Bearing fault real-time diagnosis system based on LSTM-CNN fusion model
Through the LSTM-CNN fusion model and attention mechanism, the problems of low accuracy and poor real-time accuracy of bearing fault diagnosis in the existing technology are solved, efficient fault diagnosis on embedded devices are achieved, diagnostic efficiency and real-time performance are improved, and Integrating the system into an embedded platform simplifies industrial machinery and equipment inspection.
Patent Information
- Application Number
- CN202411253911.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-09-06
- Publication Date
- 2024-11-12
AI Technical Summary
The existing bearing fault diagnosis method was not evaluated during sample set training, resulting in poor model training effect and may produce diagnostic bias. In addition, deep learning neural networks are limited in application in embedded devices with limited computing resources, making it difficult to efficiently process data. and accurately diagnose faults.
Adopt a bearing fault real-time diagnostic system based on the LSTM-CNN fusion model. Through vibration acceleration data acquisition, preprocessing and feature extraction, attention mechanism and Dropout layer are added, trained and transplanted to the Raspberry Pi 4B embedded platform to achieve real-time fault diagnosis. .
It improves the accuracy and efficiency of bearing fault diagnosis, enhances the real-time diagnosis, and integrates the system into the embedded platform, simplifies industrial mechanical equipment detection and realizes the miniaturization of the model.
Smart Images

Figure CN118936890A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of intelligent diagnosis of mechanical faults, and in particular to a real-time diagnosis system for bearing faults based on an LSTM-CNN fusion model. Background Art
[0002] In the prior art, for example, CN1106889130A discloses a bearing fault diagnosis method, which includes the following steps: obtaining a bearing target domain sample set and a source domain sample set; obtaining source weight coefficients of different source domain BN model parameters according to the size of the source domain data sample set; calculating the balance coefficient α of the target domain initial BN model parameter 1 and the balance coefficient α of the total source domain BN model parameters 2 ; Calculate each source domain BN model parameter and the target domain initial BN model parameter based on the traditional parameter learning method; Calculate the target domain final BN model parameter using the parameter fusion method. The present invention uses the learned target domain BN model and the existing reasoning algorithm to perform bearing fault diagnosis, making full use of the migration mechanism to obtain data and information from similar fields, reducing the impact of insufficient sample size on parameter learning accuracy and diagnostic results;
[0003] However, the above bearing fault diagnosis method does not evaluate the training results during the training of the sample set. Poor model training effect may lead to certain deviations in the reasoning diagnosis of bearing faults.
[0004] CN117251680B discloses a bearing fault diagnosis network, including a lightweight network ResDenseNet-CA and a high-precision network multi-scale cross-dense connection network (MSCDNet), both of which include a diagnosis network vibration signal preprocessing module, a feature extraction module and a fault diagnosis classification module; the first network reduces the number of dense blocks and dense layers, and adds residual connections layer by layer to achieve jump connections, reduce feature loss, and uses the characteristics of dense connections for feature reuse, so that high accuracy can be achieved while training quickly; the second network uses two parallel dense blocks of different convolution scales for cross-connection to perform multi-scale feature fusion, which greatly utilizes signal features, and adds SE attention mechanism between the dropout layer and the translation layer between dense blocks, so as to achieve high-precision fault diagnosis and have strong noise resistance;
[0005] The bearing fault diagnosis technology currently used in engineering can be divided into signal processing methods and artificial intelligence methods. Since signal processing methods are overly dependent on expert experience, and artificial intelligence has more advantages when facing massive data. However, there are neurons in the neural network architecture of deep learning. The more neurons there are, the stronger the learning ability of the network is. But at the same time, the design of a large number of neurons will also consume a considerable amount of computing resources and computing time, which greatly limits its application in embedded devices with limited computing power. How to process data more efficiently and diagnose faults accurately is a problem that needs to be solved urgently.
