A fault diagnosis method and system based on standard self-learning data enhancement

Abstract

The present invention proposes a fault diagnosis method and system based on standard self-learning data enhancement, which relate to the technical field of bearing fault diagnosis. The method comprises: constructing a fault diagnosis model based on a one-dimensional convolutional neural network; training the fault diagnosis model through a cross-adversarial training method of standard self-learning and data enhancement to obtain a complete data set and an intelligent fault diagnosis model under strong non-stationary working conditions; inputting the collected vibration signal to be diagnosed into the trained intelligent fault diagnosis model to obtain a bearing fault diagnosis result; the present invention takes a one-dimensional convolutional neural network as the basic framework, utilizes an incomplete training data set, and generates disturbance data through a cross-adversarial training method of standard self-learning and data enhancement to obtain a fault diagnosis model under strong non-stationary working conditions, thereby improving the accuracy of fault diagnosis.

Description

Technical Field

[0001] The present invention belongs to the technical field of bearing fault diagnosis, and in particular relates to a fault diagnosis method and system based on standard self-learning data enhancement. Background Art

[0002] The statements in this section merely provide background information related to the present invention and do not necessarily constitute prior art.

[0003] Rolling bearings, as the most widely used rotating components, are the core of internal motion conversion and power transmission in high-end equipment. Rolling bearings often operate under highly non-stationary conditions. During this process, the drastic fluctuations in load and speed lead to frequent rolling bearing failures on the one hand, and accelerate the expansion of damage on the other hand, thereby aggravating the hazards of failures. Therefore, rolling bearing fault diagnosis under highly non-stationary conditions is of great significance for ensuring the safe and efficient operation of high-end equipment.

[0004] Processing and analyzing the dynamic signals captured by health monitoring equipment is the most commonly used means of diagnosing rolling bearing faults. As health monitoring develops towards high precision, multi-dimensional and full-time, modern health monitoring equipment collects massive dynamic signals, which has brought fault diagnosis into the "big data" era. As a result, traditional fault diagnosis methods based on signal analysis are difficult to meet the requirements of diagnostic efficiency, which has given rise to deep learning intelligent fault diagnosis methods driven by "big data". At the same time, with the increase in equipment complexity, the signals collected under strong non-stationary working conditions are not only accompanied by extremely strong noise, but also the fault impact characteristics and other components are strongly coupled, overlapped, distorted and other phenomena, which greatly increases the difficulty of signal analysis. Therefore, under strong non-stationary working conditions, the demand for deep learning intelligent fault diagnosis methods is even more urgent.

[0005] Deepening the structure means that it is easier to extract accurate fault features, but it also causes the model to overfit to the training data, which highlights the importance of complete training data; for incomplete health monitoring data, deepening the model structure to extract target features can only cause it to overfit to limited diagnostic knowledge and cannot meet actual diagnostic needs. Therefore, complete health monitoring training data is the basic prerequisite for the implementation of intelligent fault diagnosis methods.

[0006] A complete training data set under strongly non-stationary working conditions requires the superposition completeness of three-dimensional continuous information: fault, instantaneous working condition (speed, load, etc.), and working condition change rate (speed, load change rate, etc.). That is, samples of each fault must be collected under any instantaneous working condition and any working condition change rate. Such a stringent requirement is impossible to achieve in practice. In practice, once a fault is found in the equipment, it must be shut down for maintenance to prevent serious accidents. The fault sample is only a dynamic signal of uniform deceleration, in which the working condition change rate information is extremely single, and a certain range of instantaneous working condition information (such as speed) is inevitably missing, which is far from meeting the completeness requirement. It can be seen that the training data collected under strongly non-stationary working conditions is extremely incomplete, which seriously restricts the development of intelligent fault diagnosis.

[0007] Data augmentation (DA) is the most direct method to deal with incomplete data sets by generating new training samples. Traditional methods originated from image recognition pre-processing, such as image rotation and magnification. In recent years, Generative Adversarial Networks (GAN), as a data intelligent generation method, has become a hot topic for data augmentation. Some GAN-based data augmentation methods have also been proposed in the field of intelligent fault diagnosis of rotating machinery. Zhou et al. designed a GAN generator and discriminator, and used a global optimization scheme to generate more samples to deal with the data imbalance problem. Shao et al. and Guo et al. developed a GAN-based auxiliary classifier framework and a multi-label one-dimensional GAN, respectively, to learn from mechanical sensor signals and generate data closer to reality to solve the problem of insufficient data.

