A gearbox intelligent fault diagnosis method based on multi-channel self-calibration convolutional neural network

Abstract

The present invention discloses a method for intelligent fault diagnosis of a gearbox based on a multi-channel self-calibrated convolutional neural network, comprising: increasing the dimension of Grami angle field data, converting one-dimensional vibration signals of multiple sensors into two-dimensional data, and converting them into grayscale images as input, establishing a data set, and dividing it into a training set and a test set. Constructing a self-calibrated convolutional neural network to extract data features. Setting a fusion layer to convert the output of the self-calibrated convolutional neural network into one-dimensional data and fuse feature information. Setting a fully connected layer to map distributed features to a sample labeling space. Constructing a Softmax feature classifier to classify the extracted features. Using a training set to let the network learn, using a test set to test the trained network, and realizing gearbox fault diagnosis. The self-calibrated convolutional neural network model proposed in the present invention combines an information fusion method, and can effectively diagnose a single fault of a gearbox under the same speed condition.

Description

Technical Field

[0001] The invention belongs to the technical field of intelligent fault diagnosis of gearbox vibration signals, and relates to an intelligent fault diagnosis method for gearboxes based on a multi-channel self-calibration convolutional neural network. Background Art

[0002] With the rapid development of China's information technology and industrial generation technology, modern mechanical industrial equipment has become larger, faster, more complex, smarter and more integrated. Its functions are constantly improving and its internal structure is becoming more sophisticated. Gearboxes are the most widely used important components in mechanical equipment, and their stable operation is crucial to the normal operation of the entire equipment. However, gearboxes are often in harsh working conditions such as variable speed and high load, which makes them very prone to damage or even failure, causing huge economic losses and casualties.

[0003] Classic fault diagnosis technologies include expert systems, model analysis, and signal processing. However, these methods have shortcomings such as poor versatility, complex modeling, weak robustness, and poor representation capabilities, which greatly limits their scope of application. In the era of big data development, intelligent fault diagnosis methods have become a hot topic and a difficult problem in the field of fault diagnosis due to their versatility, end-to-end, and powerful representation capabilities.

[0004] In fault diagnosis, the data obtained by the sensor is single, and the converted image data is also obtained by each sensor separately, so it is impossible to reflect all the characteristics of the fault from multiple angles. Data-level information fusion ensures the integrity of the original information. Therefore, data fusion of multiple sensors is of great significance for fault diagnosis.

[0005] Traditional convolutional neural networks cannot differentiate the features of each channel, which may cause the network to learn similar feature forms. The introduction of self-calibrated convolutional neural networks makes it easier for the network to learn better differentiated features and improve the attention and generalization capabilities of the neural network. Summary of the invention

[0006] The purpose of the present invention is to overcome the defects of the prior art and provide a gearbox intelligent fault diagnosis method based on a multi-channel self-calibration convolutional neural network. By combining sensor information fusion and a self-calibration convolutional neural network, an effective fault diagnosis can be performed for a single gearbox fault under the same speed condition. The present invention can realize data fusion of multi-sensor single faults, reflect all characteristics of the fault from multiple angles and ensure the integrity of the original information, effectively improving the attention and generalization ability of the neural network.

[0007] In order to solve the above technical problems, the present invention adopts the following technical solutions.

[0008] A gearbox intelligent fault diagnosis method based on a multi-channel self-calibration convolutional neural network, wherein a plurality of three-way contact acceleration sensors are fixed on the gearbox to respectively measure the vibration signals of the input shaft, intermediate shaft and output shaft of the motor at a constant speed in the y direction and the z direction; a Belgian LMS vibration and noise test system is used to connect with each sensor to obtain the value of the vibration signal measured by the sensor;

[0009] The method comprises the following steps:

[0010] Step 1: Perform Grami angle field data dimensionality enhancement on the original vibration signal of each sensor: scale the one-dimensional sequence data, convert it from the rectangular coordinate system to the polar coordinate system, and then identify the time correlation of different time points by considering the angles between different points; convert the two-dimensional matrix data into a grayscale image, establish a data set, and divide it into a training set and a test set;

