Rolling bearing fault diagnosis method based on temperature and vibration information fusion

Abstract

The application provides a rolling bearing fault diagnosis method based on temperature and vibration information fusion, adopts multi-sensor data for fault diagnosis, redundant and complementary fault information can be obtained through the multi-sensor data, and information fusion is carried out to obtain more accurate and reliable diagnosis results; temperature signals and vibration acceleration signals are used as diagnosis basis, and can be applied to the actual application of intelligent motor train unit axle box bearing fault diagnosis; a new decision fusion method is proposed to solve the temperature and vibration information fusion problem of the axle box bearing of the high-speed motor train unit, compared with the data fusion and feature fusion methods, the method proposed by the application is simple to operate and easy to implement.

Description

Technical Field

[0001] This invention relates to the field of next-generation information technology and is applicable to the axle box bearings of the running gear of intelligent EMUs, realizing a fault diagnosis method based on multi-sensor data of temperature and vibration. Background Technology

[0002] Axle box bearings are one of the key components of high-speed trains. Due to their high operating speeds and complex service environments, they are among the most vulnerable parts of the train. During train operation, bearing failure not only has a significant impact on the railway system but also poses a threat to personal safety. Therefore, research on real-time fault early warning systems for axle box bearings in high-speed trains is of great value.

[0003] Currently, all high-speed trains in my country are equipped with axle temperature sensors, and some trains also have vibration acceleration sensors installed at the axle box bearings. Existing methods for diagnosing axle box bearing faults in high-speed trains are mostly based on vibration or temperature signals. However, temperature-based bearing fault diagnosis methods are often only effective for poor lubrication and severe mechanical damage, and cannot diagnose early-stage bearing faults or their locations. Vibration-based methods can identify early-stage bearing faults and their locations, but cannot reflect the bearing's lubrication status. Therefore, considering both temperature and vibration acceleration signals simultaneously during bearing fault diagnosis could improve the accuracy of rolling bearing fault diagnosis. To this end, this invention provides a bearing fault diagnosis method based on the fusion of temperature and vibration information. Summary of the Invention

[0004] The purpose of this invention is to avoid incomplete diagnostic information caused by using only vibration or temperature signals in bearing fault diagnosis. By fusing both information and combining them with deep learning methods, a bearing fault diagnosis method based on temperature and vibration information fusion is proposed, thereby improving the accuracy of bearing fault diagnosis.

[0005] To achieve the above objectives, this invention provides a method for diagnosing rolling bearing faults based on temperature and vibration information fusion, specifically including the following steps:

[0006] Step 1: Collect bearing temperature and vibration data from the axle box, and then extract the bearing temperature and vibration characteristics in sequence as inputs for the temperature and vibration channels;

[0007] The steps for extracting bearing vibration characteristics are as follows: Select the mean z1, effective value z2, variance z3, square root amplitude z4, absolute mean z5, peak value z6, kurtosis coefficient z7, maximum value z8, minimum value z9, and peak-to-peak value z1 of the vibration time-domain signal. 10 skewness coefficient z 11 Shape factor z 12 Peak factor z13 , pulse factor z 14 , clearance factor z 15 and the center frequency of the frequency domain signal z 16 , effective value variance frequency z 17 , root variance frequency z 18 as a vibration characteristic reflecting the bearing state, form a vibration characteristic vector Z vib = [z1, z2, z3, …, z 18 ] T ,

[0008] Let the vibration signal of length N be x(t), and the calculation formula of the vibration characteristic is as follows:

[0009]

[0010]

[0011]

[0012]

[0013]

[0014] z6 = max |x i |

[0015]

[0016] z8 = max (x i )

[0017] z9 = min (x i )

[0018] z 10 = z8 - z9

[0019]

[0020]

[0021]

[0022]

[0023]

[0024]

[0025]

[0026]

[0027] The vibration feature vector is then input into a BP neural network, and combined with a Softmax classifier to achieve preliminary identification of bearing fault modes based on vibration features.

[0028] The steps for extracting bearing temperature features are as follows: A normal bearing temperature signal reconstruction model is constructed using an autoencoder. By quantifying the reconstruction error of the bearing temperature signal under different fault modes in this model, temperature features reflecting the bearing's health status are extracted. The specific operation for temperature feature extraction based on the autoencoder is as follows:

[0029] (1) Training of normal bearing temperature signal reconstruction model: Screen the temperature signals of normal bearings to construct dataset A, and randomly select 70% of the data as the training set, 10% as the validation set, and 20% as the test set to train the normal bearing temperature signal reconstruction model:

[0030] a) Model parameter initialization settings: Set the hyperparameters for training the autoencoder model, including the learning rate λ and the acceptable cost function value J. min and the upper limit of the loop n s Initialize the weight matrix W from the input layer to the hidden layer. (1) and bias matrix b (1) The weight matrix W from the hidden layer to the output layer (2) and bias matrix b (2) .