[0008] Existing data augmentation methods primarily address issues such as unbalanced datasets and small data volumes. By generating samples closer to the original data, the data volume is expanded, thereby improving the model's diagnostic accuracy. However, data generation aimed at achieving data similarity only yields convergent data. Typically, rotating machinery health monitoring data under strongly non-stationary conditions is limited to uniform deceleration datasets with missing information. Simply pursuing similarity in generated data only expands the data volume but fails to compensate for the dataset's lack of information. Only by generating diverse samples can datasets under strongly non-stationary conditions meet the requirement for superposition completeness of three-dimensional continuous information. Therefore, data generation focuses on the differences between the generated data and the original data. Summary of the Invention

[0009] In order to overcome the shortcomings of the above-mentioned prior art, the present invention provides a fault diagnosis method and system based on standard self-learning data enhancement. With a one-dimensional convolutional neural network as the basic framework and an incomplete training data set, disturbance data is generated through a cross-adversarial training method of standard self-learning and data enhancement, thereby obtaining a fault diagnosis model under strongly non-stationary working conditions and improving the accuracy of fault diagnosis.

[0010] To achieve the above objectives, one or more embodiments of the present invention provide the following technical solutions:

[0011] A first aspect of the present invention provides a fault diagnosis method based on standard self-learning data enhancement;

[0012] A fault diagnosis method based on standard self-learning data enhancement, comprising:

[0013] Based on one-dimensional convolutional neural network, a fault diagnosis model is constructed;

[0014] The fault diagnosis model is trained through a cross-adversarial training method combining standard self-learning and data enhancement to obtain an intelligent fault diagnosis model with a complete data set and under strong non-stationary working conditions.

[0015] The collected vibration signal to be diagnosed is input into the trained intelligent fault diagnosis model to obtain the bearing fault diagnosis result.

[0016] Furthermore, the one-dimensional convolutional neural network includes multiple layers of convolutional layers, pooling layers and fully connected layers;

[0017] The convolution layer uses ReLU (Rectified Linear Unit) as the activation function, and the stride of the convolution operation is 1;

[0018] All convolutional layers are connected to pooling layers to reduce the dimensionality of the output features of the convolutional layers;

[0019] The features of the input samples after multi-layer convolution and pooling are flattened into a one-dimensional vector, and then fault diagnosis is performed through three layers of full connection.

[0020] Furthermore, the standard self-learning aims to learn classification knowledge, optimizes parameters in the fault diagnosis model by repeatedly inputting updated samples, and self-learns the evaluation criteria for determining whether a sample is a disturbance sample.

[0021] Furthermore, the data enhancement is guided by the output of the model itself and generates perturbation samples through sample parameterization and model dataization methods;

[0022] Among them, the judgment criterion for the perturbation sample is whether it can interfere with the model judgment, specifically: after the sample is input into the model, it can cause a disturbance in the posterior probability of the model.

[0023] Furthermore, the sample parameterization is to regard the samples as model parameters, train the parameters that reduce the objective function through the stochastic gradient descent method, and then derive the parameters as generated samples.

[0024] Furthermore, the model digitization is to regard the parameters of the fault diagnosis model as data and fix the parameter values ​​during the training process.

[0025] Furthermore, the output of the intelligent fault diagnosis model is the posterior probability that the vibration signal to be diagnosed belongs to each fault type. The probabilities are sorted, and the fault type with the highest probability is the final bearing fault diagnosis result.

[0026] A second aspect of the present invention provides a fault diagnosis system based on standard self-learning data enhancement.

[0027] A fault diagnosis system based on standard self-learning data enhancement, including a model building module, a model training module and a fault diagnosis module;

[0028] The model building module is configured to: build a fault diagnosis model based on a one-dimensional convolutional neural network;

[0029] The model training module is configured to train the fault diagnosis model through a cross-adversarial training method combining standard self-learning and data enhancement to obtain an intelligent fault diagnosis model with a complete data set and under strong non-stationary working conditions.

[0030] The fault diagnosis module is configured to: input the collected vibration signal to be diagnosed into the trained intelligent fault diagnosis model to obtain the bearing fault diagnosis result.

[0031] A third aspect of the present invention provides a computer-readable storage medium having a program stored thereon, which, when executed by a processor, implements the steps of a fault diagnosis method based on standard self-learning data enhancement as described in the first aspect of the present invention.