[0011] Step 2: Establish a self-calibrated convolution module, build a self-calibrated convolutional neural network, and extract data features: divide the input feature F evenly into two parts {F 1 ,F 2}, and then each part is sent to a special path to focus on different types of feature information; in the first path, different convolution operations {Conv 1 ,Conv 2 ,Conv 3} 1 Perform self-calibration and obtain F 1 '; In the second path, the convolution operation Conv 4 Perform a simple convolution operation to get F' 2 , keep the original spatial feature information; after the Concate operation, {F 1 ', F' 2} connected together as output;

[0012] Step 3, setting a fusion layer to fuse the feature information extracted by the self-calibration convolution module according to a certain weight: converting the output of the self-calibration convolution module into one-dimensional data; the certain weight means that each weight value is a number between 0 and 1 and the sum of the weights is 1;

[0013] Step 4: Set up the fully connected layer: map the distributed features to the sample label space;

[0014] Step 5: Construct a Softmax feature classifier to classify the extracted features: map the input vector from the N-dimensional space to the category, and the result is given in the form of probability;

[0015] Step 6, train and test the network model: use the training set to let the network learn, and use the test set to test the trained network to achieve gearbox fault diagnosis.

[0016] Furthermore, in step 1, the original vibration signal of each sensor is subjected to Grami angle field data dimension enhancement, and the process includes:

[0017] Step 1.1 Scaling: Normalize the one-dimensional sequence data and map the one-dimensional data to [0,1] using the minmax normalization method;

[0018] Step 1.2 Coordinate conversion: Convert the scaled time series data to polar coordinates, that is, treat the value as the cosine value of the angle, and the converted angle range is [0, π / 2];

[0019] Step 1.3 Angle transformation: transform the converted data angle in polar coordinates and calculate the angle sum;

[0020] If the length of the original time series data is 2n, after the angle and transformation, a matrix of size [2n, 2n] is obtained; a piecewise aggregation approximation is used to segment the sequence, and then the subsequences in each segment are compressed into a single value by averaging to reduce parameters; 2×2 is selected as a matrix block to convert the 2n×2n matrix into an n×n matrix.

[0021] Furthermore, in step 2, a self-calibration convolution module is established, a self-calibration convolutional neural network is constructed, and data features are extracted, and the process includes:

[0022] (1) Input the original feature map F 1 ∈R H×W×C / 2 Perform a downsampling operation, and the feature map size changes from H×W×C / 2 to H / 2×W / 2×C / 2. The operation is defined as Down(*). H, W, and C represent the height, width, and number of channels of the image input to the neural network, respectively.

[0023] (2) Convolution is performed on the downsampled feature map, which is denoted as f 1 (*);

[0024] (3) Perform an upsampling operation on the feature map after convolution in (2), and restore the size of the feature map from H / 2×W / 2×C / 2 to H×W×C / 2. This operation is denoted as Up(*).

[0025] (4) Input the original feature map F 1 ∈R H×W×C / 2Added to the feature map in (3), after passing through the activation function Sigmoid, a feature map of scale H×W×C / 2 is obtained. This operation is defined as σ(*). The activation function Sigmoid calculation formula can be expressed as:

[0026]

[0027] In the formula, e is a natural constant, and its value is approximately 2.718;

[0028] (5) For the original feature map F 1 ∈R H×W×C / 2 Perform a convolution operation, which is denoted as f 2 (*);

[0029] (6) Multiply the feature map obtained in (4) by the feature map obtained in (5). This operation is recorded as

[0030] (7) Perform a convolution operation on the feature map obtained in (6) to obtain the output F 1 '∈R H×W×C / 2 , this operation is recorded as f 3 (*); The output of the self-calibrated convolution operation can be expressed as

[0031]

[0032] Furthermore, in step 3, a fusion layer is set to fuse the feature information extracted by the self-calibration convolution module according to a certain weight, and the process includes:

[0033] The output of the self-calibrated convolutional neural network is three-dimensional data. The function of the fusion layer is to weight the features extracted from different channels to obtain the feature information extracted from multiple sensor signals.

