing in the normal state. It is required to collect no less than 10 time series signals as training sets for each fault type. The more time series signals there are, the higher the quality of the generated data.

[0111] In this preferred embodiment, the first bearing vibration signal when the bearing is in a normal state and the second bearing vibration signal when the bearing is in various fault types are extracted and merged to obtain a first bearing time series signal and a second bearing time series signal.

[0112] In another preferred embodiment, the joint loss function value is calculated based on the fault identification result, the second reconstructed signal, the second bearing time series signal, the second feature signal and the second true label, including:

[0113] For each decoder after initial training, each time a second reconstructed signal is obtained, a first mean square error loss function value between the second reconstructed signal and the second bearing time series signal is calculated;

[0114] For each encoder after initial training, each time a second characteristic signal is obtained, a relative entropy divergence value between the second characteristic signal and a normal distribution is calculated;

[0115] After each fault identification result is obtained, a second cross entropy loss function value between the fault identification result and the second true label is calculated;

[0116] Specifically, the second cross entropy loss function value is calculated according to the following formula:

[0117]

[0118] In the formula, p(x i ) and q(x i ) represent the probability of the real fault identification result and the probability of the fault identification result identified by the second bearing fault diagnosis classifier, usually 1, x i Represents the i-th signal in the current Batchsize bearing time series signals.

[0119] The sum of the first mean square error loss function value, the relative entropy divergence value and the second cross entropy loss function value is calculated to obtain a joint loss function value.

[0120] Specifically, when fine-tuning the encoder and decoder, the Mask is first used to identify which signals in each batch input into the encoder belong to the second bearing time series signals under the fault state and which signals belong to the first bearing time series signals under the normal state. The Mask is a one-dimensional vector with a length of Batchsize. If the value of the element with index i is True, it means that the time series signal with index i in this batch is a signal under the fault state. Schematically, the Mask diagram corresponding to the fine-tuning of the encoder and decoder corresponding to a single fault type is as follows: Figure 6 Therefore, when calculating the first mean square error loss function value and the relative entropy divergence value, the bearing time series signal that needs to be involved in the calculation can be accurately identified.

[0121] In this preferred embodiment, a joint loss function value is calculated based on the fault identification result, the second reconstructed signal, the second bearing time series signal, the second feature signal and the second true label.

[0122] In another preferred embodiment, the training of the encoder after the initial training and the decoder after the initial training includes:

[0123] Inputting the first bearing time series signal into the encoder to be trained for encoding to obtain a third characteristic signal;

[0124] Inputting the third characteristic signal into a decoder to be trained to perform signal reconstruction, thereby obtaining a reconstructed third reconstructed time series signal;

[0125] When each third reconstructed time series signal is obtained, a second mean square error loss function value between the third reconstructed time series signal and the first bearing time series signal is calculated, and a relative entropy divergence value between the third characteristic signal and a normal distribution function is calculated;

[0126] Calculate the sum of the second mean square error loss function value and the relative entropy divergence value to obtain the first loss function value;

[0127] Every time a first loss function value is obtained, it is determined whether the above-mentioned first loss function value converges; if so, it is determined that the training of the above-mentioned initial training encoder and the initial training decoder is completed; if not, the weights of each neural network in the above-mentioned encoder and decoder are adjusted, and the above-mentioned encoder and decoder are continued to be trained.

[0128] Specifically, during the initial training of the encoder after the initial training and the decoder after the initial training, each fault type corresponds to an encoder and a decoder. Schematically, assuming that there are four fault types A, B, C and D, and there are four combinations of encoders and decoders a, b, c and d, and each combination has an encoder and a decoder, then during training, the signal data corresponding to the A fault type (normal signal data + signal data of the A fault type) is input into the a group for training, the signal data corresponding to the B fault type (normal signal data + signal data of the B fault type) is input into the b group for training, the signal data corresponding to the C fault type (normal signal data + signal data of the C fault type) is input into the c group for training, and the signal data corresponding to the D fault type (normal signal data + signal data of the D fault type) is input into the d group for training.