[0032] The fourth aspect of the present invention provides an electronic device, including a memory, a processor, and a program stored in the memory and runnable on the processor. When the processor executes the program, it implements the steps of a fault diagnosis method based on standard self-learning data enhancement as described in the first aspect of the present invention.

[0033] One or more of the above technical solutions have the following beneficial effects:

[0034] The present invention proposes a standard self-learning data enhancement method, which uses the fault diagnosis model's own prediction results as the evaluation standard for data generation, generates disturbance samples through sample parameterization and model dataization, and expands the data set to make it closer to the complete data set.

[0035] The present invention takes a one-dimensional convolutional neural network as the basic framework, utilizes an incomplete training data set, and obtains a fault diagnosis model under strong non-stationary working conditions through a cross-adversarial training method of standard self-learning and data enhancement, thereby improving the accuracy of fault diagnosis.

[0036] Advantages of additional aspects of the present invention will be given in part in the following description and in part will be obvious from the following description, or will be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] The accompanying drawings, which constitute a part of the present invention, are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute improper limitations on the present invention.

[0038] Figure 1(a)-(b) shows examples of humans trained with an incomplete dataset identifying regular targets and perturbed samples.

[0039] Figure 2 This is a flow chart of the method of the first embodiment.

[0040] Figure 3 This is the architecture diagram of the standard self-learning data enhancement method.

[0041] Figure 4 Training flowchart for the standard self-learning data augmentation method.

[0042] Figure 5 Fault test bench and faulty bearing.

[0043] Figure 6 The speed changes of samples with different health conditions in the TDR dataset.

[0044] Figure 7 Diagnosis results of different test sets.

[0045] Figure 8 This is a system structure diagram of the second embodiment. DETAILED DESCRIPTION

[0046] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0047] It should be noted that the following detailed descriptions are exemplary and intended to provide further explanation of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by those skilled in the art to which the present invention belongs.

[0048] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit the exemplary embodiments according to the present invention. As used herein, unless the context clearly indicates otherwise, the singular form is intended to include the plural form. In addition, it should be understood that when the terms "comprise" and / or "include" are used in this specification, they indicate the presence of features, steps, operations, devices, components and / or combinations thereof.

[0049] Example 1

[0050] Human object recognition is often plagued by incomplete training data sets. As shown in Figure 1, a person who has only seen conventional fish can immediately recognize fish such as "carp" and "grass carp". However, when they see "flying fish", they may hesitate in their mind whether the target is a fish or a bird, and are very likely to make a wrong identification.

[0051] In the above example, a person who has only seen common fish is equivalent to a model trained with an incomplete dataset; fish such as "carp" and "grass carp" can be regarded as samples similar to the training data, which are called common samples in this embodiment; "flying fish" are equivalent to samples that are dissimilar to the training samples (called perturbation samples in this embodiment); it can be seen that the ability of a model trained with an incomplete dataset to recognize perturbation samples is greatly reduced.

[0052] The fault data set collected under strong non-stationary working conditions is a typical incomplete data set, but the model's task is often to perform diagnosis under complex and changeable working conditions, which means that most test samples are perturbation samples; therefore, the purpose of data augmentation is to generate perturbation samples to expand the training set, thereby enhancing the completeness of the training set; the GAN-based method aims to generate regular samples that are similar to the training samples, which is obviously contrary to the purpose of data augmentation under strong non-stationary working conditions; therefore, it is urgent to change the conventional intelligent data augmentation ideas and propose intelligent data augmentation methods for perturbation data.

[0053] To generate perturbed samples, the first issue is clarifying the criteria for perturbed samples—that is, how to evaluate whether a generated sample is a perturbed sample. As can be seen from the above example, the "flying fish" perturbation sample can cause the human brain to err in its prediction results. This means that a sample can only be considered a perturbed sample when it differs sufficiently from the original training set to interfere with the model's judgment. Inspired by this idea, the present invention proposes a data augmentation method that uses the fault diagnosis model's own prediction results as the evaluation criteria for data generation. This method generates samples through sample parameterization and model dataization. Because the fault diagnosis model's own prediction results are learned by the model through training data, it is called Standard Self-Learning Data Augmentation (SSDA).

