CN114383844B - A Fault Diagnosis Method for Rolling Bearings Based on Distributed Deep Neural Network - Google Patents

Abstract

The invention belongs to the technical field of mechanical intelligent manufacturing, in particular to a rolling bearing fault diagnosis method based on a distributed deep neural network. S1: The production of the dataset. S2: Construct a distributed deep neural network model. S3: Joint training of branch model and backbone model. S4: Model inference. S5: Fault identification, summarizing the reasoning results and traffic of the branch model and the main model, and outputting the final fault diagnosis accuracy and the communication traffic consumed. The model constructed by the method of the invention can automatically extract features from the original vibration signal without manually selecting features and denoising, and can realize high-precision, low-communication, and low-time-delay rolling bearing fault diagnosis under variable load and large noise conditions.

Description

A Fault Diagnosis Method for Rolling Bearings Based on Distributed Deep Neural Network

technical field

The invention belongs to the technical field of mechanical intelligent manufacturing, in particular to a rolling bearing fault diagnosis method based on a distributed deep neural network.

Background technique

As an important part of rotating machinery and equipment, rolling bearings are widely used in various industries of the national economy. In the actual work process, the vibration signal of bearing faults changes continuously with the load, and the vibration signals of bearing faults are weak and easily covered by strong interference signals. Therefore, extracting fault features from bearing vibration signals in variable load and strong noise environments, and timely and effectively diagnosing faults, can effectively avoid continuous deterioration of faults. In recent years, data-driven fault diagnosis methods have gradually become a fault diagnosis method. important research hotspots in the field. Data-driven fault diagnosis methods mainly include fault diagnosis methods based on signal analysis, machine learning, and deep learning.

Bearing fault diagnosis based on signal analysis mainly includes: Fourier analysis method, wavelet transform method, cepstrum method, empirical mode decomposition method, etc. These fault diagnosis methods can only effectively identify a certain type of specific working conditions, have low generalization ability and are easily affected by other factors.

Bearing fault diagnosis based on machine learning mainly includes: support vector machine, extreme learning machine, artificial neural network, etc. Although these fault diagnosis methods can effectively identify multiple types of faults, they need to extract certain features of the fault signal. This process is very dependent on manual experience, and there are problems of complex and time-consuming calculations, and the feature extraction process cannot fully characterize all faults. type, there are certain limitations.

Bearing fault diagnosis based on deep learning mainly includes: convolutional neural network, deep residual network, deep Boltzmann machine, deep belief network, stacked autoencoder, etc. Although these fault diagnosis methods realize the automatic extraction of fault features and the identification of multiple types of faults, with the massive data generated by sensors and the deepening of the network depth of deep learning models, the deep learning models of centralized cloud computing have diagnostic delays and problems. The problem of a significant increase in communication costs.

Therefore, how to effectively diagnose bearing faults is an urgent problem to be solved.

Contents of the invention

In order to solve the above problems, the present invention provides a rolling bearing fault diagnosis method based on a distributed deep neural network.

The present invention adopts the following technical solutions: a rolling bearing fault diagnosis method based on a distributed deep neural network, comprising the following steps.

S1: The production of the data set, firstly, the acceleration sensor above the bearing seat is used to collect the noise-free vibration acceleration signal of the faulty bearing, and add additive Gaussian white noise to it, and then intercept the vibration signal according to a certain data point length, and It is converted into a two-dimensional image and given real fault labels to construct a data set, and finally the data set is divided into a training set and a test set according to a preset ratio.

S2: Construct a distributed deep neural network model. The model framework includes: a sample input point, a shared convolution block, a branch model, a backbone model, and two sample diagnosis result output points.

S3: Joint training of the branch model and the backbone model. During training, the weighted summation of the loss value of the cross-entropy loss function of each sample diagnosis result output point makes the entire network can be jointly trained and each sample diagnosis result output point is relatively Because of its depth, it can reach the ideal corresponding loss value, so that the predicted value is infinitely close to the real value.

S4: Model reasoning, the branch model performs rapid initial feature extraction on the input test set sample image, and when the output point of the sample diagnosis result of the branch model is confident in the prediction result, the branch exit point exits the sample and outputs the classification result, Otherwise, the image features of the non-exited samples in the shared convolution block are passed to the backbone model for feature extraction and classification in the next step.