[0006] Therefore, a real-time bearing fault diagnosis system based on an LSTM-CNN fusion model is provided to solve the above problems. Summary of the invention
[0007] The purpose of the present invention is to provide a real-time bearing fault diagnosis system based on the LSTM-CNN fusion model, which can effectively solve the problems of low accuracy and poor real-time performance of bearing fault diagnosis in industrial sites. It can not only quickly process vibration acceleration time series data, but also extract more obvious features of the data, so that the fault diagnosis system can output diagnosis results more accurately and efficiently.
[0008] To achieve the above object, the present invention provides a real-time bearing fault diagnosis system based on LSTM-CNN fusion model, which specifically includes the following steps:
[0009] S1: Build a signal acquisition experimental platform to collect bearing vibration acceleration data through vibration sensors;
[0010] S2: preprocess the bearing vibration acceleration data and construct the fault bearing data set;
[0011] S3: Input the faulty bearing dataset into the LSTM-CNN fusion model, add the attention mechanism LSTM layer, redistribute the fault weights, highlight the fault type characteristics, train and save the fusion model parameters;
[0012] S4: Porting the fusion model to the Raspberry Pi 4B embedded platform;
[0013] S5: Use Raspberry Pi 4B as the main control system to output the bearing fault type in real time;
[0014] S6: Design the host computer software.
[0015] Preferably, in step S1, the signal acquisition experimental platform includes an equipment fault prediction and health management PHM experimental platform, a faulty bearing and a vibration sensor, and specifically includes the following steps:
[0016] S11: The motor speed of the PHM experimental platform for equipment fault prediction and health management was set to 1400 rpm, and the faulty bearing contacted the motor shaft at the same position;
[0017] S12: The vibration sensor collects the vibration acceleration data of the faulty bearing. The sampling frequency of the vibration sensor is set to 25000 Hz. The highest frequency component of the industrial bearing does not exceed 6000 Hz. In the Nyquist sampling theorem, the relationship between the sampling rate and the measured signal frequency is as follows:
[0018] f s >2f N
[0019] where f s Sampling rate, f N It is the highest frequency component of the measured signal.
[0020] Preferably, in step S2, the specific process is as follows:
[0021] S21: Set seven types of bearings to construct data sets, and a single type of bearing data set includes 250,000 data points;
[0022] S22: Split the acquired bearing vibration acceleration data into single samples with a length of 1024, and the number of samples of a single type is greater than 250, to form a faulty bearing data set.
[0023] Preferably, in step S3, the specific process is as follows:
[0024] S31: normalizing the input faulty bearing data set, and dividing the faulty bearing data set into a training set, a validation set, and a test set;
[0025] S32: The convolutional neural network LSTM-CNN extracts features from the faulty bearing vibration acceleration data, adds an attention mechanism LSTM layer, redistributes fault weights, and trains and saves the fusion model parameters.
[0026] Preferably, in step S32, the specific process is as follows:
[0027] S321: Input the faulty bearing vibration acceleration data into the LSTM-CNN fusion model in the form of a four-dimensional tensor, and set the shape to (Batch_Size, 1, 1024, 1), where Batch_Size is the batch size, the first "1" represents the number of channels, "1024" represents the length of a single sample, and the second "1" represents the width. The faulty bearing vibration acceleration data passes through several convolutional layers, batch normalization layers, and activation function layers in sequence. The convolutional layers specifically include:
[0028] The first convolution layer contains 16 convolution kernels with a kernel size of (64, 1) and a stride of 1.
[0029] The second convolution layer contains 32 convolution kernels with a kernel size of (5, 1) and a stride of 2.
[0030] The third convolution layer contains 64 convolution kernels with a kernel size of (3, 1) and a stride of 2.
[0031] The fourth convolutional layer contains 128 convolution kernels with a kernel size of (3, 1) and a stride of 2.
[0032] The fifth convolutional layer contains 256 convolution kernels with a kernel size of (3, 1) and a stride of 2.