[0031] b) Encoder feature extraction: First, calculate the weighted sum of the inputs z for each neuron. (2) =W (1) X t +b (1) Then calculate its activation value a. (2) =g(z) (2) ), where g(·) is the activation function, a (2) This is the matrix composed of the values ​​of each neuron in the hidden layer, and it is also the extracted feature.

[0032] c) Decoder reconstructs the signal: Calculate the reconstructed signal The process is as shown in b), and finally... Then calculate the cost function value. And let n = n + 1;

[0033] d) Determine if n is less than n s Or is J greater than J? min If so, then update the weight matrix. bias matrix Return to b), if not, then training ends;

[0034] (3) Error feature extraction: input the temperature signal of each fault mode into the normal bearing temperature signal reconstruction model in sequence for reconstruction, obtain the corresponding reconstruction signal of each sample, and then quantize the reconstruction error of each sample in the data set A1, and use L-2 norm to quantize the difference between the to-be-detected signal and the reconstruction signal, and the formula is as follows:

[0035]

[0036] In the formula, T is X T The set of moments corresponding to each element in the middle.

[0037] At the same time, considering that poor bearing lubrication and serious mechanical damage of its parts can cause the bearing temperature to rise sharply or the bearing temperature to be significantly higher than the normal level, therefore, the maximum temperature rise rate is also used to feedback the service state, and the calculation formula is:

[0038]

[0039] Further, the temperature norm feature, the maximum temperature rise rate and the original temperature signal are combined to form a temperature feature Z temp As the input of the BP neural network, the bearing fault mode preliminary identification based on the temperature feature is realized;

[0040] Step 2: combine the preliminary mode identification result based on the temperature feature in step 1 and the preliminary mode identification result based on the vibration feature by a decision fusion method to obtain the final diagnosis result;

[0041] The specific process of the decision fusion method is as follows:

[0042] (1) collect the temperature signals and vibration signals of the rolling bearing under different fault modes, and divide the temperature signals and vibration signals into training samples, verification samples and test samples;

[0043] (2) use the training samples to construct the pre-diagnosis model composed of step 1, that is, use the temperature training samples to construct and train the temperature pre-diagnosis model, and use the vibration training samples to construct and train the vibration pre-diagnosis model;

[0044] (3) use the verification samples to calculate the weight of the pre-diagnosis model, input the temperature verification samples and vibration verification samples into the trained temperature pre-diagnosis model and vibration pre-diagnosis model in sequence, and calculate the recognition accuracy and information entropy of each model on the verification set:

[0045]

[0046]

[0047] In the formula, i is the code of the pre-diagnosis model, The total number of samples of the verification set of the i-th pre-diagnosis model, The number of samples correctly diagnosed by the i-th pre-diagnosis model on the verification set, further, the weight of calculating the recognition accuracy:

[0048]

[0049] (4) Determine the temperature-vibration decision fusion diagnosis result of the test sample, wherein the pre-diagnosis result given by each pre-diagnosis model is Calculate the score S of the c-th fault mode c , c=1, 2, …, C, C is the total number of fault modes:

[0050]

[0051] In the formula, the calculation formula of I(·) is as follows:

[0052]

[0053] Finally, the one with the highest score is selected as the temperature-vibration decision fusion diagnosis result.

[0054] The present application has the advantages and progress:

[0055] 1. The present application adopts multi-sensor data for fault diagnosis, and the existing bearing fault diagnosis method is mostly based on a single signal, which has the problem of limited single signal information. The present application can obtain redundant and complementary fault information through multi-sensor data, and fuse the information to obtain more accurate and reliable diagnosis results.

[0056] 2. The present application can be applied to the fault diagnosis of intelligent motor train unit axle box bearings. The existing multi-sensor bearing fault model mostly uses vibration signals of multiple positions or multiple directions as model inputs, which is inconsistent with the axle box bearing monitoring data that can be collected by high-speed motor train units. The model cannot be applied to high-speed motor train unit axle box bearings. The present application considers the actual situation of high-speed motor train unit axle box bearing monitoring data, uses temperature signals and vibration acceleration signals as diagnosis basis, and can be applied to the practical application of intelligent motor train unit axle box bearing fault diagnosis.