[0034] The 6 groups of vibration signals measured by the sensor are converted into grayscale images and processed by the self-calibrated convolutional neural network to obtain 6 new grayscale images; the value of each pixel at the same position on the grayscale image is x 1 、x 2 、x 3 、x 4 、x 5 、x 6 , the fusion weights are set to w 1 、w 2 、w 3 、w 4 、w 5 、w 6 , then the pixel value at the corresponding position on the grayscale image after information fusion is x 1 w 1 +x 2 w 2+x 3 w 3 +x 4 w 4 +x 5 w 5 +x 6 w 6 .

[0035] Furthermore, in step 5, a Softmax feature classifier is constructed to classify the extracted features, and the process includes:

[0036] When the self-calibrated convolutional neural network classifies an image, the input image is passed to the Softmax feature classifier after the neural network feature extraction. The parameter matrix θ is obtained after the Softmax feature classifier is trained. θ is multiplied by the image feature column vector to output the probability value of the image belonging to each category. Among them, the category corresponding to the maximum value is the determined category of the image.

[0037] The Softmax feature classifier maps the input vector from the N-dimensional space to the category, and the result is given in the form of probability. The formula is as follows:

[0038]

[0039] In the formula, is the weight, the classifier parameter corresponding to the category, and the model parameter θ is as follows:

[0040]

[0041] θ is obtained by training the Softmax classifier, and the category to which the item belongs is determined by calculating the probability of all possible categories of the item to be classified; given a data set consisting of n training samples: {(x (1) ,y (1) ),(x (2) ,y (2) ),...,(x (n) ,y (n) )},x (i) Represents the input vector, y (i) For each x (i) The category label of ; for a given test sample x (i) , the Softmax classifier estimates the probability of belonging to each category, which is calculated as follows:

[0042]

[0043] In the formula, h θ (x (i) ) is a vector, the element p(y (i) =k|x (i) ; θ) represents x(i) The probability of belonging to category k, h θ (x (i) ) the sum of the elements is equal to 1; for x (i) , select the k corresponding to the maximum probability value as the classification result of the current image; the value of the parameter θ can be obtained by minimizing the cost function, which is defined as:

[0044]

[0045] Where {·} is an indicative function, which is equal to 1 if true and 0 if false. By minimizing J(θ), the classifier parameter θ can be obtained.

[0046] Furthermore, the process of training and testing the network model in step 6 includes:

[0047] Using the vibration signal measured by the acceleration sensor of the motor at a constant speed, the original signal is dimensionalized by Grami angle field data to become a two-dimensional grayscale image; the image is divided into a training set and a test set at a certain ratio, the training set is input into the network model, and the network is trained after setting the training parameters, iterated 30 times, and the accuracy value and loss rate of each iteration are plotted into an accuracy and loss rate curve;

[0048] Then input the test set into the trained network for testing. After 30 tests, the accuracy results of each test are plotted into a test accuracy graph.

[0049] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0050] 1. The present invention adopts the Grami angular field algorithm to convert the one-dimensional vibration signal data of different sensors into two-dimensional matrix data, and then converts it into a grayscale image, establishes a data set, and divides it into a training set and a test set; after the one-dimensional sequence data is converted into two-dimensional data by the Grami angular field algorithm, the original signal timing characteristics are retained and the fault characteristics are enhanced.

[0051] 2. The present invention uses a self-calibration convolution module to extract image feature information. This structure can enlarge the receptive field of the self-calibration convolution. Each channel of the convolutional neural network learns different features differently, better extracts multi-scale information, and improves the attention and generalization ability of the neural network. Using a self-calibration convolutional neural network instead of traditional fault diagnosis methods reduces the use of labor while improving the accuracy of fault diagnosis, providing a new theoretical method and technical implementation means for state monitoring and fault diagnosis of modern complex large systems.