[0033] S322: Input the extracted acceleration features into the long short-term memory network with the attention mechanism added, and input the fault feature y obtained in step S321 into the long short-term memory network. i Multiply it by the neuron weight w, add the neuron bias, and then use the hyperbolic tangent function to convert the value u calculated above into i Mapped to the interval [-1,1], the calculation formula is as follows:
[0034] u i =tanh(wy i +b)
[0035] Where tanh is the hyperbolic tangent function, w is the neuron weight, and b is the neuron bias;
[0036] Use the Softmax function to calculate the weights of different features a i , and then to a i Normalization processing to obtain all feature attention weights a i The probability distribution of the sum of 1 is calculated as follows:
[0037]
[0038] The input of the fault feature is multiplied by the weight coefficient and summed up to obtain the final fault feature expression after the attention mechanism optimization:
[0039] F c =Σ(a i *y i )
[0040] Add Dropout layer and Dropconnect layer to the model to discard or set some neurons to 0.
[0041] Preferably, in step S4, the specific process is as follows:
[0042] S41: Building a virtual environment on Raspberry Pi 4B;
[0043] S42: Save the trained fusion model in .pkl format, compress the fusion model into a .onnx file, use the openvino toolkit to convert the fusion model into an IR intermediate model. The intermediate model format is .xml or .bin. Download and install openvino on the Raspberry Pi 4B to deploy and infer the intermediate model.
[0044] Preferably, in step S5, the specific process is as follows:
[0045] S51: The vibration sensor and the Raspberry Pi 4B communicate via Ethernet. The vibration sensor transmits the vibration acceleration data packet of the faulty bearing to the Raspberry Pi 4B, and the Raspberry Pi 4B parses the vibration acceleration data packet of the faulty bearing.
[0046] S52: Based on the computing and reasoning capabilities of the Raspberry Pi 4B and the LSTM-CNN model's ability to process and analyze time series data, the fault type of the bearing is output in real time and in a loop.
[0047] Preferably, in step S6, the specific design of the host computer software is as follows:
[0048] 1. Real-time diagnosis of faulty bearings The host computer software communicates with the vibration sensor and outputs the time-frequency domain signal of the bearing vibration acceleration in real time;
[0049] 2. The host computer software uses the deep groove ball bearing characteristic frequency calculation formula to calculate the bearing characteristic fault frequency by inputting the bearing model, inner ring outer diameter, number of balls and contact angle parameters;
[0050] 3. The upper computer software has a built-in bearing fault diagnosis algorithm, which analyzes and diagnoses the bearing fault type through the real-time collection of bearing vibration acceleration data.
[0051] Therefore, the present invention adopts the above-mentioned bearing fault real-time diagnosis system based on LSTM-CNN fusion model, which has the following beneficial effects:
[0052] (1) This scheme has high diagnostic efficiency. It adopts the LSTM-CNN fusion model, which can not only quickly process the bearing vibration acceleration data, but also extract more obvious features of the data, so that the fault diagnosis system can output diagnostic results more accurately and efficiently.
[0053] (2) This solution has strong real-time diagnosis performance. The host computer software for real-time diagnosis of faulty bearings communicates with the vibration sensor to monitor and diagnose the bearing status in real time, thus realizing system intelligence.
[0054] (3) This solution has a high degree of system integration, integrating the entire fault diagnosis system into an embedded platform, making industrial machinery equipment detection simpler and more convenient and realizing model miniaturization.
[0055] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS
[0056] Figure 1 is a system fault diagnosis flow chart of the present invention;
[0057] Figure 2 It is a structural diagram of the fault diagnosis LSTM-CNN model of the present invention;
[0058] Figure 3 It is the confusion matrix output by the LSTM-CNN fusion model training of the present invention. DETAILED DESCRIPTION
[0059] The technical solution of the present invention is further described below through the accompanying drawings and embodiments.
[0060] Unless otherwise defined, technical or scientific terms used in the present invention shall have the common meanings understood by one having ordinary skills in the field to which the present invention belongs.