S5: Fault identification, summarizing the reasoning results and traffic of the branch model and the main model, and outputting the final fault diagnosis accuracy and the communication traffic consumed.

The step of adding additive Gaussian white noise to each load vibration signal based on noise-free in step S1 is:

S11: Simulate different noise conditions by adjusting the signal-to-noise ratio. The decibel form of the signal-to-noise ratio is expressed as:

为信号功率，/>

为噪声功率，/>

为振动信号数值，N为振动信号长度；In the formula:

is the signal power, />

is the noise power, />

is the value of the vibration signal, N is the length of the vibration signal;

表示，因此对于标准正态分布噪声，其功率为1，因此，首先计算原始信号/>

的功率，然后计算在期望信噪比下产生的噪声信号/>

的功率，最后，通过下面的公式生成加性高斯白噪声，然后将其加到原始信号/>

中，使原始信号具有期望的信噪比；原始振动信号加入加性高斯白噪声的振动信号计算公式为：S12: For a signal with zero mean and known variance, its power can be calculated by variance

Represents, so for standard normally distributed noise, its power is 1, therefore, first calculate the original signal />

power, and then calculate the noise signal generated under the desired signal-to-noise ratio />

The power of , finally, the additive white Gaussian noise is generated by the following formula, and then added to the original signal />

In , the original signal has the desired signal-to-noise ratio; the formula for calculating the vibration signal by adding additive Gaussian white noise to the original vibration signal is:

X = M + x _i

In the formula: M is Gaussian white noise, and randn represents a function that generates a standard normal distribution of random numbers or matrices.

In step S1, the vibration signal is intercepted according to a certain data point length and converted into a two-dimensional image, and the calculation formula of the conversion process is expressed as:

，K表示二维图像的单边尺寸。In the formula: P represents the pixel intensity of the two-dimensional image, L represents the value of the original vibration signal added to the vibration signal of additive Gaussian white noise,

, K represents the single side size of the two-dimensional image.

In step S2, the key structure of the distributed deep neural network includes:

1) The 3×3 convolution kernel stacking method is used to extract the feature vector of the input sample image to improve the receptive field and nonlinear expression ability of the convolution layer;

2) The number of shared convolution kernels is set to 1, which is used to reduce the amount of communication between the shared convolution block and the adjacent convolution blocks of the backbone and to preserve the original image features to the greatest extent;

3) Place the convolutional feature bag behind the last convolutional block of the branch instead of the fully connected layer to quantify the final output feature vector of the convolutional block and reduce the amount of model parameters;

4) The backbone uses a residual network block and a global average pooling structure to omit the fully connected layer, increase the mobility of the feature vector, and reduce the amount of model parameters.

The joint training process of the model in step S3 is,

S31: The branch model and the backbone model have a classifier respectively, and the output point of each sample diagnosis result takes the cross-entropy loss function as the optimization target, and the cross-entropy loss function is expressed as:

表示样本的预测故障标签，C表示标签集合，/>

表示的是样本从神经网络的输入到第n个出口进行的运算，/>

表示该过程网络的权重和偏置等参数；In the formula: X represents the input sample, y represents the real fault label of the sample,

Indicates the predicted failure label of the sample, C indicates the label set, />

Indicates the operation of the sample from the input of the neural network to the nth exit, />

Represents parameters such as weights and biases of the process network;

S32: Train the weighted sum of the losses of the output points of each sample diagnosis result, and use the SGD method to update the parameters of the distributed neural network. The loss function of the distributed neural network is expressed as:

表示每个出口的权重，/>

表示第n个出口的估计值。In the formula: N represents the number of classified exports,

Indicates the weight of each outlet, />

represents the estimated value of the nth exit.

In step S4, the sample confidence is used as the basis for judging whether the branch exit point has confidence in the prediction result. If the sample confidence is less than a given threshold, it is confident, otherwise, there is no confidence. The formula for calculating the sample confidence is:

。In the formula: C is the set of all real labels, x is the probability vector,

.

