Fault diagnosis method based on data fusion and multi-channel convolutional network

Abstract

The application discloses a kind of based on data fusion and the fault diagnosis method of multichannel convolution network, first input bearing acceleration's multiple-source homogeneous sensor data, through kriging interpolation method, the missing data is completed, then the contribution rate of each data is calculated, and these data signals are fused by contribution rate;Then the normalized processing is carried out to the data after fusion, is input into multichannel convolution network, and the preliminary classification decision of fault is output;Finally, the output of multichannel convolution network is weighted, to obtain final decision.The application uses multichannel convolution network, and has good extraction effect on data feature trend, and solves the problem that traditional fault diagnosis method is not used properly for multiple-source data in the prior art.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of fault diagnosis, and particularly relates to a fault diagnosis method based on data fusion and a multi-channel convolution network. BACKGROUND

[0002] In the field of developing fault diagnosis, in order to improve the accuracy of the fault diagnosis method, the multi-sensor system has gradually matured, the quality of multi-source sensor data has a decisive role on the sufficiency and accuracy of the decision of the fault diagnosis system, however, when obtaining high-quality multi-source data, problems such as data missing, existing pseudo data, and too large data quantity are inevitable, how to effectively use multi-source data and improve the accuracy of fault diagnosis needs to establish a reasonable data processing method, therefore, the fault diagnosis research combined with the data fusion method has been widely valued in recent years. SUMMARY

[0003] The purpose of the application is to provide a fault diagnosis method based on data fusion and a multi-channel convolution network, which uses the multi-channel convolution network to have a good extraction effect on the data feature trend, and solves the problem of improper use of multi-source data by the traditional fault diagnosis method in the prior art.

[0004] The technical scheme adopted by the application is that the fault diagnosis method based on data fusion and a multi-channel convolution network is implemented according to the following steps:

[0005] Step 1, input the multi-source homogeneous sensor data of bearing acceleration, complete the missing data by the kriging interpolation method, then calculate the contribution rate of each data, and fuse these data signals by the contribution rate;

[0006] Step 2, normalize the fused data, input into the multi-channel convolution network, and output the preliminary classification decision of fault;

[0007] Step 3, perform weighted processing on the output of the multi-channel convolution network, so as to obtain the final decision.

[0008] The application has the characteristics that,

[0009] Step 1 is specifically as follows:

[0010] Step 1.1, first, filter the input multi-source homogeneous sensor data of bearing acceleration, perform Gaussian filtering on the obtained multi-source homogeneous data, and obtain the multi-source homogeneous data after Gaussian filtering;

[0011] Step 1.2, adopt the kriging interpolation method in the spatial interpolation method to perform data completion processing on the data obtained by different sensors, obtain a plurality of groups of homogeneous data with the same dimension, and calculate the estimated value of the unknown data point by formula (1):

[0012]

[0013] where H(z o ) is the estimated value of the property of the unknown sample point z o , c is the mean value of all sample points, h(z i ) is the property value of the known sample point z i , N is the number of known sample points within 5 units of the unknown sample point; i represents the i-th data point;

[0014] Step 1.3, calculate the contribution rate of the completed multi-source data obtained in step 1.2, the specific method is as follows:

[0015] First, calculate the correlation between the data signals, as shown in formula (2):

[0016]

[0017] where R i,j represents the correlation between the i-th signal and the j-th signal, x i (t0) represents the t0-th data point in signal i, t represents the selected data point, and n represents the number of data points in the signal.

[0018] Then, quantitatively calculate the obtained signal correlation as shown in formula (3):

[0019]

[0020] where E i is the correlation degree of signal i, m is the total number of signals, and j represents the j-th signal. The correlation degree is analyzed for importance as shown in formula (4):

[0021]

[0022] where y i is the information redundancy of signal i. By calculating the importance analysis, the representation of the amount of signal information is obtained. The mean and variance of the representation of the importance analysis after the correlation degree is calculated. The mean and variance are used to calculate the contribution rate of each signal as shown in formula (5):

[0023]

[0024] where k i represents the contribution rate of the i-th signal, y i is the importance analysis representation, μ is the mean of the importance analysis representation, and σ is the variance of the importance analysis representation. The normalization method is shown in formula (6):

[0025]

[0026] wherein, K i (t) is the contribution rate of data point t of the normalized signal i, and then the data points of each signal are multiplied by the weight in time sequence, as shown in formula (7):

[0027]

[0028] wherein, x(h) is the signal value of the fused data point h, and the linear addition of the corresponding points can obtain the signal after the data level fusion.