[0054] This embodiment discloses a fault diagnosis method based on standard self-learning data enhancement, such as Figure 2 As shown, specifically including:

[0055] Step S1: constructing a fault diagnosis model based on a one-dimensional convolutional neural network;

[0056] The basic model architecture of SSDA adopts the widely used one-dimensional convolutional neural network (1-D-CNN). 1-D-CNN consists of multiple layers of convolutional layers, pooling layers, and fully connected layers. The parameter sets of each layer are shown in Table 1:

[0057] Table 1 Layer-by-layer parameters of 1-D-CNN

[0058]

[0059] In order to reduce manual workload, the original measured vibration signal is segmented and directly input into the network without going through signal processing methods such as Fourier transform. The input sample of the constructed model is defined as Where N is the sample dimension, and this study sets N to 1200 dimensions.

[0060] Convolutional layer

[0061] For the lth convolutional layer, its characteristics can be obtained by the following formula:

[0062]

[0063] in, is the convolution kernel, and K l is the length of the convolution kernel, M l-1 is the number of channels of the previous feature layer, M l is the number of channels of the current layer; is the output feature of the previous feature layer, Yes N l-1 ×M l-1 The vector space of N l-1 is the characteristic dimension; b l is the bias vector; f(·) is the activation function. In this embodiment, all convolutional layers use ReLU (Rectified Linear Unit) as the activation function; v l-1 *k l is the convolution operation, which can be calculated by the following formula:

[0064]

[0065] Among them, the subscript [·] represents the sequence number of the elements in the matrix, and the stride of the convolution operation is 1. Therefore, The dimension is (N l-1 -K l +1)×M l .

[0066] Pooling layer

[0067] All convolutional layers in this model are connected to the pooling layer, and the output features of the lth convolutional layer are Perform dimensionality reduction, the output feature v of the pooling layer l for:

[0068]

[0069] in, S is the pooling length.

[0070] Fully connected layer

[0071] The input sample is flattened into a one-dimensional vector u1 after multiple layers of convolution and pooling. Then, fault diagnosis is performed through three layers of full connection. The forward propagation of the full connection is:

[0072] u l =f(w l u l-1 +b l )……(4)

[0073] Among them, w l and b l They are the weight matrix and bias vector of the fully connected layer respectively; the activation function of the first two fully connected layers is the ReLU activation function, and the features of the last fully connected layer are activated by the Softmax function to obtain the output of the model (C represents the number of fault types), that is, the elements in the output o can be calculated as follows:

[0074]

[0075] in, represents w3u3+b3, which is the feature of the output layer without activation function.

[0076] The output o of the model represents the posterior probability that the sample belongs to each fault type. Therefore, the fault type of the sample can be determined based on the model output. For the convenience of description, the process of converting the sample x into the feature u1 after inputting the model is abstracted as the mapping Φ f , the process of transforming feature u1 into output o is abstracted as mapping Φ c , that is, u1=Φ f (x), o = Φ c (u1); all parameters of the model are represented by θ.

[0077] Step S2: The fault diagnosis model is trained by a cross-adversarial training method combining standard self-learning and data enhancement to obtain a complete data set and an intelligent fault diagnosis model under strong non-stationary working conditions;

[0078] SSDA includes two training steps: standard self-learning and data enhancement. In standard self-learning, the parameters of the 1-D-CNN model are optimized by repeatedly inputting updated samples with the goal of learning classification knowledge. This process is equivalent to the model self-learning the evaluation criteria for judging whether a sample is a perturbation sample. In data enhancement, the posterior probability of the model output result is disturbed by the method of sample parameterization and model dataization, thereby generating diversified samples. After alternating between the two training steps, not only a complete training data set can be obtained, but also a fault diagnosis model for strongly non-stationary working conditions can be established. The method architecture and training process of SSDA are as follows: Figure 3 、 4 shown.

[0079] Standard self-learning

[0080] The main purpose of the standard self-learning step is to train a model that can perform fault diagnosis. Since the judgment criterion of the perturbation sample is whether it can interfere with the model judgment, the judgment of the model will be regarded as the evaluation criterion and applied to the data enhancement step.

[0081] Model (1-D-CNN) through training dataset To train, among them, is the number of samples in the data set, x i represents the i-th sample in the dataset, Represents its label, y i is a one-hot vector, and the assignment rule of its elements is:

[0082]

[0083] parameter In the example, r∈[0,1,2,…,R] represents the number of adversarial training cycles, and R is the total number of cycles. From the training data set And the dataset generated by the rth data augmentation Composition, and is the original training dataset.

[0084] In the standard self-learning process, the model is trained with a cross-entropy objective function, which is defined as:

[0085]

[0086] Among them,i =Φ c (Φ f (x i )).