In step S5, the communication volume calculation formula for module fault identification is:

In the formula: l is the percentage of branch exit samples to all input samples, f is the feature image size output from the shared convolution block to the backbone model, o is the number of feature image channels output from the shared convolution block to the backbone model, and the constant 4 means Yes, on a 64-bit normal Windows system, a 32-bit floating point number occupies 4 bytes.

Compared with the prior art, the present invention has the following beneficial effects:

The present invention converts the vibration signal of the rolling bearing into an image by graying data to construct an image sample; it proposes to add a branch model at the bottom layer of the model, withdraw some simple samples in advance, maximize the use of computer computing resources of the backbone model, and replace the convolution feature bag The fully connected layer of the branch model improves the defect of insufficient computer computing power of the branch model; through the collaborative calculation of the branch and the main model, different fault types and fault severity of the rolling bearing can be identified; the model constructed by the method of the present invention can be automatically obtained from the original vibration signal Extracting features does not require manual feature selection and denoising. Under variable load and large noise conditions, it can achieve high-precision, low-communication, and low-latency rolling bearing fault diagnosis.

Description of drawings

Fig. 1 is the fault diagnosis flowchart of the inventive method;

Fig. 2 is the vibration signal conversion image schematic diagram of the inventive method;

Fig. 3 is a two-dimensional image schematic diagram corresponding to each fault type in the data set of the method of the present invention;

Fig. 4 is the distributed deep neural network structural representation of the inventive method;

Figure 5 is a schematic diagram of the confusion matrix of model branch reasoning alone, model backbone reasoning alone, and model branch backbone collaborative reasoning;

Figure 6 is a schematic diagram of the t-SNE visualization results of the model branch alone inference, the model trunk alone reasoning and the model branch trunk collaborative reasoning.

Detailed ways

A rolling bearing fault diagnosis method based on a distributed deep neural network, comprising the following steps.

S1: The production of the data set, firstly, the acceleration sensor above the bearing seat is used to collect the noise-free vibration acceleration signal of the faulty bearing, and add additive Gaussian white noise to it, and then intercept the vibration signal according to a certain data point length, and It is converted into a two-dimensional image and given real fault labels to construct a data set, and finally the data set is divided into a training set and a test set according to a preset ratio.

S2: Construct a distributed deep neural network model. The model framework includes: a sample input point, a shared convolution block, a branch model, a backbone model, and two sample diagnosis result output points.

S3: Joint training of the branch model and the backbone model. During training, the weighted summation of the loss value of the cross-entropy loss function of each sample diagnosis result output point makes the entire network can be jointly trained and each sample diagnosis result output point is relatively Because of its depth, it can reach the ideal corresponding loss value, so that the predicted value is infinitely close to the real value.

S4: Model reasoning, the branch model performs rapid initial feature extraction on the input test set sample image, and when the output point of the sample diagnosis result of the branch model is confident in the prediction result, the branch exit point exits the sample and outputs the classification result, Otherwise, the image features of the non-exited samples in the shared convolution block are passed to the backbone model for feature extraction and classification in the next step.

S5: Fault identification, summarizing the reasoning results and traffic of the branch model and the main model, and outputting the final fault diagnosis accuracy and the communication traffic consumed.

The step of adding additive Gaussian white noise to each load vibration signal based on noise-free in step S1 is:

S11: Simulate different noise conditions by adjusting the signal-to-noise ratio. The decibel form of the signal-to-noise ratio is expressed as:

为信号功率，/>

为噪声功率，/>

为振动信号数值，N为振动信号长度；In the formula:

is the signal power, />

is the noise power, />

is the value of the vibration signal, N is the length of the vibration signal;

表示，因此对于标准正态分布噪声，其功率为1，因此，首先计算原始信号/>

的功率，然后计算在期望信噪比下产生的噪声信号/>

的功率，最后，通过下面的公式生成加性高斯白噪声，然后将其加到原始信号/>

中，使原始信号具有期望的信噪比；原始振动信号加入加性高斯白噪声的振动信号计算公式为：S12: For a signal with zero mean and known variance, its power can be calculated by variance

Represents, so for standard normally distributed noise, its power is 1, therefore, first calculate the original signal />

power, and then calculate the noise signal generated under the desired signal-to-noise ratio />

The power of , finally, the additive white Gaussian noise is generated by the following formula, and then added to the original signal />

In , the original signal has the desired signal-to-noise ratio; the formula for calculating the vibration signal by adding additive Gaussian white noise to the original vibration signal is:

X = M + x _i

In the formula: M is Gaussian white noise, and randn represents a function that generates a standard normal distribution of random numbers or matrices.

