CN112308038A - Mechanical equipment fault signal identification method based on classroom type generation confrontation network model - Google Patents

Abstract

A method for identifying fault signals of mechanical equipment based on a classroom-type generative adversarial network model relates to the field of identifying fault signals of machinery. The invention aims to solve the problem of low accuracy of the existing mechanical equipment fault signal identification method. The method for identifying mechanical equipment fault signals based on a classroom-based generative adversarial network model comprising a generator and a plurality of discriminators according to the present invention includes: acquiring normal vibration signals of mechanical equipment and mechanical equipment fault vibration signals; Divide into test set and training set; set classroom-style generative adversarial network structure parameters; obtain a batch of samples; calculate the improvement value of generation ability; calculate the influence weight of each generator on the value of the discriminator loss function; calculate the loss function of the discriminator ; Calculate the loss function of the generator; test the accuracy of the discriminator; input the mechanical equipment vibration signal into the classification model with the highest accuracy to obtain the recognition result.

Description

Identification method of mechanical equipment fault signal based on classroom-based generative adversarial network model

technical field

The present invention relates to the field of fault signal identification, in particular to a method for identifying fault signals of mechanical equipment based on a classroom-based generative confrontation network model.

Background technique

In modern industry, the occurrence of mechanical equipment failures will bring greater safety hazards to the equipment. Therefore, in order to ensure the safety of the equipment, the fault information should be effectively analyzed to understand and master the state of the equipment during use, and determine its overall status. Or part of it is normal or abnormal, it is very necessary to detect faults early, and to judge and eliminate the occurrence of faults.

Among the existing fault diagnosis, the most common method is to detect and locate faults according to the vibration signal characteristics of equipment fault signals. According to the abnormal vibration signal characteristics of the equipment when the fault occurs, it is judged whether there is a fault. Most of the traditional machine learning-based fault signal identification methods are based on balanced data, but in real life, mechanical equipment fault signal data collection is difficult and the volume is small, which leads to the current mechanical equipment fault signal identification methods are accurate. low rate problem.

SUMMARY OF THE INVENTION

The purpose of the present invention is to solve the problem of low accuracy of the existing mechanical equipment fault signal identification methods, and propose a mechanical equipment fault signal identification method based on a classroom-type generative confrontation network model.

A method for identifying fault signals of mechanical equipment based on a classroom-based generative adversarial network model. The specific process is as follows:

Step 1. Obtain the normal vibration signal of the mechanical equipment and the vibration signal of the mechanical equipment failure;

Step 2: Divide the mechanical equipment vibration signal obtained in step 1 into a test set and a training set;

The normal vibration signal of the mechanical equipment is divided into a training set and a test set by random sampling;

All the vibration signals of the mechanical equipment failure are used as a test set;

Step 3: Set the structural parameters of the classroom-style generative adversarial network;

Step 4: Sampling from the training set according to the prior probability distribution to obtain a batch of samples;

Step 5. Use the obtained samples to train the generator and calculate the improvement value of each generator's generation ability in the previous training;

The improvement value of the generation ability is the value of the change in the loss function value of the discriminator for each generated data before and after the previous training.

Step 6, utilize the weight function to calculate the influence weight λ _i,t of each generator to the discriminator loss function value according to the boost value of the generation ability calculated in step 5;

Step 7: Calculate the loss function of the discriminator using the obtained samples, and use the loss function value to train the discriminator;

Step 8: Calculate the loss function value of each generator by using the obtained samples and train each generator by using the loss function value;

Step 9. Use the discriminator to classify the test set data and test the accuracy of the discriminator;

Step 10: Obtain the vibration signal of the mechanical equipment to be detected and input the discriminator model with the highest classification accuracy for the test set data.

The beneficial effects of the present invention are:

The invention improves the existing generative confrontation network model, constructs the generative confrontation network model into a model with multiple generators and a discriminator, and proposes a weight distribution function to adaptively adjust the difference between the generators and the discriminator loss function. Influence the weights, so that the generators work together to improve the fit of the discriminator and the training sample space, train a discriminator with excellent performance, and apply it to the task of mechanical equipment fault signal identification, thereby improving the detection of mechanical equipment faults. The accuracy of signal recognition.

Description of drawings

Figure 1 is a structural diagram of a classroom-based generative adversarial network model;

Fig. 2 is the structure diagram of the discriminator used for identifying the fault signal of mechanical equipment.