[0029] Step 2 is specifically as follows:

[0030] Step 2.1, the fused signal calculated in step 1.3 is first processed by dimension reduction, and the fused multi-dimensional signal is reduced to 3 dimensions by using principal component analysis, and then the signal is normalized, and the specific method is as follows:

[0031] First, the singular value decomposition matrix is obtained:

[0032] X = U∑V T (8)

[0033] wherein, T is a transpose symbol, X is the original data, the value dimension of X is 17968*21, U is an orthogonal matrix, the value dimension of U is 17968*21, ∑ is a diagonal matrix, the value dimension of ∑ is 21*21, V is an orthogonal matrix, the value dimension of V is 21*21, then the diagonal matrix is divided into 3 columns, and the dimension-reduced data is obtained by formula (9):

[0034] Y3 = U∑3 (9)

[0035] wherein, Y3 is the dimension-reduced data, the subscript is the set dimension, ∑3 is the selected three-column diagonal matrix, and finally the normalization operation shown in formula (6) is performed;

[0036] Step 2.2, the data after the normalization in step 2.1 is input into three convolutional networks respectively, the three convolutional networks have the same structure, the structure of each convolutional network includes three convolutional layers, three pooling layers and a first-level fully connected layer, the outputs of the three convolutional networks are merged into a second-level fully connected layer, the number of channels of the network is proportional to the dimension of the data, the output of each network is a probability value matrix of the fault category, the matrix is 5*1, each fault probability in the probability matrix ranges from 0 to 1, and the fault represented by the maximum probability in the probability matrix is the primary fault classification decision obtained by the probability matrix.

[0037] Step 3 is specifically as follows:

[0038] Step 3.1, each convolutional network in the step 2 will calculate a loss, and the weight influence factor of each convolutional network output to the second full connection layer is determined by the loss, as formula (10):

[0039]

[0040] Wherein, W a is the channel weight factor of the a-th convolutional network, loss a is the loss size of the a-th convolutional network output, c is a constant, and then the weight influence factor is normalized as formula (11):

[0041]

[0042] Wherein, w a is the channel weight of the a-th convolutional network, and b is the total number of network channels;

[0043] After obtaining the normalized weight, the matrix in step 2.2 is multiplied by the weight and input to the full connection layer, the probability of the fault category is obtained, and the loss of the convolutional network with the maximum weight is output;

[0044] Step 3.2, the probability matrix obtained in step 2.2 is multiplied by the weight obtained in step 3.1 and input to the second full connection layer, and a final fault probability matrix is output through the second full connection layer, and the matrix size is 5*1;

[0045] Step 3.3, the final fault probability matrix obtained in step 3.2 is used for decision analysis to obtain a specific fault category, and the bearing data set has five fault types, which are:

[0046] a, inner ring fault, b, rolling element fault, c, outer ring 3 o'clock position fault, d, outer ring 6 o'clock position fault, e, outer ring 12 o'clock position fault, the fault probability of each row in the probability matrix is 0 to 1, the maximum value in the final fault probability matrix is used for final decision analysis, and all are 0, indicating no fault.

[0047] The beneficial effects of the present application are that the fault diagnosis method based on data fusion and multi-channel convolutional network combines the multi-level fusion strategy of data level fusion, feature level fusion and decision level fusion, improves the utilization rate of original data, enhances the robustness of the system, effectively solves the problems of data redundancy and existence of pseudo data in multi-source homogeneous data, and the multi-channel convolutional network also effectively solves the problem of overfitting in the deep learning process, and improves the prediction accuracy. BRIEF DESCRIPTION OF DRAWINGS

[0048] Figure 1 is the overall flowchart of the present application;

[0049] Figure 2 is the schematic diagram of the multi-channel convolution network in the application. DETAILED DESCRIPTION

[0050] The application will be described in detail below in combination with the drawings and specific embodiments.