[0057] 3. The present application solves the problem of temperature-vibration information fusion of high-speed motor train unit axle box bearings by proposing a new decision fusion method. Compared with the data fusion and feature fusion methods, the method proposed by the present application is simple to operate and easy to implement. DETAILED DESCRIPTION

[0058] Figure 1 The diagnostic method flowchart of the present application;

[0059] Figure 2 The BP neural network structure schematic diagram;

[0060] Figure 3 is a schematic diagram of a self-encoder structure;

[0061] Figure 4 is a temperature signal reconstruction model hidden layer number screening process. DETAILED DESCRIPTION

[0062] The technical solution of the present application is first to collect the bearing temperature data and vibration data of the shaft box, then to extract the bearing temperature features and vibration features in turn, to train the bearing fault pre-diagnosis model based on the temperature features and the bearing fault pre-diagnosis model based on the vibration features by using the BP neural network respectively, to obtain the pre-diagnosis results based on the temperature features and the pre-diagnosis results based on the vibration features, and finally to combine the two to obtain the final diagnosis result, as shown in the diagnosis flowchart Figure 1 , which can be divided into three parts of vibration feature extraction and fault mode pre-identification, temperature feature extraction and fault mode pre-identification, and temperature-vibration decision fusion diagnosis.

[0063] When the bearing fails during service, it will directly reflect in its vibration signal, and effective vibration features can be extracted to capture the bearing state information reflected by the signal. The present application selects the time domain mean z1, effective value z2, variance z3, square root amplitude value z4, absolute average value z5, peak value z6, kurtosis coefficient z7, maximum value z8, minimum value z9, peak-peak value z10, skewness coefficient z11, shape factor z12, peak factor z13, pulse factor z14, clearance factor z15, center frequency z16, effective value variance frequency z17, root variance frequency z18 of the vibration time domain signal and the center frequency z16, effective value variance frequency z17, root variance frequency z18 of the frequency domain signal as the vibration features reflecting the bearing state, to form the vibration feature vector Z = [z1, z2, z3, …, z18] 10 . 11 . 12 . 13 . 14 . 15 . 16 . 17 . 18 . vib . 18 . T .

[0064] Let the vibration signal of length N be x(t), and the calculation formula of the vibration feature is as follows:

[0065]

[0066] Then the vibration feature vector is input into the BP neural network, and the Softmax classifier is combined to realize the preliminary identification of the bearing fault mode based on the vibration features.

[0067] The BP artificial neural network structure is composed of an input layer, a hidden layer and an output layer, and the error is transmitted from the output layer to the hidden layer and then to the input layer according to the error back propagation process of the BP neural network, so that the optimal weight can be obtained by combining the error with the set learning rate.

[0068] The BP neural network structure is as shown in Figure 2 , wherein k represents the dimension of the output layer, i represents the dimension of the input layer, and j represents the number of hidden layers.

[0069]

[0070] When the network structure is determined, the network is composed of a set of input and output training sets, that is, the error is transmitted backward, the error is corrected, the weight and threshold are continuously adjusted, and the network can realize or approach the required input and output mapping relationship.

[0071] The learning process of the bearing fault diagnosis BP network driven by the vibration characteristics generally includes the following two steps:

[0072] The first step is a forward process of transmitting the vibration characteristics as input parameters through the input layer to the hidden layer and then to the output layer, wherein the output result of each layer is obtained by comparing the weight and threshold defined in the network structure with the previous propagation process.

[0073] The second step is a backward propagation process, which calculates the difference between the actual output and the expected output, wherein each layer is modified to minimize the error signal.

[0074] The two processes are alternated before convergence. The algorithm is described as an error back propagation algorithm because the error is fed back in each layer to correct the weighting coefficients and thresholds between the layers.

[0075] Bearing temperature anomalies occur less frequently in bearing bench tests, so the temperature anomaly data is generally less in the experimental data. In order to extract temperature characteristics that can effectively reflect the bearing state, the present application uses an autoencoder to construct a normal bearing temperature signal reconstruction model, and quantifies the reconstruction error of the bearing temperature signal under different fault modes in the model to extract temperature characteristics reflecting the bearing health state.