[0061] The words "include" or "comprises" and the like used in the present invention mean that the elements before the word include the elements listed after the word, and do not exclude the possibility of also including other elements. The orientation or position relationship indicated by the terms "inside", "outside", "upper", "lower", etc. is based on the orientation or position relationship shown in the drawings, which is only for the convenience of describing the present invention and simplifying the description, and does not indicate or imply that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation. Therefore, it cannot be understood as a limitation of the present invention. When the absolute position of the described object changes, the relative position relationship may also change accordingly. In the present invention, unless otherwise clearly specified and limited, the terms "attachment" and the like should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral body; it can be directly connected, or indirectly connected through an intermediate medium, and it can be the internal connection of two elements or the interaction relationship between two elements. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.
[0062] Example
[0063] like Figure 1 As shown, the present invention provides a real-time bearing fault diagnosis system based on the LSTM-CNN fusion model, which specifically includes the following steps:
[0064] S1: Build a signal acquisition experimental platform to collect bearing vibration acceleration data through vibration sensors;
[0065] S2: preprocess the bearing vibration acceleration data and construct the fault bearing data set;
[0066] S3: Input the faulty bearing dataset into the LSTM-CNN fusion model, add the attention mechanism LSTM layer, redistribute the fault weights, highlight the fault type characteristics, train and save the fusion model parameters;
[0067] S4: Porting the fusion model to the Raspberry Pi 4B embedded platform;
[0068] S5: Use Raspberry Pi 4B as the main control system to output the bearing fault type in real time;
[0069] S6: Design the host computer software.
[0070] In step S1, the signal acquisition experimental platform includes an equipment fault prediction and health management PHM experimental platform, a faulty bearing and a vibration sensor, and specifically includes the following steps:
[0071] S11: The motor speed of the PHM experimental platform for equipment fault prediction and health management was set to 1400 rpm, and the faulty bearing contacted the motor shaft at the same position;
[0072] S12: The vibration sensor collects the vibration acceleration data of the faulty bearing. The sampling frequency of the vibration sensor is set to 25000 Hz. The highest frequency component of the industrial bearing does not exceed 6000 Hz. In the Nyquist sampling theorem, the relationship between the sampling rate and the measured signal frequency is as follows:
[0073] f s >2f N
[0074] where f s Sampling rate, f N It is the highest frequency component of the measured signal. According to the Nyquist sampling theorem, the sampling frequency of the vibration sensor is set to 25000 Hz to meet the sampling requirements of the bearing acceleration signal.
[0075] The specific process of step S2 is as follows:
[0076] S21: Set seven types of bearings to construct data sets, namely normal bearings, inner ring 0.8mm scratch fault bearings IR8, outer ring 0.8mm scratch fault bearings OR8, rolling element 0.8mm dot fault bearings BALL4, inner ring 0.4mm scratch fault bearings IR4, outer ring 0.4mm scratch fault bearings OR4, rolling element 0.4mm dot fault bearings BALL4 and inner and outer ring fault bearings IR+OR. The sampling time for a single type of bearing is 10 seconds, and the data set constructed for a single type of bearing includes 250,000 data points.
[0077] S22: Split the acquired bearing vibration acceleration data into single samples with a length of 1024, and the number of samples of a single type is 280, forming a faulty bearing data set.
[0078] The specific process of step S3 is as follows:
[0079] S31: normalizing the input faulty bearing data set, and dividing the faulty bearing data set into a training set, a validation set, and a test set;
[0080] S32: The convolutional neural network LSTM-CNN extracts features from the faulty bearing vibration acceleration data, adds an attention mechanism LSTM layer, redistributes fault weights, and trains and saves the fusion model parameters.
[0081] like Figure 2 As shown, the specific process of step S32 is as follows:
[0082] S321: Input the faulty bearing vibration acceleration data into the LSTM-CNN fusion model in the form of a four-dimensional tensor, and set the shape to (Batch_Size, 1, 1024, 1), where Batch_Size is the batch size, the first "1" represents the number of channels, "1024" represents the length of a single sample, and the second "1" represents the width. The faulty bearing vibration acceleration data passes through several convolutional layers, batch normalization layers, and activation function layers in sequence. The convolutional layers specifically include:
[0083] The first convolution layer contains 16 convolution kernels with a kernel size of (64, 1) and a stride of 1.