In step S1, the vibration signal is intercepted according to a certain data point length and converted into a two-dimensional image, and the calculation formula of the conversion process is expressed as:

，K表示二维图像的单边尺寸。In the formula: P represents the pixel intensity of the two-dimensional image, L represents the value of the original vibration signal added to the vibration signal of additive Gaussian white noise,

, K represents the single side size of the two-dimensional image.

In step S2, the key structure of the distributed deep neural network includes:

5) The feature vector of the input sample image is extracted by stacking 3×3 convolution kernels to improve the receptive field and nonlinear expression ability of the convolution layer;

6) The number of shared convolution kernels is set to 1, which is used to reduce the amount of communication between the shared convolution block and the adjacent convolution blocks of the backbone and to preserve the original image features to the greatest extent;

7) Place the convolutional feature bag behind the last convolutional block of the branch instead of the fully connected layer to quantify the final output feature vector of the convolutional block and reduce the amount of model parameters;

8) The backbone uses a residual network block and a global average pooling structure to omit the fully connected layer, increase the mobility of the feature vector, and reduce the amount of model parameters.

The joint training process of the model in step S3 is,

S31: The branch model and the backbone model have a classifier respectively, and the output point of each sample diagnosis result takes the cross-entropy loss function as the optimization target, and the cross-entropy loss function is expressed as:

表示样本的预测故障标签，C表示标签集合，/>

表示的是样本从神经网络的输入到第n个出口进行的运算，/>

表示该过程网络的权重和偏置等参数；In the formula: X represents the input sample, y represents the real fault label of the sample,

Indicates the predicted failure label of the sample, C indicates the label set, />

Indicates the operation of the sample from the input of the neural network to the nth exit, />

Represents parameters such as weights and biases of the process network;

S32: Train the weighted sum of the losses of the output points of each sample diagnosis result, and use the SGD method to update the parameters of the distributed neural network. The loss function of the distributed neural network is expressed as:

表示每个出口的权重，/>

表示第n个出口的估计值。In the formula: N represents the number of classified exports,

Indicates the weight of each outlet, />

represents the estimated value of the nth exit.

In step S4, the sample confidence is used as the basis for judging whether the branch exit point has confidence in the prediction result. If the sample confidence is less than a given threshold, it is confident, otherwise, there is no confidence. The formula for calculating the sample confidence is:

。In the formula: C is the set of all real labels, x is the probability vector,

.

In step S5, the communication volume calculation formula for module fault identification is:

In the formula: l is the percentage of branch exit samples to all input samples, f is the feature image size output from the shared convolution block to the backbone model, o is the number of feature image channels output from the shared convolution block to the backbone model, and the constant 4 means Yes, on a 64-bit normal Windows system, a 32-bit floating point number occupies 4 bytes.

Experimental case

Experimental data

28的二维图像，转换图像如图3所示。数据集共45396个样本，将其按照5:1的比例分为训练集和测试集。样本具体组成信息见表1。The experimental data set is the bearing data set released by Case Western Reserve University. This experiment uses the driving end bearing with a sampling frequency of 12Khz. In this data set, there are three types of faults, and each fault type has three different types of damage. size. There are nine fault states and one normal state. The three fault types are roller fault (RF), outer race fault (OF) and inner race fault (IF). The damage sizes are 0.18mm, 0.36mm and 0.54mm, respectively. Adding Gaussian white noise with signal-to-noise ratios of -3dB, 0dB, 3dB, 6dB and 9dB to the vibration signal of the driving end under four loads (0, 1, 2, 3 HP) conditions, will add the variable load of the noise The vibration signal is intercepted every 784 data points and converted into 28

28 two-dimensional image, the converted image is shown in Figure 3. The data set has a total of 45396 samples, which are divided into training set and test set according to the ratio of 5:1. The specific composition information of the samples can be seen in Table 1.