[0129] Specifically, the training process of the encoder after the initial training and the decoder after the initial training is as follows: Figure 7 As shown. Using the multi-layer convolutional neural network ConvNet_1 as the encoder, the input first bearing time series signal TS is encoded to obtain an abstract feature, namely the third feature signal TS_enc, size = (Batchsize, 1,512). The first bearing time series signal is essentially a two-dimensional matrix, shape = (N, 2), and the encoder encodes the first bearing time series signal into a two-dimensional matrix of shape = (1,512). This two-dimensional matrix is ​​equivalent to the image of the first bearing time series signal in a low-dimensional space (coding space), which is the compressed information of the first bearing time series signal; each component of the third feature signal can describe a certain aspect of the attribute of the entire sequence in the first bearing time series signal, and the specific attribute meaning represented by each component is automatically learned by the encoder. During the entire training process, the sum of the second mean square error loss function value and the relative entropy divergence value is used as the joint loss function value during training, that is, the above-mentioned first loss function value. Schematically, the encoding process of the encoder is as follows Figure 8 shown.

[0130] Specifically, the second convolutional neural network ConvNet_2 is used as a decoder to deconvolve the third feature signal to obtain a reconstructed third time series signal TS_dec, the size of which is equal to (Batchsize, N, 2). Schematically, the decoding process of the decoder is shown in the figure below: Fig. 9 shown. Figure 8 and Fig. 9 Where k = 7, which can be adjusted according to the length of the specific time series. When the convolution step is 1 and there is no padding, a one-dimensional vector f of length n is convolved with a one-dimensional vector w of length k to obtain a one-dimensional vector g of length n-k+1. The mathematical expression of the convolution is as follows:

[0131]

[0132] In the formula, f represents the information of the first bearing time series signal in a single attribute dimension, that is, the horizontal time series signal or the vertical time series signal in the first bearing time series signal, and the length is N. Taking the horizontal time series signal as an example, f[i+j] represents the horizontal time series signal at the [i+j]th moment in the first bearing time series signal, the one-dimensional vector w[j] represents the jth weight of the convolution kernel w, and g[i] represents the weighted sum of the horizontal time series signals from the i-th to the i+k-1-th moments in the horizontal time series signal.

[0133] Specifically, deconvolution is the inverse operation of the mathematical expression of the above convolution. After each convolution layer and deconvolution layer, it needs to be processed by an activation function, generally using the Tanh function, which is expressed as follows:

[0134]

[0135] Where x represents the output of a single neuron, and tanh represents the Tanh activation function value.

[0136] Specifically, the second mean square error loss function value Loss_dec is calculated according to the input first bearing time series signal and the third reconstructed time series signal TS_dec obtained by the decoder reconstruction. The formula for calculating the second mean square error loss function value is as follows:

[0137]

[0138] In the formula, y i represents the signal data at the i-th moment in the time series signal of the first bearing, Represents the signal data at the i-th moment in the third reconstructed time series signal. It should be noted that the above formula for calculating the second mean square error loss function value only represents the calculation of the second mean square error loss function value of the horizontal time series signal or the vertical time series signal in the first bearing time series signal, so the final second mean square error loss function value is the sum of the second mean square error loss function values ​​corresponding to the horizontal time series signal and the vertical time series signal.

[0139] Specifically, the similarity between the third characteristic signal TS_enc obtained by the encoder and the normal distribution N(0,1) is calculated, and the similarity is used as the relative entropy divergence value Loss_kl between the two. The calculation of the relative entropy divergence value here can be understood as calculating the information loss from the normal distribution to the third characteristic signal. KL divergence is used here as the similarity function. KL divergence (Kullback-Leibler divergence), also known as relative entropy, is an indicator used to measure the difference between two probability distributions. The relative entropy divergence value is calculated according to the following formula:

[0140]

[0141] In the formula, p(x) represents the normal distribution, q(x) does not affect the probability distribution of each component in the third characteristic signal, D KL (p‖q) represents the relative entropy divergence value.

[0142] Specifically, when adjusting the weights of each neural network in the encoder and decoder, the back propagation algorithm is used to calculate the gradient, and the gradient descent algorithm is used according to the gradient to optimize the neural network weights of the encoder and decoder.