[0084] The second convolution layer contains 32 convolution kernels with a kernel size of (5, 1) and a stride of 2.
[0085] The third convolution layer contains 64 convolution kernels with a kernel size of (3, 1) and a stride of 2.
[0086] The fourth convolutional layer contains 128 convolution kernels with a kernel size of (3, 1) and a stride of 2.
[0087] The fifth convolutional layer contains 256 convolution kernels with a kernel size of (3, 1) and a stride of 2.
[0088] S322: Input the extracted acceleration features into the long short-term memory network with the attention mechanism added, and input the fault feature y obtained in step S321 into the long short-term memory network. i Multiply it by the neuron weight w, add the neuron bias, and then use the hyperbolic tangent function to convert the value u calculated above into i Mapped to the interval [-1,1], the calculation formula is as follows:
[0089] u i =tanh(wy i +b)
[0090] Where tanh is the hyperbolic tangent function, w is the neuron weight, and b is the neuron bias;
[0091] Use the Softmax function to calculate the weights of different features a i , and then to a i Normalization processing to obtain all feature attention weights a i The probability distribution of the sum of 1 is calculated as follows:
[0092]
[0093] The input of the fault feature is multiplied by the weight coefficient and summed up to obtain the final fault feature expression after the attention mechanism optimization:
[0094] F c =∑(a i *y i )
[0095] Add Dropout layer and Dropconnect layer to the model to discard or set some neurons to 0 to prevent overfitting of the model and increase generalization ability.
[0096] The specific process of step S4 is as follows:
[0097] S41: Building a virtual environment on Raspberry Pi 4B;
[0098] S42: Save the trained fusion model in .pkl format, compress the fusion model into a .onnx file, use the openvino toolkit to convert the fusion model into an IR intermediate model. The intermediate model format is .xml or .bin. Download and install openvino on the Raspberry Pi 4B to deploy and infer the intermediate model.
[0099] The specific process of step S5 is as follows:
[0100] S51: Place the vibration sensor at the driving end of the motor shaft of the PHM experimental platform. The vibration sensor and Raspberry Pi 4B communicate through Ethernet. Enter the "ping 192.168.1.164" command on the Raspberry Pi 4B to check whether the vibration sensor and Raspberry Pi 4B are successfully connected. After successful communication, enter the "source myenv / bin / activate" command on the Raspberry Pi 4B to enter the virtual environment that has been built. In the virtual environment, enter the "cd 1 / pytorch / demo / test" command on the Raspberry Pi 4B to enter the folder for implementing diagnostic tests. In the "test" folder, enter the "python3test.py" command in the Raspberry Pi 4B terminal to run the "test.py" python file in the "test" folder to start collecting bearing vibration acceleration data. The vibration sensor transmits the faulty bearing vibration acceleration data packet to the Raspberry Pi 4B, and the Raspberry Pi 4B parses the faulty bearing vibration acceleration data packet.
[0101] S52: Based on the computing and reasoning capabilities of the Raspberry Pi 4B and the LSTM-CNN model's ability to process and analyze time series data, the fault type of the bearing is output in real time and cyclically for every 100 sample data collected.
[0102] like Figure 3 As shown in the figure, the trained LSTM-CNN model is tested on the data set, the confusion matrix is output, and the performance is compared with that of the CNN, LSTM, VGG, and LeNet-5 models. According to the experimental results, it can be concluded that the training speed and accuracy of the LSTM-CNN model in this scheme are ahead of other models. The specific experimental results are shown in the table:
[0103] Table 1
[0104]
[0105] The average training accuracy of the LSTM-CNN model is 99.97%, and the average test set accuracy is 99.83%. Therefore, it can be concluded that the LSTM-CNN model has a strong ability to process and analyze time series data, and is superior to the CNN, LSTM, VGG and LeNet-5 models.