Table 1

model structure

The built model structure is shown in Figure 4. The model framework includes an input point, a shared convolution block, a branch model, a backbone model, and two exit points. The model uses a 3×3 convolution kernel stacking method to extract the feature vector of the input image to improve the receptive field and nonlinear expression ability of the convolution layer. The number of shared convolution kernels is set to 1, which is used to reduce the amount of communication between the shared convolution block and the adjacent convolution blocks of the backbone and to preserve the original image features to the greatest extent. CBoF is placed behind the last convolutional block of the branch instead of the fully connected layer to quantify the feature vectors of the final output of the convolutional block and reduce the amount of model parameters. The backbone uses a residual network block and a global average pooling structure to omit a fully connected layer, increase the mobility of feature vectors, and reduce the amount of model parameters. The model parameters are shown in Table 2.

Table 2

model training

,=0.8分支模型加权值/>

=0.2，batch size为16，优化器为SGD，动量参数为0.9，初始学习率为0.1，学习率衰减值为0.0001，迭代次数为100。The loss weighted sum of the two outlets is combined for joint training, and the model hyperparameters are set as follows: backbone model weighted value

,=0.8 branch model weighted value/>

=0.2, the batch size is 16, the optimizer is SGD, the momentum parameter is 0.9, the initial learning rate is 0.1, the learning rate decay value is 0.0001, and the number of iterations is 100.

(1) Test of the inventive method in variable load, multi-noise mixed scene

Table 3 shows the inference accuracy and communication cost of the trained model in variable load, multi-noise and variable load single noise environments. According to the test results in variable load and multi-noise environment, the model branch can handle 47.98% of the samples and the recognition accuracy is 100%, and the overall recognition accuracy of the model reaches 99.27%, which proves that the model has good anti-interference ability; According to the test results of variable load and single noise environment, the number of model branch processing samples is inversely proportional to the communication cost required by the model in the collaborative reasoning process. The number of model branch processing samples varies in various noises, but the recognition accuracy is uniform. It can reach 100%; -3dB noise interferes the most with model inference, but the overall recognition accuracy of the model can also be maintained at 96%. In the process of changing the test environment from -3dB to no noise, the overall accuracy of model inference continues to improve. Communication costs are constantly decreasing.

table 3

(2) The inventive method is compared with other neural network models

表示模型分支与主干协同推理。从各模型单独推理的结果中可以看出，DDNN（T=0）在变载荷、多噪声和变载荷、单一噪声环境中，其识别精度均优于其他单独推理的模型，虽然在推理速度方面稍落后于其他模型，但其参数量分别是AlexNet、LeNet-5、Vgg16模型参数量的54%、37%、20%，DDNN（T=1）模型参数量最少，推理速度最快，但其识别精度低；从模型分支单独推理、模型主干单独推理和模型分支和主干的协同推理结果中可以看出，DDNN（协同）在保证识别精度的前提下，在增加560个参数后，其相比于模型主干单独推理，推理速度提高了18%，通讯成本降低了32%（模型主干单独推理的平均通讯量为4×28×28B），模型分支单独推理的精度表现最差。Table 4 shows the inference results of the constructed model and other neural network models in terms of inference speed, model parameters, communication cost, variable load, multiple noise and variable load, and recognition accuracy in a single noise environment. T=0 means The model trunk reasoning alone, T=1 means the model branch reasoning alone, T

Indicates that the model branch is reasoned in conjunction with the trunk. It can be seen from the results of individual reasoning of each model that the recognition accuracy of DDNN (T=0) is better than that of other independent reasoning models in variable load, multi-noise, variable load, and single noise environments, although in terms of inference speed It is slightly behind other models, but its parameters are 54%, 37%, and 20% of AlexNet, LeNet-5, and Vgg16 models, respectively. DDNN (T=1) model has the least number of parameters and the fastest inference speed, but its The recognition accuracy is low; it can be seen from the results of separate reasoning of the model branch, single reasoning of the model backbone, and collaborative reasoning of the model branch and the backbone. Under the premise of ensuring the recognition accuracy, after adding 560 parameters, the DDNN (collaboration) is better than the Based on the reasoning of the model backbone alone, the reasoning speed is increased by 18%, and the communication cost is reduced by 32% (the average communication volume of the model backbone reasoning alone is 4×28×28B), and the accuracy of the model branch reasoning alone is the worst.