Detailed ways

Embodiment 1: The specific process of the method for identifying fault signals of mechanical equipment based on the classroom-based generative adversarial network model in this embodiment is as follows:

Step 1. Obtain the normal vibration signal of the mechanical equipment and the vibration signal of the mechanical equipment failure;

Step 2: Divide the mechanical equipment vibration signal obtained in step 1 into a test set and a training set;

The normal vibration signal of the mechanical equipment is divided into a training set and a test set by random sampling;

All the mechanical equipment fault vibration signals are used as a test set;

Step 3. Set classroom-style generative adversarial network structure parameters:

Step 31. Establish a classroom-style generative adversarial network model:

The classroom-style generative adversarial network model contains a discriminator and multiple generators:

The generator that builds the model:

where G is the generator, N is the number of generators, X is the generated data, Z is the noise variable, and _Gi is the ith generator;

The input data and the discriminant network are shared among the generators, and the generator hybrid structure provides the learning signal for the discriminator;

Step 32: Setting the structural parameters of the classroom-based GAN, including: the number of generators, the classification model structure based on the classroom-based GAN, the number of training times, the number of times to start training, adjustment parameters, and batch size.

Step 4: Randomly sample from the training set according to the prior probability distribution to obtain a batch of samples:

where x is the mechanical equipment vibration signal in the training set, z is the noise sample, i is the sample number, px is the probability distribution of _x , _pz is the probability distribution of z, and m is the batch.

Step 5. Use the obtained samples to train the generator, and calculate the improvement value of each generator's generation ability in the previous training from the third training (set Qi _{and t} to 0 in the first two trainings):

是当z_t-1～p_z时的期望值，

是当z_t-2～p_z时的期望值。Among them, D _t-1 is the discriminator after the t-1th training, D _t-2 is the discriminator after the t-2th training, G _i,t-1 is the t-1th training i generators, G _{i, t-2} is the i-th generator after the t-2th training, z _t-1 is the noise sample sampled in the t-1th training; z _t-2 is the t-th - 2 noisy samples sampled in training,

is the expected value when z _t-1 ~ p _z ,

is the expected value when z _t-2 ~ p _z .

Step 6: Calculate the influence weight λ _i,t of each generator on the discriminator loss function value by using the weight function according to the improvement value of the generation ability calculated in step 5:

Among them, Q _i,t is the generation ability improvement value of the i-th generator before the t-th training, α≥0 is the hyperparameter that can be adjusted, and N is the number of generators in the model.

Step 7: Calculate the loss function of the discriminator using the obtained samples and use the loss function value to train the discriminator. The loss function of the discriminator is:

是判别器的损失函数，，G_i(z)是第i个生成器的生成样本，D_t-1(x)是第t-1次训练后的判别器对输入样本来源的预测值，z_t是第t次训练中采样的噪声样本；in,

is the loss function of the discriminator, G _i (z) is the generated sample of the ith generator, D _t-1 (x) is the predicted value of the input sample source of the discriminator after the t-1th training, z _t is the noise sample sampled in the t-th training;

也就是设置超参数α＝0，则：When t<T _start-up , each generator has the same influence weight on the discriminator loss function value, namely

That is, set the hyperparameter α=0, then:

Among them, T _start-up is the number of training times in the model training startup phase, and D _loss is the loss function of the discriminator when t < T _start-up .

Step 8. Use the obtained samples to calculate the loss function value of each generator and use the loss function value to train each generator. The loss function of the generator is:

是生成器的损失函数。where z _t is the noise sample sampled in the t-th training, G _i,t-1 is the ith generator after the t-1th training, D _t is the discriminator after the t-th training,

is the loss function of the generator.

Step 9. Use the discriminator to classify the test set data and test the accuracy of the discriminator:

Input the test sample into the discriminator, if the output value is 1, the sample is the normal vibration signal of the mechanical equipment; if the output value is 0, it is the vibration signal of the mechanical equipment failure;

Compare the output of the discriminator with the real sample labels, if they are the same, the judgment is accurate, if they are different, the judgment is wrong, and then calculate the average accuracy of all samples.

Step 10: Obtain the vibration signal of the mechanical equipment to be detected and input the discriminator model with the highest classification accuracy for the test set data.

Example:

According to the technical solution of the specific embodiment, a classroom-based generative adversarial network model is obtained, and the task of identifying mechanical fault signals is completed:

Aiming at the classroom-type generative adversarial network structure proposed by the present invention, experimental analysis is carried out with the data set provided by the Bearing Data Center of Case Western Reserve University (CWRU) in the United States to verify the discrimination based on classroom-type generative adversarial network training. The ability of the device to identify fault signals of mechanical equipment. The detected bearing in the CWRU dataset supports the motor shaft, the drive end bearing is SKF6205, the fan end bearing is SKF6203, the bearing is single-point damaged by EDM, and an acceleration sensor is placed above the motor fan end and the drive end bearing seat to collect the bearing. The vibration acceleration signal is obtained, the processed bearing is installed in the test motor, and the vibration acceleration signal data is collected under the motor load conditions of 0, 1, 2 and 3 horsepower respectively, and the sampling frequency is 12KHz. The fault and inner ring fault diameters are 0.007 inches (0.1778mm), 0.014 inches (0.3556mm) and 0.021 inches (0.5334mm) of the drive end bearing fault data for experiments. Five common neural network structures are selected in the experiment, namely feedforward neural network. Network (FF, Feedforward neural networks), deconvolution network (DN, Deconvolutional networks), long short-term memory network (LSTM, Long short-term memory), radial basis neural network (RBF, Radial basis function) and residual network (RN, Residual networks) is used as the generator model, and the discriminator network uses convolutional neural networks (CNN, Convolutional neural networks).