[0087] The model uses the adaptive moment estimation algorithm (Adam) as the optimizer, and the number of back propagation iterations is recorded as T s , the learning rate is ε s By minimizing L s (θ), the model will have the ability to The ability to make a correct diagnosis of samples from

[0088] Data augmentation

[0089] Generating perturbation samples is the goal of data augmentation, and the criterion is whether the generated samples can interfere with the judgment of the model; therefore, guided by the output of the model itself, perturbation samples are generated through sample parameterization and model dataization.

[0090] Sample parameterization, that is, treating samples as model parameters, training the parameters that reduce the objective function through the stochastic gradient descent method, and then deriving the parameters as generated samples; model dataization is treating the parameters θ of the 1-D-CNN model as data, that is, fixing the parameters θ during the training process.

[0091] Therefore, first of all, the dataset Initialize parameters to initial values in, is the last generated sample, and This means that the model further performs data augmentation based on previously generated samples.

[0092] The criterion for perturbing samples is that they will cause perturbations in the posterior probability of the model after being input into the model. Therefore, the first objective function of data enhancement is:

[0093]

[0094] in, and represents the number of samples in the original initialized data set; formula (8) shows that the objective function of data enhancement is antagonistic to the objective function of standard self-learning, so a complete data set and a diagnostic model can be obtained at the same time.

[0095] If the data enhancement process only focuses on perturbations, it is easy to make the posterior probability deviation too large and generate meaningless samples. It is necessary to restrict the sample generation process. Therefore, the second objective function of data enhancement is

[0096]

[0097] in, u 1, =Φ f (x i ), λ>0 is the adjustment coefficient; Participating in optimization means that the data augmentation process limits excessive sample changes but allows reasonable sample diversity.

[0098] The final objective function of the data enhancement process is:

[0099]

[0100] Among them, the parameters Adam is also used as the optimizer, and the number of back propagation iterations is recorded as T g , the learning rate is ε g By minimizing Parameter Dataset will be transformed into a dataset different from its initialization perturbation sample dataset.

[0101] Training strategy

[0102] like Figure 4 As shown in Figure 2, in the standard self-learning data augmentation fault diagnosis method, standard self-learning and data augmentation processes are performed alternately to obtain a complete data set and an intelligent fault diagnosis model under strong non-stationary working conditions. The specific training process is as follows:

[0103] (1) Initialize the dataset And randomly initialize the model parameters θ 0 , set the hyperparameter T s 、T g , ε s , ε g , λ, R, and the additional training times E of the model after the adversarial cycle R m , initialize r=0.

[0104] (2) Based on the training data set Perform standard self-learning until the maximum number of iterations T is reached s , let r = r + 1, and then get the training model parameter θ r .

[0105] (3) Use model dataization and sample parameterization methods to enhance data, and use data sets After T g Iterations generate new perturbation sample data sets

[0106] (4) Merger and Create a new training set

[0107] (5) Determine whether the maximum number of cycles R is reached. If r <, return to step (2). Otherwise, based on the training set Perform standard self-learning until the number of additional training times E is reached m .

[0108] (6) Complete training and obtain a complete data set and has the optimal parameter set θ R+1 A fault diagnosis model for strongly non-stationary working conditions.

[0109] Step S3: input the collected vibration signal to be diagnosed into the trained intelligent fault diagnosis model to obtain the bearing fault diagnosis result.

[0110] The output of the intelligent fault diagnosis model is the posterior probability that the vibration signal to be diagnosed belongs to each fault type. The probabilities are sorted, and the fault type with the highest probability is the final bearing fault diagnosis result.

[0111] Through experiments and analysis of the results, the accuracy of the fault diagnosis method based on standard self-learning data enhancement proposed in the present invention under strong non-stationary working conditions is verified.

[0112] Data Description

[0113] A motor-driven strong non-stationary bearing fault test bench was selected for verification experiment. The test bench and the faulty parts were as follows: Figure 5 As shown in the figure, the test bench consists of a motor, tachometer, coupling, bearing housing, and double-disc rotor. The target fault bearing is the end bearing, model NU205EM, with an accelerometer (PCB315A) placed on the end bearing housing. Three single faults are pre-set for the bearing: inner race fault (IF), rolling element fault (RF), and outer race fault (OF), as well as one combined fault: outer race and rolling element combined fault (ORF). The motor speed range is 0 to 1500 rpm, and vibration signals are collected using an LMS data acquisition system at a sampling frequency of 12.8 kHz.