[0143] Preferably, the gradient descent algorithm, as a built-in component in deep learning frameworks such as Pytorch and Tensorflow, can calculate the gradient of each neuron weight based on the loss value. Taking any convolutional layer in the encoder ConvNet_1 as an example, assuming that W is the learnable neural network weight of the current convolutional layer, it is necessary to calculate the derivatives of the second mean square error loss function value and the relative entropy divergence value with respect to W, and then update W as follows:

[0144]

[0145] Where C represents the value of the first loss function, and η represents the learning rate, which is a hyperparameter that controls the update step size, usually 0.01. During the training process, each iteration needs to update the learnable neural network weight W according to the new first loss function value until the training is completed.

[0146] In this preferred embodiment, the encoder and the decoder are trained according to the first bearing time series signal to obtain an encoder after initial training and a decoder after initial training.

[0147] Step S104: taking the result corresponding to the maximum confidence level among the above confidence levels as the fault diagnosis result of the above bearing to be diagnosed.

[0148] Specifically, if the confidence corresponding to the fault type of inner ring damage is the largest among the currently obtained confidences, then the fault diagnosis result of the bearing to be diagnosed is that the inner ring is damaged.

[0149] Preferably, since a minor bearing fault usually does not immediately cause a machine shutdown, and this method can identify whether a bearing is faulty before the bearing is completely damaged, this method can perform predictive fault diagnosis. Even if the time from the occurrence of a bearing fault to complete damage is extremely short, this method can perform real-time diagnosis, so this method has a high application value.

[0150] Preferably, the above steps and embodiments can diagnose and identify whether the bearing is fault-free and the specific fault type when a fault occurs based on a preset first bearing fault diagnosis classifier, and based on a method similar to diagnosing bearing faults based on the preset first bearing fault diagnosis classifier, a preset third bearing fault diagnosis classifier can be used to diagnose whether the bearing has a fault.

[0151] Specifically, the method for diagnosing whether a bearing fault occurs according to the preset third bearing fault diagnosis classifier is as follows:

[0152] Acquire a second bearing vibration signal of the bearing to be diagnosed in a second preset time period; wherein the vibration signal includes: a horizontal vibration signal and a vertical vibration signal; extract the second bearing vibration signal according to a second preset time window length, merge the extracted bearing vibration signals to obtain a third bearing time series signal; input the third bearing time series signal into a preset second bearing fault diagnosis classifier constructed based on a neural network, so that the preset second bearing fault diagnosis classifier obtains a first confidence level corresponding to each third bearing time series signal according to the third bearing time series signal; wherein the first confidence level includes: a first confidence level for no fault and a first confidence level for a fault; use the result corresponding to the largest first confidence level among the first confidence levels as the fault diagnosis result of the bearing to be diagnosed.

[0153] Specifically, the training of the preset third bearing fault diagnosis classifier is similar to the training of the preset first bearing fault diagnosis classifier. The difference is that when training the preset third bearing fault diagnosis classifier, the number of encoders, decoders and the corresponding fourth bearing fault diagnosis classifier involved is 1, and the real labels carried in the signal data are faulty and no fault, without specific fault type labels. At the same time, the classifier involved outputs 2 probability values ​​for each time series signal, corresponding to the confidence of faulty and no faulty, respectively. In the above-mentioned preset first bearing fault diagnosis classifier and second bearing fault diagnosis classifier, the number of probability values ​​output is the total number of fault types plus 1.

[0154] In another preferred embodiment, after obtaining the above fault diagnosis result, the method further includes:

[0155] Extracting the confidence level corresponding to the fault diagnosis result, and sending the diagnosis result and the confidence level to a human-computer interaction interface for display;

[0156] If the fault diagnosis result is not fault-free, the confidence level corresponding to the fault diagnosis result is compared with a preset confidence threshold;

[0157] When the above confidence level is higher than the above preset confidence threshold, an early warning is issued.

[0158] Specifically, the diagnosis results are returned to the human-computer interaction interface, and the report shows whether each bearing is damaged and the confidence level of the damage. If the confidence level is higher than the set threshold, the relevant staff will be notified via email or the Andon system.

[0159] In this preferred embodiment, the confidence level corresponding to the fault diagnosis result that does not correspond to no fault is compared with a preset confidence threshold, and an early warning is issued when the confidence level is higher than the preset confidence threshold.