[0076] The autoencoder is an unsupervised learning network composed of an input layer, a hidden layer and an output layer, and a schematic diagram thereof is as shown in Figure 3 Unlike the BP neural network, the output of the autoencoder network is the reconstructed input signal, wherein the input signal is x, and the output signal is To make the output signal as similar as possible to the input signal, the network weights are updated by minimizing the average error to complete the training of the reconstructed model. In this invention, the specific operation of temperature feature extraction based on the autoencoder is as follows:

[0077] (1) Training of the normal bearing temperature signal reconstruction model. Temperature signals from normal bearings were selected to construct dataset A. Approximately 70% of the data was randomly selected as the training set, approximately 10% as the validation set, and approximately 20% as the test set to train the normal bearing temperature signal reconstruction model.

[0078] a) Model parameter initialization settings: Set the hyperparameters for training the autoencoder model, including the learning rate λ and the acceptable cost function value J. min and the upper limit of the loop n s Initialize the weight matrix W from the input layer to the hidden layer. (1) and bias matrix b (1) The weight matrix W from the hidden layer to the output layer (2) and bias matrix b (2) .

[0079] b) Encoder extracts features. First, the input weighted sum z of each neuron is calculated. (2) =W (1) X t +b (1) Then calculate its activation value a. (2) =g(z) (2) ), where g(·) is the activation function, a (2) This is the matrix composed of the values ​​of each neuron in the hidden layer, and it is also the extracted feature.

[0080] c) Decoder reconstructs the signal. Calculate the reconstructed signal. The process is as shown in b), and finally... Then calculate the cost function value. And let n = n + 1;

[0081] d) Determine if n is less than n s Or is J greater than J? min If so, then update the weight matrix. bias matrix Return to b), if not, then training ends;

[0082] (3) Error Feature Extraction. The temperature signals of each fault mode are sequentially input into the normal bearing temperature signal reconstruction model for reconstruction, obtaining the corresponding reconstructed signal for each sample. Then, the reconstruction error of each sample in dataset A1 is quantified, and the difference between the detected signal and the reconstructed signal is quantified using the L-2 norm, as shown in the following formula:

[0083]

[0084] In the formula, T is X T The set of time points corresponding to each element in the middle.

[0085] Meanwhile, considering that bearing lubrication is poor and its parts are severely mechanically damaged, both of which can cause the bearing temperature to rise sharply or the bearing temperature to be significantly higher than the normal level, therefore, the maximum temperature rise rate is also used to feedback the service state, and the calculation formula is:

[0086]

[0087] Further, the temperature norm feature, the maximum temperature rise rate and the original temperature signal are combined to form a temperature feature Z temp As the input of the BP neural network, the bearing fault mode preliminary identification based on the temperature feature is realized.

[0088] The voting method is a classical decision-level fusion method, after a series of pre-identification models are obtained, the pre-diagnosis results are given, the weight of each pre-diagnosis result is given, the score of the fault category is calculated, and then the highest one is selected as the final diagnosis result, so the key lies in the weight distribution of the pre-diagnosis result, for this, the present application proposes a new decision fusion method.

[0089] The method takes the information entropy of the historical diagnosis accuracy rate of the temperature and vibration model as the weight distribution index, and distributes the weight of each pre-diagnosis model.

[0090] The specific process is as follows:

[0091] (1) Collect the temperature signals and vibration signals under different fault modes of the rolling bearing, and divide the temperature signals and vibration signals into training samples, verification samples and test samples;

[0092] (2) Construct a pre-diagnosis model. The temperature training sample and the vibration training sample are used to construct a temperature pre-diagnosis model and a vibration pre-diagnosis model in turn;

[0093] (3) Calculate the weight of the pre-diagnosis model. The temperature verification sample and the vibration verification sample are input into the trained temperature pre-diagnosis model and vibration pre-diagnosis model in turn, and the identification accuracy and information entropy of each model on the verification set are calculated:

[0094]

[0095]

[0096] In the formula, i is the code of the pre-diagnosis model, is the total number of samples of the verification set of the i th pre-diagnosis model, The number of samples correctly diagnosed on the validation set for the ith pre-diagnosis model. Further, the weight of the recognition accuracy is calculated:

[0097]

[0098] (4) Determine the temperature-vibration decision fusion diagnosis result. Let the pre-diagnosis result given by each pre-diagnosis model be Calculate the score S of the cth fault mode c :

[0099]

[0100] In the formula, the calculation formula of I(·) is as follows:

[0101]

[0102] Finally, the one with the highest score is selected as the temperature-vibration decision fusion diagnosis result.