[0106] In step S6, the specific design of the host computer software is as follows:
[0107] 1. Real-time diagnosis of faulty bearings The host computer software communicates with the vibration sensor and outputs the time-frequency domain signal of the bearing vibration acceleration in real time;
[0108] 2. The host computer software uses the deep groove ball bearing characteristic frequency calculation formula to calculate the bearing characteristic fault frequency by inputting the bearing model, inner ring outer diameter, number of balls and contact angle parameters;
[0109] 3. The upper computer software has a built-in bearing fault diagnosis algorithm, which analyzes and diagnoses the bearing fault type through the real-time collection of bearing vibration acceleration data.
[0110] Therefore, the present invention adopts the above-mentioned real-time bearing fault diagnosis system based on the LSTM-CNN fusion model. The LSTM-CNN fusion model can not only quickly process the bearing vibration acceleration data, but also extract more obvious features of the data, so that the fault diagnosis system can output the diagnosis results more accurately and efficiently, thereby improving the fault diagnosis efficiency; the real-time diagnosis host computer software of the faulty bearing is connected and communicated with the vibration sensor, and the bearing status is monitored and diagnosed in real time to realize the intelligence of the system and enhance the real-time nature of the diagnosis; the entire fault diagnosis system is integrated into the embedded platform, so that the detection of industrial machinery and equipment is simpler and more convenient, the model is miniaturized, and the system is more integrated.
[0111] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that they can still modify or replace the technical solution of the present invention with equivalents, and these modifications or equivalent replacements cannot cause the modified technical solution to deviate from the spirit and scope of the technical solution of the present invention.
Claims
1. A real-time bearing fault diagnosis system based on LSTM-CNN fusion model, characterized by: The specific steps include: S1: Build a signal acquisition experimental platform to collect bearing vibration acceleration data through vibration sensors; S2: preprocess the bearing vibration acceleration data and construct the fault bearing data set; S3: Input the faulty bearing dataset into the LSTM-CNN fusion model, add the attention mechanism LSTM layer, redistribute the fault weights, highlight the fault type characteristics, train and save the fusion model parameters; S4: Porting the fusion model to the Raspberry Pi 4B embedded platform; S5: Use Raspberry Pi 4B as the main control system to output the bearing fault type in real time; S6: Design the host computer software.
2. A bearing fault real-time diagnosis system based on LSTM-CNN fusion model according to claim 1, characterized in that: In step S1, the signal acquisition experimental platform includes an equipment fault prediction and health management PHM experimental platform, a faulty bearing and a vibration sensor, and specifically includes the following steps: S11: The motor speed of the PHM experimental platform for equipment fault prediction and health management was set to 1400 rpm, and the faulty bearing contacted the motor shaft at the same position; S12: The vibration sensor collects the vibration acceleration data of the faulty bearing. The sampling frequency of the vibration sensor is set to 25000 Hz. The highest frequency component of the industrial bearing does not exceed 6000 Hz. In the Nyquist sampling theorem, the relationship between the sampling rate and the measured signal frequency is as follows: in s >2f N where f s Sampling rate, f N It is the highest frequency component of the measured signal.
3. A bearing fault real-time diagnosis system based on LSTM-CNN fusion model according to claim 1, characterized in that: In step S2, the specific process is as follows: S21: Set seven types of bearings to construct data sets, and a single type of bearing data set includes 250,000 data points; S22: Split the acquired bearing vibration acceleration data into single samples with a length of 1024, and the number of samples of a single type is greater than 250, to form a faulty bearing data set.
4. A bearing fault real-time diagnosis system based on LSTM-CNN fusion model according to claim 1, characterized in that: In step S3, the specific process is as follows: S31: normalizing the input faulty bearing data set, and dividing the faulty bearing data set into a training set, a validation set, and a test set; S32: The convolutional neural network LSTM-CNN extracts features from the faulty bearing vibration acceleration data, adds an attention mechanism LSTM layer, redistributes fault weights, and trains and saves the fusion model parameters.