[0051] The application is based on a fault diagnosis method based on data fusion and a multi-channel convolution network, and a flowchart is as shown in Figure 1 The specific implementation is as follows:

[0052] Step 1, input the multi-source homogeneous sensor data of bearing acceleration, complete the missing data by the Kriging interpolation method, then calculate the contribution rate of each data, and fuse these data signals through the contribution rate;

[0053] Step 1 is specifically as follows:

[0054] Step 1.1, first, filter the input multi-source homogeneous sensor data of bearing acceleration, since in industrial production, most common noises are subject to normal distribution, so the obtained multi-source homogeneous data is first subjected to Gaussian filtering to obtain multi-source homogeneous data after Gaussian filtering;

[0055] Step 1.2, the Kriging interpolation method in the spatial interpolation method is used to complete the data of different sensors to obtain multiple groups of homogeneous data with the same dimension, since the sampling frequencies of different sensors are different, so the data lengths obtained are different, in order to make up for the missing values, the estimated value of the unknown data point is calculated by formula (1):

[0056]

[0057] Wherein, H(z o ) is the attribute estimation value of the unknown sample point z o , c is the mean value of all sample points, h(z i ) is the attribute value of the known sample point z i , N is the number of known sample points within 5 units of the unknown sample point, and i represents the i th data point.

[0058] Step 1.3, calculate the contribution rate of the completed multi-source data obtained in step 1.2, the specific method is as follows:

[0059] First, calculate the correlation between the data signals, as formula (2):

[0060]

[0061] Wherein, R i,jrepresents the correlation between the ith signal and the jth signal, x i (t0) represents the t0th data point in the signal i, t represents the selected data point, and n represents the number of data points in the signal;

[0062] Then, the signal correlation is quantitatively calculated as formula (3):

[0063]

[0064] where E i is the correlation degree of signal i, m is the total number of signals, j represents the jth signal, and the correlation degree is analyzed for importance as shown in formula (4):

[0065]

[0066] where y i is the information redundancy of signal i, the information amount of the signal is obtained by calculating the importance analysis, the mean and variance of the correlation degree after importance analysis are calculated, and the contribution rate of each signal is calculated using the mean and variance as shown in formula (5):

[0067]

[0068] where k i represents the contribution rate of the ith signal, y i is the importance analysis representation, μ is the mean of the importance analysis representation, and σ is the variance of the importance analysis representation. After the contribution rate is calculated, the weight coefficient of each signal in the fusion signal is obtained by normalizing the contribution rate, and the normalization method is shown in formula (6):

[0069]

[0070] where K i (t) is the contribution rate of data point t of signal i after normalization, and then each signal data point is multiplied by the weight in time sequence, as shown in formula (7):

[0071]

[0072] where x(h) is the signal value after fusion of data point h, and the corresponding point linear addition can obtain the signal after data-level fusion.

[0073] Step 2, combining Figure 2 , the fused data is normalized and input into a multi-channel convolution network to output a preliminary classification decision of the fault;

[0074] Step 2 is as follows:

[0075] Step 2.1, the fusion signal calculated in step 1.3 is first processed by dimension reduction, and the multi-dimensional signal after fusion is reduced to 3 dimensions by using principal component analysis, and then the signal is normalized, the specific method is as follows:

[0076] First, the singular value decomposition matrix is obtained:

[0077] X = U∑V T (8)

[0078] Wherein, T is the transpose symbol, X is the original data, the value dimension of X is 17968*21, U is an orthogonal matrix, the value dimension is 17968*21, ∑ is a diagonal matrix, the value dimension is 21*21, V is an orthogonal matrix, the value dimension is 21*21, then the diagonal matrix is divided into 3 columns, and the dimension-reduced data is obtained by formula (9):

[0079] Y3 = U∑3 (9)

[0080] Wherein, Y3 is the dimension-reduced data, the subscript is the set dimension, ∑3 is the selected three-column diagonal matrix, and finally the normalization operation shown in formula (6) is performed;

[0081] Step 2.2, the data after normalization in step 2.1 is input into three convolutional networks respectively, the three convolutional networks have the same structure, and the structure of each convolutional network includes three convolutional layers, three pooling layers and a first-level fully connected layer. The outputs of the three convolutional networks are merged into a second-level fully connected layer. The number of channels of the network is proportional to the dimension of the data. The output of each network is a probability value matrix of fault categories, and the matrix is 5*1. Each fault probability in the probability matrix ranges from 0 to 1. The fault represented by the maximum probability in the probability matrix is the primary fault classification decision obtained by the probability matrix.