[0103] The specific application examples of the present application are as follows:

[0104] 1. Bearing data source

[0105] The present application requires a sufficient amount of temperature signals and vibration signals of various fault modes of bearings for training to obtain a better model. Therefore, the present application uses the bearing vibration acceleration signals and temperature signals collected by the bearing test bench of Southwest Jiaotong University to verify the effectiveness of the present method. Five bearing health states are considered, the bearing speed is 1100 rpm, the load is 1200 kg, the excitation table frequency is 10 Hz, the acceleration sensor sampling frequency is 25600 Hz, the total number of fault modes C involved in the present case is 4, and the dataset information is shown in Table 1. The specific parameters of the experimental rolling bearing are shown in Table 2.

[0106] Table 1 Detailed information of dataset A

[0107]

[0108] Table 2 Specific parameters of rolling bearing

[0109] Parameter name Inner ring diameter Outer ring diameter Rolling element diameter Pitch diameter Number of rollers Contact angle Parameter value 130 mm 230 mm 26.68 mm 196.38 mm 17 12.08°

[0110] The vibration features of dataset A in Table 1 are extracted and the fault mode is pre-identified to obtain an 18-dimensional vibration feature vector. After normalization, the vibration feature vector is input into the BP neural network. In the present case, the structure of the bearing fault pre-diagnosis model driven by the vibration feature is: 1 layer of hidden layer, and the number of hidden layer neurons is 16. The diagnosis accuracy on the training set and the validation set is 90% and 85%, respectively.

[0111] Meanwhile, temperature feature extraction and fault mode pre-identification are performed on the data set A in Table 1, in the process, the number of hidden layers of the temperature signal reconstruction model is 4, and the screening process is as shown in Figure 4 Taking data set A as an example, the mean and variance of the error eigenvalues in the bearing normal state, inner ring fault state, outer ring fault state and rolling element fault state are calculated respectively, and details are shown in Table 3, and then a 5-dimensional temperature feature vector is formed by combining the maximum temperature difference and the original temperature signal, and after normalization, the BP neural network is input, in this case, the structure of the bearing fault pre-diagnosis model driven by the temperature feature is: 1 hidden layer, the number of hidden layer neurons is 8, and finally the diagnostic accuracy on the training set and the validation set is 73%, 71%.

[0112] Table 3 Comparison of L-2 norms of rolling bearings under different working conditions

[0113]

[0114]

[0115] Further, the temperature prediction result and the vibration prediction result are fused on the test set, in order to verify the effectiveness of the rolling bearing fault diagnosis model of the temperature-vibration information fusion proposed, the effect of the diagnosis model using only the temperature feature (denoted as model I) and the diagnosis model using only the vibration feature (denoted as model II) is compared. The comparison result is shown in Table 4:

[0116] Table 4 Accuracy rate comparison of different models in rolling bearing fault mode identification

[0117]

[0118] From Table 4, it can be seen that compared with the bearing fault diagnosis method using only the temperature signal or only the vibration signal, the diagnosis accuracy of the rolling bearing fault diagnosis method of temperature-vibration fusion proposed in the application is higher.

[0119] The application mainly solves the following problems:

[0120] (1) Improve the accuracy of fault diagnosis, the bearing vibration signal and the temperature signal are simultaneously used as the basis for bearing fault diagnosis in the application, based on the complementarity of temperature and vibration data, more perfect information reflecting the bearing health state is obtained, the accuracy of bearing fault diagnosis is improved, and the stable and safe operation of the EMU is ensured.

[0121] (2) Ensure the practicality of the bearing fault model, the application carries out fault diagnosis around the axle box bearing of the high-speed EMU, and comprehensively considers the actual situation of the axle box bearing monitoring data, and through the research on the multi-sensor signal decision fusion method, the practicality of the model can be ensured.