Experiments are carried out on CWRU data, and the final classification accuracy of each discriminator for the category data that participates in training (signal data of non-faulty bearing vibration data) and the category data that does not participate in training (signal data of faulty bearing vibration data) is shown in Table 1. DC, FF, LSTM, RBF and RN represent the generator model used in the experiment, and the coloring below indicates the use of the generator in the model. For example, the generator model used in the model No. 17 includes DC, FF and RBF, the three accuracy rates are the classification accuracy rates of the discriminator for normal samples and fault samples with three different fault diameters, respectively. Rank 1 indicates the ranking of the model's classification accuracy among all models, and rank 2 indicates the model. Ranking in the same number of generator models; the last two rows show the comparison between the One-Class Support Vector Machine (OCSVM) and the Support Vector Data Description (SVDD) as the comparative classifiers in the data classification results on the set. Table 2 summarizes the training time of the most complex classroom-style generative adversarial network (the number of generators is 5, and the number of generators is 31 in Tables 1 and 2), OCSVM and SVDD, and the prediction time for test samples.

Table 1

Table 2

It can be seen from the experimental results that the classifier trained by the classroom-based generative adversarial network can effectively distinguish the vibration data signal of the fault-free bearing from the vibration data signal of the faulty bearing, and obtain a good identification effect. From the classification effect of the discriminator on normal samples and fault samples with a fault diameter of 0.014in in the experiment, it can be seen that the discriminator of the classroom-style generative adversarial network combining RN, DC and LSTM with the performance rankings of 1, 2 and 4 The performance (No. 21) is better than the discriminator for the RN, RBF and LSTM combination (No. 25) ranked 1, 3 and 4, and similarly, the generator ranking 2, 4 and 5 DC, LSTM and FF combination ( No. 16) is better than the discriminator obtained by the combination of RBF, LSTM and FF ranked 3, 4 and 5 (No. 22), which is due to the better performance of the generator's generation space in fitting fault-free bearing vibration data. The signal sample space is closer to the signal sample space of the faultless bearing vibration data, and the obtained discriminator is more "fit" to the training sample space, and the signal recognition performance is better.

The combination of generators with small differences in performance may obtain unexpected results, and the corresponding discriminator of the classroom-style generative adversarial network has a strong signal recognition ability. In the classification results of the samples, the generation performance of DC and RBF ranked second (accuracy rate 96.070%) and third (accuracy rate 95.985%). The performance ranks 1st among all dual generator models (98.985% accuracy), and the recognition accuracy has a certain improvement, which is better than the combination of RN (1st, 96.855% accuracy) and RBF (8th, accurate rate 94.605%), this is because the generators with similar performance "close" to the sample space of the fault-free bearing vibration data signal at a similar speed during the training process, the generators cooperate with each other, the generation performance is improved, and the corresponding discriminator also has better good classification ability.

However, for the generator combinations with large performance differences, due to the large difference between the time when the discriminator performance is optimal or the sample space of the generator fitting the fault-free bearing vibration data signal, it is impossible to achieve synchronization "close" in the training process without faults. In the sample space of bearing vibration data signal, the classification ability of the obtained discriminator is not ideal, even inferior to the discriminator in the single generator model. For example, 2. In the classification results of the discriminator for normal samples and fault samples with a fault diameter of 0.021in in the experiment, the generation performance of FF (accuracy rate 83.750%) is higher than that of DC (accuracy rate 98.500%) and RN (accuracy rate 99.005%) The generation performance of these generators is quite different. The discriminator performance of the dual-generator classroom-style generative adversarial network combined with FF is not as good as the discriminator performance in the single-generator model. Therefore, choosing the correct combination of generators can improve the classification ability of the discriminator. have a greater impact.