[0160] Preferably, experiments have shown that, compared with the prior art, the method has a more stable fault recognition rate and higher generalization, and the fault recognition rate reaches 98±0.5%, which is valuable for use.

[0161] Based on the above method embodiment, the present invention provides a corresponding device embodiment.

[0162] like Fig.10 As shown, an embodiment of the present invention provides a bearing fault diagnosis device based on a neural network, comprising:

[0163] Signal acquisition module, signal merging module, fault diagnosis module and fault diagnosis result determination module;

[0164] The signal acquisition module is used to acquire the bearing vibration signal of the bearing to be diagnosed in a preset time period; wherein the vibration signal includes: a horizontal vibration signal and a vertical vibration signal;

[0165] The signal merging module is used to extract the bearing vibration signal according to a preset time window length, and merge the extracted bearing vibration signals to obtain a bearing time series signal;

[0166] The above-mentioned fault diagnosis module is used to input the above-mentioned bearing time series signal into a preset first bearing fault diagnosis classifier constructed based on a neural network, so that the above-mentioned preset first bearing fault diagnosis classifier obtains the confidence corresponding to each bearing time series signal according to the above-mentioned bearing time series signal; wherein the above-mentioned confidence includes: the confidence of no fault and the confidence of each fault type;

[0167] The fault diagnosis result determination module is used to use the result corresponding to the maximum confidence level among the above confidence levels as the fault diagnosis result of the bearing to be diagnosed.

[0168] It should be noted that the device embodiments described above are merely schematic, wherein the modules described above as separate components may or may not be physically separated, and the components displayed as modules may or may not be physical modules, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the present embodiment. In addition, in the accompanying drawings of the device embodiments provided by the present invention, the connection relationship between the modules indicates that there is a communication connection between them, which can be specifically implemented as one or more communication buses or signal lines. A person of ordinary skill in the art can understand and implement it without creative work. The above schematic diagram is only an example of a bearing fault diagnosis device based on a neural network, and does not constitute a limitation on a bearing fault diagnosis device based on a neural network, and may include more or fewer components than shown in the figure, or a combination of certain components, or different components.

[0169] Based on the above method item embodiments, the present invention provides corresponding terminal device item embodiments.

[0170] Another embodiment of the present invention provides a terminal device, including a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, and when the processor executes the computer program, it implements a neural network-based bearing fault diagnosis method described in any embodiment of the present invention.

[0171] Exemplarily, in this embodiment, the computer program may be divided into one or more modules, which are stored in the memory and executed by the processor to implement the present invention. The one or more module elements may be a series of computer program instruction segments capable of implementing specific functions, which are used to describe the execution process of the computer program in the device;

[0172] The terminal device may be a computing device such as a desktop computer, a notebook, a PDA, a cloud server, etc. The device may include, but is not limited to, a processor, a memory;

[0173] The processor referred to herein may be a central processing unit (CPU), other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or any conventional processor, etc. The above-mentioned processor is the control center of the above-mentioned device, and uses various interfaces and lines to connect various parts of the entire device;

[0174] The above-mentioned memory can be used to store the above-mentioned computer programs and / or modules. The above-mentioned processor realizes various functions of the above-mentioned device by running or executing the computer programs and / or modules stored in the above-mentioned memory, and calling the data stored in the memory. The above-mentioned memory can mainly include a program storage area and a data storage area, wherein the program storage area can store an operating system, at least one application required for a function, etc.; in addition, the memory can include a high-speed random access memory, and can also include a non-volatile memory, such as a hard disk, a memory, a plug-in hard disk, a smart memory card (Smart Media Card, SMC), a secure digital (Secure Digital, SD) card, a flash card (Flash Card), at least one disk storage device, a flash memory device, or other volatile solid-state storage devices.

[0175] Based on the above method item embodiments, the present invention provides a corresponding storage medium item embodiment.

[0176] Another embodiment of the present invention provides a storage medium, which includes a stored computer program, wherein when the computer program is running, the device where the storage medium is located is controlled to execute a neural network-based bearing fault diagnosis method according to any embodiment of the present invention.