[0052] 3. The present invention uses multiple sensors to collect vibration signals of the gearbox. Compared with the signals collected by a single sensor, the fault information reflected by the signals collected by multiple sensors is more comprehensive and specific. These fault information can reflect the characteristics of the fault from multiple angles, reducing the probability of fault diagnosis errors caused by incomplete signals collected by a single sensor. Inputting these fault information into the self-calibration convolutional neural network for training and testing can effectively perform fault diagnosis. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Figure 1 It is a schematic diagram of sensor arrangement and information acquisition according to an embodiment of the present invention.

[0054] Figure 2 It is a flow chart of a gearbox fault diagnosis method based on a multi-channel self-calibration convolutional neural network of the present invention.

[0055] Figure 3 It is a schematic diagram of the Grami angle field dimension enhancement principle of an embodiment of the present invention.

[0056] Figure 4 It is a schematic diagram of the structure of a self-calibrated convolutional neural network according to an embodiment of the present invention.

[0057] Figure 5 It is a schematic diagram of information fusion principle of an embodiment of the present invention.

[0058] Figure 6 It is a training accuracy curve diagram of an embodiment of the present invention.

[0059] Figure 7 It is a training loss rate curve diagram of an embodiment of the present invention.

[0060] Figure 8 1 is a test accuracy diagram of an embodiment of the present invention. DETAILED DESCRIPTION

[0061] The present invention provides a gearbox intelligent fault diagnosis method based on a multi-channel self-calibrated convolutional neural network, which combines information fusion and a self-calibrated convolutional neural network to effectively diagnose a single fault of a gearbox under the same speed condition. The method includes: converting a one-dimensional vibration signal measured by multiple sensors into two-dimensional data through Grami angle field data dimension increase, converting the two-dimensional data into a grayscale image as input, establishing a data set, and dividing it into a training set and a test set. Establishing a self-calibrated convolution module, constructing a self-calibrated convolutional neural network, and extracting data features. Setting a fusion layer, converting the output of the self-calibrated convolutional neural network into one-dimensional data, and fusing feature information. Setting a fully connected layer, mapping distributed features to a sample labeling space. Constructing a Softmax feature classifier to classify the extracted features. Using a training set to let the network learn, using a test set to test the trained network, and realizing gearbox fault diagnosis. The self-calibrated convolutional neural network model proposed by the present invention combines an information fusion method, and can effectively diagnose a single fault of a gearbox under the same speed condition.

[0062] The present invention is further described in detail below with reference to the accompanying drawings.

[0063] like Figure 1 As shown, the gearbox in the present invention is a bevel gearbox, including an input shaft, an intermediate shaft, an output shaft, four planetary gears and a number of bearings. The present invention uses a plurality of three-way contact acceleration sensors, which are fixed on the gearbox through a magnetic base, and respectively measure the vibration signals of the input shaft, intermediate shaft and output shaft of the motor at a certain constant speed. Each sensor can measure the signals in the y direction and the z direction of the corresponding axis. The present invention uses the Belgian LMS vibration and noise test system to collect signals, connects it with each sensor, and can obtain the value of the signal measured by the sensor.

[0064] like Figure 2 As shown, a gearbox intelligent fault diagnosis method based on a multi-channel self-calibration convolutional neural network of the present invention comprises the following steps:

[0065] Step 1: Perform Grami angular field data dimensionality enhancement on the original vibration signal of each sensor: scale the one-dimensional sequence data and convert it from the rectangular coordinate system to the polar coordinate system. Then, consider the angles between different points to identify the time correlation at different time points. Convert the two-dimensional matrix data into a grayscale image, establish a data set, and divide it into a training set and a test set.

[0066] Step 2: Establish a self-calibrated convolution module, build a self-calibrated convolutional neural network, and extract data features: divide the input feature F evenly into two parts {F 1 ,F 2}, and then each part is sent to a special path to focus on different types of feature information. In the first path, different convolution operations {Conv 1 ,Conv 2 ,Conv 3} 1 Perform self-calibration and obtain F 1 '. In the second path, the convolution operation Conv 4 Perform a simple convolution operation to get F' 2 , keeping the original spatial feature information. After the Concate operation, {F 1 ', F' 2} are connected together as output.