Claims (8)

1. The mechanical equipment fault signal identification method based on the classroom type generation confrontation network model is characterized by comprising the following steps of:

acquiring a normal vibration signal of mechanical equipment and a fault vibration signal of the mechanical equipment;

step two, dividing the vibration signals of the mechanical equipment obtained in the step one into a test set and a training set;

dividing the normal vibration signals of the mechanical equipment into a training set and a test set in a random sampling mode;

all the mechanical equipment fault vibration signals are used as a test set;

step three, setting classroom type generation confrontation network structure parameters;

step four, sampling from the training set line according to prior probability distribution to obtain a batch of samples;

fifthly, training the generators by using the obtained samples and calculating the improvement value of the generating capacity of each generator in the previous training;

the raising value of the generating capacity is a value of the change of the loss function value judged by the judger on each generating data before and after the previous training;

step six, calculating the influence weight lambda of each generator on the loss function value of the discriminator by using a weight function according to the lifting value of the generating capacity calculated in the step five_i,t；

Step seven, calculating a loss function of the discriminator by using the obtained samples and training the discriminator by using the loss function value;

step eight, calculating a loss function value of each generator by using the obtained samples and training each generator by using the loss function value;

classifying the test set data by using a discriminator, and testing the accuracy of the discriminator to obtain a discriminator model with the highest accuracy;

step ten, acquiring a vibration signal of the mechanical equipment to be detected and inputting the vibration signal into a discriminator model with the highest classification accuracy of the data of the test set.

2. The method for identifying fault signals of mechanical equipment based on the classroom-type generated countermeasure network model as claimed in claim 1, wherein the classroom-type generated countermeasure network structure parameters are set in the third step, and the specific process is as follows:

step three, establishing a classroom-based generation confrontation network model:

the classroom-based generation confrontation network model comprises a discriminator and a plurality of generators:

a generator for constructing a model:

where G is the generator, N is the number of generators, X is the generated data, Z is the noise variance, G is the noise variance_iIs the ith studentForming a device;

wherein, the generators share input data and a discrimination network, and the generator mixed structure provides a learning signal for the discriminator;

step two, setting the classroom type generation confrontation network structure parameters, including: the number of generators, a class-based generation of a classification model mechanism of the countermeasure network, training times, starting training times, parameter adjustment and batch size.

3. The method for identifying fault signals of mechanical equipment based on classroom-based generation of countermeasure network models as claimed in claim 2, wherein the step four is randomly sampling from the training set according to prior distribution to obtain a batch of samples:

where x is the true mechanical equipment vibration signal in the training set, z is the noise sample, i is the sample number, p_xIs the probability distribution of x, p_zIs the probability distribution of z, and m is the batch.

4. The method for identifying fault signals of mechanical equipment based on classroom-type generation countermeasure network model according to claim 3, wherein the fifth step is to calculate the improvement value of the generating capacity of each generator in the previous training by:

wherein D is_t-1Is the discriminator after the t-1 training, D_t-2Is the discriminator after the t-2 training, G_i,t-1Is the ith generator after the t-1 training, G_i,t-2Is the ith generator after the t-2 training, z_t-1Is the noise sample sampled in the t-1 st training; z is a radical of_t-2Is the noise sample sampled in the t-2 training,

is when z is_t-1～p_zThe expected value of the time of day,

is when z is_t-2～p_zThe expected value of the time.

5. The method as claimed in claim 4, wherein the sixth step calculates the weight λ of the effect of each generator on the function value of the loss function of the discriminator by using a weight function according to the boost value of the generation capability calculated in the fifth step_i,tThe specific process is as follows:

wherein Q is_i,tThe generation capacity improvement value of the ith generator before the t training is shown, alpha is more than or equal to 0 and is a super parameter for adjustment, and N is the number of generators in the model.

6. The class-based method for identifying faults of mechanical equipment for generating countermeasure network models according to claim 5, wherein the loss function of the discriminator in the seventh step is:

wherein,

is the loss function of the arbiter, D (x) is the predicted value of the arbiter for the sample source, G_i(z) is a generated sample of the i-th generator, D_t-1(x) Is the predicted value of the discriminator after the t-1 training on the source of the input sample, z_tIs the noise sample sampled in the t-th training;

when T < T_start-upEach generator has the same weight of influence on the value of the discriminator loss function, i.e.

That is, if the super parameter α is set to 0, then:

wherein, T_start-upIs the number of training times in the starting phase of the model training, D_lossIs T < T_start-upThe penalty function of the time arbiter.

7. The class-based generation of mechanical equipment failure signal identification of countermeasure network models of claim 6, wherein the loss function of each generator in step eight is:

wherein z is_tIs the noise sample sampled in the t-th training, G_i,t-1Is the ith generator after the t-1 training, D_tIs the discriminator after the t-th training,

is a loss function of the generator.

8. The class-based mechanical equipment fault signal identification method for generating the countermeasure network model according to claim 7, wherein in the ninth step, a discriminator is used for classifying the test set data, and the accuracy of the discriminator is tested by the following specific processes:

inputting a test sample into a discriminator, wherein if the output value is 1, the sample is a normal signal of the mechanical equipment, and if the output value is 0, the sample is a fault signal of the mechanical equipment;

and comparing the output result of the discriminator with the label of the real sample, judging accurately if the output result is the same as the label of the real sample, judging wrongly if the output result is different from the label of the real sample, and then calculating the average accuracy of all samples.

Images