[0082] Step 3, the output of the multi-channel convolutional network is weighted to obtain the final decision.

[0083] Step 3 is as follows:

[0084] Step 3.1, each convolutional network in step 2 will calculate a loss, and the weight influence factor of each convolutional network output to the second-level fully connected layer is determined by the loss, as shown in formula (10):

[0085]

[0086] Wherein, W a represents the channel weight factor of the a-th convolutional network, loss aLoss size of the a-th convolutional network output, c is a constant, generally set to a random number between 0-1, then the weight influence factor is normalized as formula (11):

[0087]

[0088] Where, w a is the channel weight of the a-th convolutional network, b is the total number of network channels;

[0089] After getting the normalized weight, multiply the matrix in step 2.2 with the weight and input it to the fully connected layer to get the probability of the fault class, and output the loss of the convolutional network with the largest weight;

[0090] Step 3.2, multiply the probability matrix obtained in step 2.2 with the weight obtained in step 3.1 and input it into the second fully connected layer, and output a final fault probability matrix through the second fully connected layer, with a size of 5*1;

[0091] Step 3.3, decision analysis is performed on the final fault probability matrix obtained in step 3.2 to get the specific fault class. There are 5 kinds of fault types in the bearing data set, which are:

[0092] a, inner ring fault, b, rolling element fault, c, outer ring 3 o'clock position fault, d, outer ring 6 o'clock position fault, e, outer ring 12 o'clock position fault, the probability in each row of the probability matrix is 0 to 1, the maximum value in the final fault probability matrix is used to represent the final decision analysis, and 0 means no fault.

[0093] The specific structure of the single-channel convolutional network is shown in Table 1. In the convolutional layer 1, the convolution kernel size is 64*2*1, 8 convolution kernels are set, the bias value is 8, and the convolution step is 2. In the pooling layer 1, the pooling method is set to up-sampling, and the sampling interval is 2. In the convolutional layer 2, the convolution kernel size is 3*8*1, 16 convolution kernels are set, the bias value is 16, and the convolution step is 2. In the pooling layer 2, the pooling method is set to up-sampling, and the sampling interval is 2. In the convolutional layer 3, the convolution kernel size is 3*16*1, 32 convolution kernels are set, the bias value is 32, and the convolution step is 2. In the pooling layer 3, the pooling method is set to up-sampling, and the sampling interval is 2. Finally, the result of the third pooling is input into the first fully connected layer, the number of neurons in the first fully connected layer is set to 400, the activation function is ReLU, and the probability integration output function is softmax.

[0094] Table 1 single convolutional network structure

[0095]

[0096] (Note: (Convolution kernel width, Kernel width, KW); (Convolution kernel height, Kernel height, KH); (Convolution kernel channel, Kernel channel, KC); (Number of kernels, KN); (Strides); (Bias, B)).