[0067] Step 3: Set up a fusion layer to fuse the feature information extracted by the self-calibration convolution module according to a certain weight. The output of the self-calibration convolution module is converted into one-dimensional data, where the certain weight means that each weight value is a number between 0 and 1 and the sum of the weights is 1.

[0068] Step 4: Set up a fully connected layer to map the distributed features to the sample label space.

[0069] Step 5: Build a Softmax feature classifier to classify the extracted features, map the input vector from the N-dimensional space to the category, and give the result in the form of probability.

[0070] Step 6, train and test the network model: use the training set to let the network learn, and use the test set to test the trained network to achieve gearbox fault diagnosis.

[0071] In step 1, the original vibration signal of each sensor is subjected to Grami angle field data dimension enhancement, and the process includes:

[0072] The Grami angle field method is used to convert one-dimensional data into two-dimensional data. The principle is as follows Figure 3 As shown. It scales the one-dimensional sequence data, converts it from the rectangular coordinate system to the polar coordinate system, and identifies the time correlation of different time points by considering the angles between different points. The implementation steps are as follows:

[0073] Step 1.1 Scaling

[0074] Normalize the one-dimensional sequence data and map the one-dimensional data to [0,1] using the minmax normalization method.

[0075] Step 1.2 Coordinate transformation

[0076] Convert the scaled time series data to polar coordinates, that is, regard the value as the cosine value of the angle, and the converted angle range is [0, π / 2].

[0077] Step 1.3 Angle transformation

[0078] Transform the transformed data angles in polar coordinates and calculate the angle sum.

[0079] If the length of the original time series data is 2n, after the angle and transformation, a matrix of size [2n, 2n] is obtained. A piecewise aggregation approximation is used to segment the sequence, and then the subsequences in each segment are compressed into a single value by averaging to reduce the parameters. A 2×2 matrix block is selected to convert the 2n×2n matrix into an n×n matrix.

[0080] In step 2, a self-calibration convolution module is established, a self-calibration convolutional neural network is constructed, and data features are extracted. The process includes:

[0081] The advantage of convolutional neural networks is that they extract the spatial association of input data, while the spatial relationship of a simple one-dimensional vector is not obvious. Converting one-dimensional vibration signal sequence data into two-dimensional data can achieve better diagnostic results in intelligent fault diagnosis.

[0082] Convolution operation is a common operation to extract feature information in neural networks, but the uniform convolution kernel size makes it difficult to extract feature information of different scales in one convolution operation. The self-calibration convolution effectively solves this problem. The network structure diagram is shown in the figure below. Figure 4 shown.

[0083] The self-calibration convolution operation steps are as follows:

[0084] (1) Input the original feature map F 1 ∈R H×W×C / 2 After downsampling, the feature map size changes from H×W×C / 2 to H / 2×W / 2×C / 2. The operation is defined as Down(*). H, W, and C represent the height, width, and number of channels of the image input to the neural network, respectively.

[0085] (2) Convolution is performed on the downsampled feature map, which is denoted as f 1 (*).

[0086] (3) An upsampling operation is performed on the feature map after convolution in (2), and the size of the feature map is restored from H / 2×W / 2×C / 2 to H×W×C / 2. This operation is denoted as Up(*).

[0087] (4) Input the original feature map F 1 ∈R H×W×C / 2Added to the feature map in (3), after passing through the activation function Sigmoid, the calculation formula obtains a feature map of scale H×W×C / 2. This operation is defined as σ(*). The activation function Sigmoid calculation formula can be expressed as:

[0088]

[0089] In the formula, e is a natural constant, and its value is approximately 2.718.

[0090] (5) For the original feature map F 1 ∈R H×W×C / 2 Perform a convolution operation, which is denoted as f 2 (*).