Table 4

Confusion Matrix Analysis

The confusion matrix of model branch independent reasoning, model backbone independent reasoning, and model branch backbone collaborative reasoning is shown in Figure 5. The overall verification accuracy of the former is only 91.74%. Label 8 is the most likely to be misidentified, and label 0 is the least likely to be misidentified. The misrecognition rates of the two are 28.17% and 0 respectively; the overall accuracy of the latter two verifications is 99.27%, label 8 is also the most likely to be misrecognized, and labels 0, 1, 3, 6, and 7 are the least likely to be misrecognized , their false recognition rates are 3.33% and 0, respectively.

t-SNE Visual Analysis

Figure 6 shows the t-SNE visualizations of model branch independent reasoning, model backbone independent reasoning, and model branch-background collaborative reasoning. The former has a large number of label samples with obvious decision boundary blurring. In the results of model collaborative reasoning, the model branch processing There is no decision boundary model phenomenon in the sample, the result of the model backbone processing the sample is similar to the result of the model backbone's independent reasoning, and there is a phenomenon that the decision boundary of a very small number of labeled samples is blurred.

Aiming at the problems of low fault diagnosis accuracy and large diagnosis time delay of rolling bearings under complex working conditions of variable load and large noise, the present invention proposes a rolling bearing fault diagnosis method based on a distributed deep neural network. This method uses data grayscale to convert the bearing vibration signal into an image to construct an image data set; it proposes to add a branch at the bottom of the model as a branch model, withdraw some simple samples in advance, and replace the fully connected layer of the branch model with a convolutional feature bag ; Through the collaborative calculation of the branch model and the main model, the fault diagnosis of rolling bearings with high precision, low communication cost and low delay under the complex working conditions of variable load and large noise is finally realized. In addition, the constructed model still has a good effect in the environment of variable load and single noise, showing good generalization ability.

Claims (7)

1. A rolling bearing fault diagnosis method based on a distributed deep neural network is characterized in that: comprises the steps of,

the method comprises the steps of S1, manufacturing a data set, namely firstly, collecting a noise-free vibration acceleration signal of a fault bearing through an acceleration sensor above a bearing seat, adding additive Gaussian white noise to the vibration signal, then intercepting the vibration signal according to a certain data point length, converting the vibration signal into a two-dimensional image, endowing the two-dimensional image with a real fault label to construct the data set, and finally dividing the data set into a training set and a testing set according to a preset proportion;

s2, constructing a distributed deep neural network model, wherein the model framework comprises: one sample input point, one shared convolution block, one branch model, one trunk model and two sample diagnosis result output points;

s3, carrying out combined training on the branch model and the trunk model, wherein during training, the whole network can be combined trained by weighting and summing the loss values of the cross entropy loss functions of the output points of the diagnosis results of each sample, and the output points of the diagnosis results of each sample can reach ideal corresponding loss values relative to the depth of the output points of the diagnosis results of each sample, so that the predicted values are infinitely close to the true values;

s4, model reasoning, namely quickly extracting initial characteristics of an input test set sample image by a branch model, exiting the sample by a branch outlet point and outputting a classification result under the condition that a sample diagnosis result output point of the branch model has confidence to a prediction result, otherwise, transmitting the image characteristics of the sample which is not exited in a shared convolution block to a trunk model, and extracting and classifying the characteristics of the next step;

and S5, fault identification, namely summarizing the results and the communication quantity of the reasoning of the branch model and the trunk model, and outputting the final fault diagnosis precision and the consumed communication quantity.