5. A bearing fault real-time diagnosis system based on LSTM-CNN fusion model according to claim 4, characterized in that: In step S32, the specific process is as follows: S321: Input the faulty bearing vibration acceleration data into the LSTM-CNN fusion model in the form of a four-dimensional tensor, and set the shape to (Batch_Size, 1, 1024, 1), where Batch_Size is the batch size, the first "1" indicates the number of channels, "1024" indicates the length of a single sample, and the second "1" indicates the width. The faulty bearing vibration acceleration data passes through several convolutional layers, batch normalization layers, and activation function layers in sequence. The convolutional layers specifically include: The first convolution layer contains 16 convolution kernels with a kernel size of (64, 1) and a stride of 1. The second convolution layer contains 32 convolution kernels with a kernel size of (5, 1) and a stride of 2. The third convolution layer contains 64 convolution kernels with a kernel size of (3, 1) and a stride of 2. The fourth convolutional layer contains 128 convolution kernels with a kernel size of (3, 1) and a stride of 2. The fifth convolutional layer contains 256 convolution kernels with a kernel size of (3, 1) and a stride of 2. S322: Input the extracted acceleration features into the long short-term memory network with the attention mechanism added, and input the fault feature y obtained in step S321 into the long short-term memory network. i Multiply it by the neuron weight w, add the neuron bias, and then use the hyperbolic tangent function to convert the value u calculated above into i Mapped to the interval [-1,1], the calculation formula is as follows: u i =tanh(wy i +b) Where tanh is the hyperbolic tangent function, w is the neuron weight, and b is the neuron bias; Use the Softmax function to calculate the weights of different features a i , and then to a i Normalization processing to obtain all feature attention weights a i The probability distribution of the sum of 1 is calculated as follows: The input of the fault feature is multiplied by the weight coefficient and summed up to obtain the final fault feature expression after the attention mechanism optimization: F c =∑(a i *y i ) Add Dropout layer and Dropconnect layer to the model to discard or set some neurons to 0.
6. A bearing fault real-time diagnosis system based on LSTM-CNN fusion model according to claim 1, characterized in that: In step S4, the specific process is as follows: S41: Building a virtual environment on Raspberry Pi 4B; S42: Save the trained fusion model in .pkl format, compress the fusion model into a .onnx file, use the openvino toolkit to convert the fusion model into an IR intermediate model. The intermediate model format is .xml or .bin. Download and install openvino on the Raspberry Pi 4B to deploy and infer the intermediate model.
7. A bearing fault real-time diagnosis system based on LSTM-CNN fusion model according to claim 1, characterized in that: In step S5, the specific process is as follows: S51: The vibration sensor and the Raspberry Pi 4B communicate via Ethernet. The vibration sensor transmits the vibration acceleration data packet of the faulty bearing to the Raspberry Pi 4B, and the Raspberry Pi 4B parses the vibration acceleration data packet of the faulty bearing. S52: Based on the computing and reasoning capabilities of the Raspberry Pi 4B and the LSTM-CNN model's ability to process and analyze time series data, the fault type of the bearing is output in real time and in a loop.
8. A bearing fault real-time diagnosis system based on LSTM-CNN fusion model according to claim 1, characterized in that: In step S6, the specific design of the host computer software is as follows:
1. Real-time diagnosis of faulty bearings The host computer software communicates with the vibration sensor and outputs the time-frequency domain signal of the bearing vibration acceleration in real time; 2. The host computer software uses the deep groove ball bearing characteristic frequency calculation formula to calculate the bearing characteristic fault frequency by inputting the bearing model, inner ring outer diameter, number of balls and contact angle parameters; 3. The upper computer software has a built-in bearing fault diagnosis algorithm, which analyzes and diagnoses the bearing fault type through the real-time collection of bearing vibration acceleration data.
Citation Information
Cited By
Forging and pressing equipment detection method and system based on multi-source data fusion
CN120372253A
Industrial bearing vibration time sequence signal fault prediction method and system fusing attention mechanism and LSTM
CN121144702A