[0091] (6) Multiply the feature map obtained in (4) by the feature map obtained in (5). This operation is recorded as

[0092] (7) Perform a convolution operation on the feature map obtained in (6) to obtain the output F 1 '∈R H×W×C / 2 , this operation is recorded as f 3 (*). The output of the self-calibrated convolution operation can be expressed as

[0093]

[0094] In step 3, a fusion layer is set to fuse the feature information extracted by the self-calibration convolution module according to a certain weight, and the process includes:

[0095] The output of the self-calibrated convolutional neural network is three-dimensional data. The function of the fusion layer is to weight the features extracted from different channels to obtain the feature information extracted from multiple sensor signals. The information fusion principle diagram is shown in the figure. Figure 5 shown.

[0096] After the six groups of vibration signals measured by the sensor are converted into grayscale images and processed by the self-calibrated convolutional neural network, six new grayscale images are obtained. The value of each pixel at the same position on the grayscale image is x 1 、x 2 、x 3 、x 4 、x 5 、x 6 , the fusion weights are set to w 1 、w 2 、w 3 、w 4 、w 5 、w 6 , then the pixel value at the corresponding position on the grayscale image after information fusion is x 1 w 1 +x2 w 2 +x 3 w 3 +x 4 w 4 +x 5 w 5 +x 6 w 6 .

[0097] The step 4 sets the fully connected layer, including:

[0098] The fully connected layer maps the distributed features to the sample label space, which essentially linearly transforms one feature space into another feature space.

[0099] The step 5 is to construct a Softmax feature classifier to classify the extracted features, and the process includes:

[0100] The principle of Softmax classifier is probability calculation. When the self-calibrated convolutional neural network classifies an image, the input image is passed into the classifier after the neural network feature extraction. After training the Softmax classifier, the parameter matrix θ can be obtained. θ is multiplied by the image feature column vector to output the probability value of the image belonging to each category. Among them, the category corresponding to the maximum value is the determined category of the image.

[0101] Softmax regression is an extension of logistic regression, which provides more possibilities for category labels and is suitable for multi-classification problems. The Softmax classifier maps the input vector from the N-dimensional space to the category, and the result is given in the form of probability. The formula is as follows:

[0102]

[0103] In the formula, is the weight, the classifier parameter corresponding to the category, and the model parameter θ is as follows:

[0104]

[0105] θ is obtained by training the Softmax classifier, which determines the category to which the item belongs by calculating the probability of all possible categories of the item to be classified. Given a data set consisting of n training samples: {(x (1) ,y (1) ),(x (2) ,y (2) ),...,(x (n) ,y (n) )},x (i) Represents the input vector, y (i) For each x (i) For a given test sample x (i), the Softmax classifier estimates the probability of belonging to each category, and its calculation principle is as follows:

[0106]

[0107] In the formula, h θ (x (i) ) is a vector, the element p(y (i) =k|x (i) ; θ) represents x (i) The probability of belonging to category k, h θ (x (i) ) the sum of the elements is equal to 1. (i) , select the k corresponding to the maximum probability value as the classification result of the current image. The value of the parameter θ can be obtained by minimizing the cost function, which is defined as:

[0108]

[0109] Where {·} is an indicative function, which is equal to 1 if true and 0 if false. By minimizing J(θ), the classifier parameter θ can be obtained.

[0110] The process of step 6, training and testing the network model, includes:

[0111] Use the training set to let the network learn, and use the test set to test the trained network to achieve fault diagnosis.

[0112] Vibration sensor layout scheme Figure 1 As shown. The vibration signals of the input shaft, intermediate shaft and output shaft in the y-axis and z-axis directions are collected respectively. The data used includes 9 types of fault data and normal data. The 9 types of fault data are the crack width of the inner and outer rings of the bearings with a width of 1.2mm and a crack depth of 2mm, the crack width of the inner and outer rings of the bearings with a width of 0.6mm and a crack depth of 2mm, 2 broken teeth faults of bevel gears, pitting faults of 4 pits and 6 pits, 0.5 times tooth thickness crack faults and 0.8 times tooth thickness wear faults.

[0113] Using the vibration signal measured by the acceleration sensor of the motor at a constant speed, the original signal is multiplied by the Grami angle field data to become a two-dimensional grayscale image. After dividing the image into a training set and a test set at a certain ratio, the training set is input into the network model, and the network is trained after setting the training parameters. It is iterated 30 times, and the accuracy value and loss rate of each iteration are plotted into a graph to obtain the following: Figure 6 and Figure 7 Accuracy and loss rate graphs are shown.