Claims

1. A method for fault diagnosis based on data fusion and multi-channel convolutional network, characterized in that, Specifically, the following steps are implemented: Step 1, input the multi-source homogeneous sensor data of bearing acceleration, complete the missing data by Kriging interpolation method, then calculate the contribution rate of each data, and fuse these data signals through the contribution rate; The step 1 is specifically as follows: Step 1.1, first, filter the input multi-source homogeneous sensor data of bearing acceleration, first, Gaussian filter the obtained multi-source homogeneous data to obtain the Gaussian filtered multi-source homogeneous data; Step 1.2, adopt Kriging interpolation method in spatial interpolation method to complete the data of different sensors, obtain a plurality of groups of homogeneous data with same dimension, and calculate the estimated value of unknown data point through formula (1): （1） in, unknown sample points The attribute estimates, The mean of all sample points. For known sample points The attribute value, N, is the number of known sample points within 5 units of the unknown sample point; Indicates the first One data point; Step 1.3, calculate the contribution rate through the completed multi-source data obtained in step 1.2, the specific method is as follows: First, calculate the correlation between data signals, as formula (2): （2） wherein, represents a correlation between the first signal and the second signal, represents a first data point in the signal , represents the selected data point, represents a number of data points in the signal. Then, quantitatively calculate the signal correlation as formula (3): （3） in, For signal Relevance The total number of signals. Indicates the first For each signal, the relevance is analyzed for importance, as shown in formula (4): （4） wherein, is the information redundancy of the signal The contribution rate of each signal is calculated using the mean and variance of the correlation representation after the importance analysis, as shown in equation (5): （5） wherein, represents the contribution rate of the importance analysis, importance analysis, importance analysis, the normalization mode is shown in equation (6): （6） wherein, is the normalized signal is the data point of the contribution rate, and then the data point of each signal is multiplied by the weight in the time sequence, as shown in equation (7): （7） wherein, is the data point The signal value after fusion, corresponding to the point linear addition, can obtain the signal after data level fusion; Step 2, normalize the fused data and input into the multi-channel convolution network to output the preliminary classification decision of fault; Step 2.1, the fused signal calculated in step 1.3 is first processed by dimension reduction, and the fused multi-dimensional signal is reduced to 3 dimensions by using principal component analysis, and then the signal is normalized; Step 2.2, the normalized data of step 2.1 is input into three convolutional networks respectively, the three convolutional networks have the same structure, each convolutional network includes three convolutional layers, three pooling layers and a first-level fully connected layer, the outputs of the three convolutional networks are merged into a second-level fully connected layer, the number of channels of the network is proportional to the dimension of the data, the output of each network is a probability value matrix of fault categories, the matrix is , each fault probability range in the probability matrix is 0 to 1, the fault represented by the maximum probability in the probability matrix is the primary fault classification decision obtained by the probability matrix; Step 3, the output of the multi-channel convolution network is weighted to obtain the final decision. 2.The data fusion and multi-channel convolution network based fault diagnosis method according to claim 1, characterized in that, The step 2.1 is specifically as follows: First, get the matrix through singular value decomposition: （8） wherein, is a transpose symbol, is original data, the value dimension is 17968 21, is an orthogonal matrix, and the value dimension is 17968 21, is a diagonal matrix, and the value dimension is 21 21, is an orthogonal matrix, and the value dimension is 21 21, and then the diagonal matrix is divided into 3 columns, and the reduced data is obtained through formula (9): （9） wherein, is the reduced dimension data, and the subscript is the set dimension, is the selected three-column diagonal matrix, and then the normalization operation shown in equation (6) is performed. 3.The data fusion and multi-channel convolution network based fault diagnosis method according to claim 2, characterized in that, The step 3 is specifically as follows: Step 3.1, each convolution network in the step 2 will calculate a loss, and the weight influence factor of each convolution network output to the second full connection layer is determined through the loss, as formula (10): （10） wherein, represents the channel weight factor of the th convolutional network, is the loss size of the output of the th convolutional network, is a constant, and then the weight influence factor is normalized as formula (11): （11） wherein, is the channel weight of the th convolutional network, is the total number of network channels. After obtaining the normalized weight, multiply the matrix in step 2.2 by the weight and input into the full connection layer to obtain the probability of fault category, and output the convolution network loss with the maximum weight; Step 3.2, the probability matrix obtained in step 2.2 is multiplied with the weight corresponding to step 3.1 and input into the second full connection layer, and a final fault probability matrix is output through the second full connection layer, and the matrix size is 5 1; Step 3.3, the final fault probability matrix obtained through step 3.2 is used for decision analysis to obtain the specific fault category, and the bearing data set has 5 fault types, which are: a, inner ring fault, b, rolling element fault, c, outer ring 3 o'clock position fault, d, outer ring 6 o'clock position fault, e, outer ring 12 o'clock position fault, the probability in each row of the probability matrix is 0 to 1, the maximum value in the final fault probability matrix is used to represent the final decision analysis, and 0 represents no fault.

Images