Claims

1. A rolling bearing fault diagnosis method based on temperature and vibration information fusion, characterized in that: Specifically comprising the following steps: Step 1: Collecting the temperature data and vibration data of the bearing box bearing, and then extracting the bearing temperature features and vibration features in turn as the inputs of the temperature channel and the vibration channel; The extracting bearing vibration characteristic step is as follows: selecting mean value z1, effective value z2, variance z3, square root amplitude value z4, absolute average value z5, peak value z6, kurtosis coefficient z7, maximum value z8, minimum value z9, peak-peak value z10, skewness coefficient z11, shape factor z12, peak factor z13, pulse factor z14, clearance factor z15 and center frequency z16, effective value variance frequency z17, root variance frequency z18 of time domain signal and frequency domain signal. 10 11 12 13 14 15 16 17 18 The vibration characteristic vector Z vib = [z1, z2, z3, …, z 18 ] T ,​​​​​​​​ The vibration signal with a length of N is x(t), and the calculation formula of the vibration feature is as follows: z6 = max |x i | z8 = max(x i ) z9 = min(x i ) z 10 = z8 - z9 Then the vibration feature vector is input into the BP neural network, and the Softmax classifier is combined to realize the preliminary identification of the bearing fault mode based on the vibration feature; The bearing temperature feature extraction step is as follows: using an autoencoder to construct a normal bearing temperature signal reconstruction model, quantifying the reconstruction error of the bearing temperature signal under different fault modes in the model to extract the temperature feature reflecting the bearing health state; The specific operation of the temperature feature extraction based on the autoencoder is as follows: (1) Normal bearing temperature signal reconstruction model training; select the temperature signal of the normal bearing, construct a data set A, randomly select 70% of the data as the training set, 10% as the validation set, and 20% as the test set, and train the normal bearing temperature signal reconstruction model: a) Model parameter initialization settings: Set hyperparameters in the autoencoder model training, including learning rate λ, acceptable cost function value J min and upper limit of the loop n s , initialize the weight matrix W (1) and bias matrix b (1) from the input layer to the hidden layer (2) , the weight matrix W (2)； and bias matrix b from the hidden layer to the output layer b) Encoder feature extraction: First, calculate the weighted sum of the inputs z for each neuron. (2) =W (1) X t +b (1) Then calculate its activation value a. (2) =g(z) (2) ), where g(·) is the activation function, a (2) This is the matrix composed of the values ​​of each neuron in the hidden layer, and it is also the extracted feature. c) Decoder reconstructs signal: Compute reconstructed signal Process as in b) finally Further compute cost function value and let n = n + 1; d) determine if n is less than n s or J is greater than J min , if so, update the weight matrix bias matrix return to b), if not, training is complete; (2) Error feature extraction: input the temperature signal of each fault mode into the normal bearing temperature signal reconstruction model in turn for reconstruction, and then quantify the reconstruction error of each sample in the data set A1, and use L-2 norm to quantify the difference between the to-be-detected signal and the reconstructed signal, the formula is as follows: In the formula, T is X T a set of time instants corresponding to each element in the middle At the same time, considering that poor lubrication of the bearing and serious mechanical damage of its parts will cause the bearing temperature to rise sharply or the bearing temperature to be significantly higher than the normal level, therefore, the maximum temperature rise rate is also used to feedback the service state, and the calculation formula is as follows: Further, the temperature characteristic Z is formed by combining the temperature norm characteristic, the maximum temperature rising rate and the original temperature signal temp , as the input of BP neural network, the bearing fault mode preliminary identification based on temperature characteristic is realized. Step 2: Combine the preliminary mode identification results based on the temperature features and the preliminary mode identification results based on the vibration features in step 1 by using a decision fusion method to obtain the final diagnosis result; The specific process of the decision fusion method is as follows: (1) Collecting the temperature signal and vibration signal of the rolling bearing under different fault modes, and dividing the temperature signal and vibration signal into training samples, validation samples and test samples; (2) Using the training samples to construct the pre-diagnosis model formed in step 1, that is, using the temperature training samples to construct and train the temperature pre-diagnosis model, and using the vibration training samples to construct and train the vibration pre-diagnosis model; (3) Using the validation samples to calculate the weight of the pre-diagnosis model, inputting the temperature validation samples and vibration validation samples into the trained temperature pre-diagnosis model and vibration pre-diagnosis model in turn, and calculating the recognition accuracy and information entropy of each model on the validation set: In the formula, i is the code of the pre-diagnosis model, is the total number of samples in the validation set of the i th pre-diagnosis model, is the number of samples correctly diagnosed by the i th pre-diagnosis model on the validation set, and further, the weight for calculating the recognition accuracy is calculated: (4) determining the temperature-vibration decision fusion diagnosis result of the test sample, assuming that the pre-diagnosis results given by each pre-diagnosis model are calculating the score S of the cth fault mode c , c = 1, 2, …, C, C is the total number of fault modes: Wherein, the calculation formula of I(·) is as follows: Finally, the one with the highest score is selected as the temperature-vibration decision fusion diagnosis result.

Images