[0114] Then input the test set into the trained network for testing. After 30 tests, the accuracy results of each test are plotted in the figure, as shown below: Figure 8 The test accuracy graph shown.

Claims

1. A gearbox intelligent fault diagnosis method based on a multi-channel self-calibrated convolutional neural network, characterized in that: A plurality of three-way contact acceleration sensors are fixed on the gearbox to measure the vibration signals of the input shaft, intermediate shaft and output shaft of the motor in the y direction and the z direction at a constant speed respectively; a vibration and noise test system is connected to each sensor to obtain the value of the vibration signal measured by the sensor; The method comprises the following steps: Step 1: Perform Grami angle field data dimensionality enhancement on the original vibration signal of each sensor: scale the one-dimensional sequence data, convert it from the rectangular coordinate system to the polar coordinate system, and then consider the angles and between different points to identify the time correlation at different time points; convert the two-dimensional matrix data into a grayscale image, establish a data set, and divide it into a training set and a test set; Step 2: Establish a self-calibrated convolution module, build a self-calibrated convolutional neural network, and extract data features: evenly divide the input feature F into two parts {F1, F2}, and then send each part to a special path to focus on different types of feature information; in the first path, use different convolution operations {Conv1, Conv2, Conv3} to perform self-calibration operations on F1 to obtain F1'; in the second path, use the convolution operation Conv4 to perform a simple convolution operation to obtain F2', keeping the original spatial feature information; connect {F1', F2'} together through the Concate connection operation as the output; Step 3, setting a fusion layer to fuse the feature information extracted by the self-calibration convolution module according to a certain weight: converting the output of the self-calibration convolution module into one-dimensional data; the certain weight means that each weight value is a number between 0 and 1 and the sum of the weights is 1; Step 4: Set up the fully connected layer: map the distributed features to the sample label space; Step 5: Construct a Softmax feature classifier to classify the extracted features: map the input vector from the N-dimensional space to the category, and the result is given in the form of probability; Step 6, train and test the network model: use the training set to let the network learn, and use the test set to test the trained network to achieve gearbox fault diagnosis.

2. The gearbox intelligent fault diagnosis method based on a multi-channel self-calibrated convolutional neural network according to claim 1 is characterized in that: In step 1, the original vibration signal of each sensor is subjected to Grami angle field data dimension enhancement, and the process includes: Step 1.1 Scaling: Normalize the one-dimensional sequence data and map the one-dimensional data to [0,1] using the minmax normalization method; Step 1.2 Coordinate conversion: Convert the scaled time series data to polar coordinates, that is, treat the value as the cosine value of the angle, and the converted angle range is [0, π / 2]; Step 1.3 Angle transformation: transform the converted data angle in polar coordinates and calculate the angle sum; If the length of the original time series data is 2n, after the angle and transformation, a matrix of size [2n, 2n] is obtained; a piecewise aggregation approximation is used to segment the sequence, and then the subsequences in each segment are compressed into a single value by averaging to reduce parameters; 2×2 is selected as a matrix block to convert the 2n×2n matrix into an n×n matrix.