[0177] In this embodiment, the storage medium is a computer-readable storage medium, the computer program includes computer program code, and the computer program code may be in source code form, object code form, executable file or some intermediate form, etc. The computer-readable medium may include: any entity or device capable of carrying the computer program code, recording medium, USB flash drive, mobile hard disk, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electric carrier signal, telecommunication signal and software distribution medium, etc.

[0178] Compared with the prior art, by implementing the above-mentioned embodiments of the present invention, the specific fault type occurring in the bearing can be identified.

[0179] The above is a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principle of the present invention. These improvements and modifications are also considered to be within the scope of protection of the present invention.

Claims

1. A bearing fault diagnosis method based on neural network, characterized in that: include: Acquire a bearing vibration signal of the bearing to be diagnosed in a preset time period; wherein the vibration signal includes: a horizontal vibration signal and a vertical vibration signal; Extracting the bearing vibration signal according to a preset time window length, and combining the extracted bearing vibration signals to obtain a bearing time series signal; The bearing time series signal is input into a preset first bearing fault diagnosis classifier constructed based on a neural network, so that the preset first bearing fault diagnosis classifier obtains the confidence corresponding to each bearing time series signal according to the bearing time series signal; wherein the confidence includes: the confidence of no fault and the confidence of each fault type; The result corresponding to the maximum confidence degree among the confidence degrees is used as the fault diagnosis result of the bearing to be diagnosed.

2. A method for diagnosing bearing faults based on a neural network according to claim 1, characterized in that: The training of the preset first bearing fault diagnosis classifier includes: Acquire a first bearing time series signal when the bearing is in a normal state within a first preset time period; Randomly collect a number of one-dimensional latent vectors from a normal distribution to obtain a random sampling signal; For each random sampling signal, input the random sampling signal into a corresponding preset decoder to generate a derived fault time series signal, and then obtain a plurality of derived fault time series signals and a fault type corresponding to each derived fault time series signal; Input the several derivative fault time series signals with the first true label and the first bearing time series signal into the first bearing fault diagnosis classifier to be trained, and obtain the confidence of each derivative fault time series signal and the first bearing time series signal being judged as fault-free and the confidence of each fault type by the first bearing fault diagnosis classifier to be trained; wherein the first true label is: the fault type label of the derivative fault time series signal and the fault-free label of the first bearing time series signal; For each confidence level, calculating a first cross entropy loss function between the confidence level and the first true label, and determining whether the first cross entropy loss function converges; If the first cross entropy loss function converges, it is determined that the training of the first bearing fault diagnosis classifier is completed, and the preset first bearing fault diagnosis classifier is obtained; if the first cross entropy loss function does not converge, the weights of each neural network in the first bearing fault diagnosis classifier are adjusted, and the first bearing fault diagnosis classifier continues to be trained.

3. A method for diagnosing bearing faults based on a neural network according to claim 2, characterized in that: The training of the preset decoder includes: Acquire a second bearing time series signal of the bearing under each fault type within a first preset time period; For each fault type, each second bearing time series signal with a second true label and the first bearing time series signal are respectively input into the corresponding encoder after initial training for encoding, so as to obtain the second feature signal corresponding to each second bearing time series signal and the first feature signal corresponding to the first bearing time series signal; wherein the second true label is: the fault type label corresponding to each second bearing time series signal and the no-fault label corresponding to the first bearing time series signal; Inputting the first characteristic signal and the second characteristic signal into the initially trained decoder corresponding to each initially trained encoder for signal reconstruction, to obtain a first reconstructed signal corresponding to the first bearing time series signal, and a second reconstructed signal corresponding to each second bearing time series signal; Inputting the first reconstructed signal and the second reconstructed signal into the corresponding second bearing fault diagnosis classifier to be trained for classification, and obtaining the fault recognition results of the second bearing fault diagnosis classifier to be trained for the first reconstructed signal and the second reconstructed signal; Calculate a joint loss function value according to the fault identification result, the second reconstructed signal, the second bearing time series signal, the second feature signal and the second true label; Determine whether the value of the joint loss function converges; if so, determine that the training of each initially trained decoder is completed to obtain the preset decoder; if not, adjust the encoder after the initial training, each initially trained decoder and each neural network weight in the second bearing fault diagnosis classifier, and continue to train the initially trained decoders.