2. The rolling bearing fault diagnosis method based on the distributed deep neural network according to claim 1, wherein: the step S1 of adding additive Gaussian white noise to each load vibration signal based on no noise comprises the following steps:

s11: different noise conditions are simulated by adjusting the signal-to-noise ratio, which is expressed in decibel form as:

wherein:

for signal power, +.>

Is noise power +.>

In order to be a value of the vibration signal,Nis the vibration signal length;

s12: for a signal with zero mean and known variance, the power can be calculated using the variance

It means that, therefore, for a standard normal distributed noise, its power is 1, and therefore, the original signal +.>

And then calculate the power at the desired signal to noiseNoise signal generated by comparison->

Finally, additive white gaussian noise is generated by the following formula and then added to the original vibration signal +.>

In, let the original vibration signal +.>

With a desired signal-to-noise ratio; the calculation formula of the vibration signal with the original vibration signal added with the additive Gaussian white noise is as follows:

X=M+x _i

wherein: m is gaussian white noise and randn represents a function that produces a random number or matrix of standard normal distribution.

3. The rolling bearing fault diagnosis method based on the distributed deep neural network according to claim 2, characterized in that: in the step S1, the vibration signal is intercepted according to a certain data point length and is converted into a two-dimensional image, and the calculation formula of the conversion process is expressed as follows:

wherein:Prepresenting the pixel intensities of a two-dimensional image,La value representing the addition of the original vibration signal to the additive white gaussian noise vibration signal,

，Krepresenting the single-sided size of the two-dimensional image.

4. The rolling bearing fault diagnosis method based on the distributed deep neural network according to claim 1, wherein: in the step S2, the key structure of the distributed deep neural network includes:

1) Extracting feature vectors of an input sample image by adopting a 3×3 convolution kernel stacking mode so as to improve the receptive field and nonlinear expression capability of a convolution layer;

2) The number of the shared convolution kernels is set to be 1, so that the communication quantity between the shared convolution blocks and the adjacent convolution blocks of the trunk is reduced, and the original image characteristics are reserved to the maximum extent;

3) The convolution feature bag is placed behind the last convolution block of the branch to replace a full connection layer, so that the feature vector finally output by the convolution block is quantized, and the model parameter quantity is reduced;

4) The trunk adopts a residual network block and a global average pooling structure, so that a full connection layer is omitted, the fluidity of the feature vector is increased, and the number of model parameters is reduced.

5. The rolling bearing fault diagnosis method based on the distributed deep neural network according to claim 1, wherein: the combined training process of the model in the step S3 is that,

s31: the branch model and the trunk model are respectively provided with a classifier, and each sample diagnosis result output point takes a cross entropy loss function as an optimization target, wherein the cross entropy loss function is expressed as:

wherein:

representing input samples, +_>

True failure label representing sample, +.>

A predictive failure tag representing a sample, C represents a tag set, <' > and->

Representing the operation of the sample from the input of the neural network to the nth exit, +.>

Weights and bias parameters representing the process network;

s32: training the loss weighted summation of the output points of the diagnosis results of each sample, and updating the parameters of the distributed neural network by adopting an SGD method, wherein the loss function of the distributed neural network is expressed as follows:

wherein: n represents the number of sort outlets,

representing the weight of each exit, +.>

Representing an estimate of the nth exit.

6. The rolling bearing fault diagnosis method based on the distributed deep neural network according to claim 1, wherein: in the step S4, the confidence coefficient of the sample is used as a basis for judging whether the confidence coefficient of the branch outlet point has the confidence to the prediction result, if the confidence coefficient of the sample is smaller than the given threshold value, the confidence coefficient is the confidence, otherwise, the confidence coefficient of the sample is not the confidence, and the calculation formula of the confidence coefficient of the sample is:

wherein:Cfor the set of all the real tags,xas a vector of the probability that the vector of the probability,

。

7. the rolling bearing fault diagnosis method based on the distributed deep neural network according to claim 1, wherein: in the step S5, the traffic calculation formula for identifying the mode fault is as follows:

wherein:

for branch exit samples as a percentage of the total input samples, +.>

To share the feature image size of the convolution block output to the backbone model, o is the number of feature image channels of the shared convolution block output to the backbone model, and constant 4 means that in a 64-bit normal Windows system, one 32-bit floating point number occupies 4 bytes.

Images