3. The gearbox intelligent fault diagnosis method based on a multi-channel self-calibrated convolutional neural network according to claim 1 is characterized in that: In step 2, a self-calibration convolution module is established, a self-calibration convolutional neural network is constructed, and data features are extracted. The process includes: (1) Input the original feature map F1∈R H×W×C / 2 Perform a downsampling, and the feature map size changes from H×W×C / 2 to H / 2×W / 2×C / 2. The operation is defined as Down(*). H, W, and C represent the height, width, and number of channels of the image input to the neural network, respectively. (2) Convolution is performed on the downsampled feature map, and this operation is recorded as f1(*); (3) Perform an upsampling operation on the feature map after convolution in (2), and restore the size of the feature map from H / 2×W / 2×C / 2 to H×W×C / 2. This operation is denoted as Up(*). (4) Input the original feature map F1∈R H×W×C / 2 Added to the feature map in (3), after passing through the activation function Sigmoid, a feature map of scale H×W×C / 2 is obtained. This operation is defined as σ(*), and the activation function Sigmoid calculation formula is expressed as: In the formula, e is a natural constant, and its value is 2.718; (5) For the original feature map F1∈R H×W×C / 2 Perform a convolution operation, which is denoted as f2(*); (6) Multiply the feature map obtained in (4) by the feature map obtained in (5). This operation is recorded as (7) Perform a convolution operation on the feature map obtained in (6) to obtain the output F1'∈R H×W×C / 2 , the operation is recorded as f3(*); the output of the self-calibrated convolution operation can be expressed as 4. The gearbox intelligent fault diagnosis method based on a multi-channel self-calibrated convolutional neural network according to claim 1 is characterized in that: In step 3, a fusion layer is set to fuse the feature information extracted by the self-calibration convolution module according to a certain weight, and the process includes: The output of the self-calibrated convolutional neural network is three-dimensional data. The function of the fusion layer is to weight the features extracted from different channels to obtain the feature information extracted from multiple sensor signals. The six groups of vibration signals measured by the sensor are converted into grayscale images and processed by the self-calibrated convolutional neural network to obtain six new grayscale images; the values ​​of each pixel at the same position on the grayscale image are x1, x2, x3, x4, x5, and x6, and the fusion weights are set to w1, w2, w3, w4, w5, and w6 respectively. The values ​​of the pixels at the corresponding positions on the grayscale image after information fusion are x1w1+x2w2+x3w3+x4w4+x5w5+x6w6.

5. The gearbox intelligent fault diagnosis method based on a multi-channel self-calibrated convolutional neural network according to claim 1 is characterized in that: The step 5 is to construct a Softmax feature classifier to classify the extracted features, and the process includes: When the self-calibrated convolutional neural network classifies an image, the input image is passed to the Softmax feature classifier after the neural network feature extraction. The parameter matrix θ is obtained after the Softmax feature classifier is trained. θ is multiplied by the image feature column vector to output the probability value of the image belonging to each category. Among them, the category corresponding to the maximum value is the determined category of the image. The Softmax feature classifier maps the input vector from the N-dimensional space to the category, and the result is given in the form of probability. The formula is as follows: In the formula, is the weight, the classifier parameter corresponding to the category, and the model parameter θ is as follows: θ is obtained by training the Softmax classifier, and the category to which the item belongs is determined by calculating the probability of all possible categories of the item to be classified; given a data set consisting of n training samples: {(x (1) ,y (1) ),(x (2) ,y (2) ),...,(x (n) ,y (n) )},x (i) Represents the input vector, y (i) For each x (i) The category label of ; for a given test sample x (i) , the Softmax classifier estimates the probability of belonging to each category, which is calculated as follows: In the formula, h θ (x (i) ) is a vector, the element p(y (i) =k|x (i) ; θ) represents x (i) The probability of belonging to category k, h θ (x (i) ) the sum of the elements is equal to 1; for x (i) , select the k corresponding to the maximum probability value as the classification result of the current image; the value of the parameter θ can be obtained by minimizing the cost function, which is defined as: In the formula, {y (i) =j} is an indicative function, which is equal to 1 if true and 0 if false; minimizing J(θ) can yield the classifier parameter θ.

6. The gearbox intelligent fault diagnosis method based on a multi-channel self-calibrated convolutional neural network according to claim 1 is characterized in that: The process of step 6, training and testing the network model, includes: Using the vibration signal measured by the acceleration sensor of the motor at a constant speed, the original signal is dimensionalized by Grami angle field data to become a two-dimensional grayscale image; the image is divided into a training set and a test set at a certain ratio, the training set is input into the network model, and the network is trained after setting the training parameters, iterated 30 times, and the accuracy value and loss rate of each iteration are plotted into an accuracy and loss rate curve; Then input the test set into the trained network for testing. After 30 tests, the accuracy results of each test are plotted into a test accuracy graph.

Images