4. A method for diagnosing bearing faults based on a neural network according to claim 3, characterized in that: The acquisition of the first bearing time series signal and the second bearing time series signal includes: Acquire a first bearing vibration signal when the bearing is in a normal state and a second bearing vibration signal when the bearing is in each fault type state within a first preset time period; Extracting the first bearing vibration signal and the second bearing vibration signal respectively according to the preset time window length to obtain a plurality of first bearing vibration signals and a plurality of second bearing vibration signals under each fault type; Combining the plurality of first bearing vibration signals to obtain a first bearing time series signal; A plurality of second bearing vibration signals under each fault type are combined to obtain a second bearing time series signal corresponding to each fault type.

5. A method for diagnosing bearing faults based on a neural network according to claim 4, characterized in that: The method of calculating a joint loss function value according to the fault identification result, the second reconstructed signal, the second bearing time series signal, the second feature signal and the second true label includes: For each decoder after initial training, each time a second reconstructed signal is obtained, a first mean square error loss function value between the second reconstructed signal and the second bearing time series signal is calculated; For each encoder after initial training, each time a second characteristic signal is obtained, a relative entropy divergence value between the second characteristic signal and a normal distribution is calculated; After each fault identification result is obtained, a second cross entropy loss function value between the fault identification result and the second true label is calculated; Calculate the sum of the first mean square error loss function value, the relative entropy divergence value, and the second cross entropy loss function value to obtain a joint loss function value.

6. A method for diagnosing bearing faults based on a neural network according to claim 5, characterized in that: The training of the encoder after the initial training and the decoder after the initial training includes: Inputting the first bearing time series signal into an encoder to be trained for encoding to obtain a third characteristic signal; Inputting the third characteristic signal into a decoder to be trained to perform signal reconstruction, thereby obtaining a reconstructed third reconstructed time series signal; When each third reconstructed time series signal is obtained, a second mean square error loss function value between the third reconstructed time series signal and the first bearing time series signal is calculated, and a relative entropy divergence value between the third characteristic signal and a normal distribution function is calculated; Calculate the sum of the second mean square error loss function value and the relative entropy divergence value to obtain a first loss function value; Every time a first loss function value is obtained, it is determined whether the first loss function value converges; if so, it is determined that the training of the initially trained encoder and the initially trained decoder is completed; if not, the weights of each neural network in the encoder and decoder are adjusted, and the encoder and decoder are continued to be trained.

7. A method for diagnosing bearing faults based on a neural network according to claim 6, characterized in that: After obtaining the fault diagnosis result, the method further includes: Extracting the confidence level corresponding to the fault diagnosis result, and sending the diagnosis result and the confidence level to a human-computer interaction interface for display; If the fault diagnosis result is not fault-free, comparing the confidence level corresponding to the fault diagnosis result with a preset confidence threshold; When the confidence level is higher than the preset confidence threshold, an early warning is issued.

8. A bearing fault diagnosis device based on a neural network, characterized in that: include: Signal acquisition module, signal merging module, fault diagnosis module and fault diagnosis result determination module; The signal acquisition module is used to acquire the bearing vibration signal of the bearing to be diagnosed in a preset time period; wherein the vibration signal includes: a horizontal vibration signal and a vertical vibration signal; The signal merging module is used to extract the bearing vibration signal according to a preset time window length, and merge the extracted bearing vibration signals to obtain a bearing time series signal; The fault diagnosis module is used to input the bearing time series signal into a preset first bearing fault diagnosis classifier constructed based on a neural network, so that the preset first bearing fault diagnosis classifier obtains the confidence corresponding to each bearing time series signal according to the bearing time series signal; wherein the confidence includes: the confidence of no fault and the confidence of each fault type; The fault diagnosis result determination module is used to use the result corresponding to the maximum confidence level among the confidence levels as the fault diagnosis result of the bearing to be diagnosed.

9. A terminal device, characterized in that: It comprises a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, wherein when the processor executes the computer program, a neural network-based bearing fault diagnosis method as described in any one of claims 1 to 7 is implemented.

10. A storage medium, characterized in that: The storage medium includes a stored computer program, wherein when the computer program is running, the device where the storage medium is located is controlled to execute a bearing fault diagnosis method based on a neural network as described in any one of claims 1 to 7.

Images