[0114] To verify the effectiveness of the proposed method, the data contains the following three types of working conditions.

[0115] (1) Uniform deceleration condition: The motor is uniformly decelerated from 1500 rpm to a standstill. This process simulates the incomplete data set collected when a fault occurs during actual operation and serves as the training data for this method.

[0116] (2) Strong non-stationary working condition: This working condition simulates the strong non-stationary working condition of the equipment in actual operation. The speed change is as follows: Figure 6 As shown in FIG, the test data used to verify this method is expressed in TDR.

[0117] (3) Constant speed condition: The speed change rate of the constant speed sample is 0. Compared with the strong non-stationary condition, the difference between it and the training sample is greater. It can be considered that all the samples of the constant speed condition are disturbance samples. The experiment collected data at speeds of 800 rpm, 1000 rpm and 1500 rpm (represented by TD1, TD2 and TD3 respectively) to test the validity of the generated data.

[0118] Experimental results analysis

[0119] Model undetermined parameter T s 、T g , ε s , ε g , λ, R, and E m The values ​​are preset to 100, 100, 0.01, 1, 1, 10, and 2000 respectively. After the model is trained using the incomplete training data set, it is tested using the TD1, TD2, TD3, and TDR data sets. In order to verify the effectiveness of the proposed method, a 1-D-CNN model with the same structure as the 1-D-CNN of this method is used. Only the training samples are used for training and the test data are diagnosed. For comparison, the results are shown in the figure below. Figure 7 shown.

[0120] Figure 7 It can be seen that the diagnostic accuracy of the two methods for diagnosing the TD1, TD2 and TD3 data sets is significantly lower than the accuracy for diagnosing TDR. This is because the constant speed data set has more disturbance samples than the training data set. Although the model structures of 1-D-CNN and SSDA in the fault diagnosis process are exactly the same, there is a significant gap in the diagnostic results of the two methods. The accuracy of 1-D-CNN is less than 90% when diagnosing the constant speed data set, and the accuracy is only 91.67% to 92.54% when diagnosing the strongly non-stationary working condition data set. Compared with 1-D-CNN, the SSDA method proposed in this invention has an accuracy improvement of more than 10% when diagnosing the constant speed data set, and the accuracy of TDR is improved to 98.55% to 99.07%. This shows that the proposed method can generate disturbance samples to expand the data set to make it closer to the complete data set.

[0121] Example 2

[0122] This embodiment discloses a fault diagnosis system based on standard self-learning data enhancement;

[0123] like Figure 8 As shown, a fault diagnosis system based on standard self-learning data enhancement includes a model building module, a model training module and a fault diagnosis module;

[0124] The model building module is configured to: build a fault diagnosis model based on a one-dimensional convolutional neural network;

[0125] The model training module is configured to train the fault diagnosis model through a cross-adversarial training method combining standard self-learning and data enhancement to obtain an intelligent fault diagnosis model with a complete data set and under strong non-stationary working conditions.

[0126] The fault diagnosis module is configured to: input the collected vibration signal to be diagnosed into the trained intelligent fault diagnosis model to obtain the bearing fault diagnosis result.

[0127] Example 3

[0128] The purpose of this embodiment is to provide a computer-readable storage medium.

[0129] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of a fault diagnosis method based on standard self-learning data enhancement as described in Example 1 of the present disclosure.

[0130] Example 4

[0131] The purpose of this embodiment is to provide an electronic device.

[0132] An electronic device includes a memory, a processor, and a program stored in the memory and executable on the processor. When the processor executes the program, the steps of a fault diagnosis method based on standard self-learning data enhancement as described in Example 1 of the present disclosure are implemented.

[0133] The foregoing description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Those skilled in the art will readily appreciate that various modifications and variations of the present invention are possible. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention are intended to be within the scope of protection of the present invention.

Claims

1. A fault diagnosis method based on standard self-learning data enhancement, characterized in that: include: Based on one-dimensional convolutional neural network, a fault diagnosis model is constructed; The fault diagnosis model is trained through standard self-learning and data enhancement cross-adversarial training. The specific training process is as follows: Initialize the data set , and randomly initialize the model parameters , set the hyperparameters 、 、 、 、 、 and fighting cycles Additional training times for the model after ,initialization ;Based on the training data set Perform standard self-learning until the maximum number of iterations is reached ,make , and then get the training model parameters ;Use model data and sample parameterization methods to enhance data, and use data sets ,go through Iterations generate new perturbation sample data sets ;merge and Create a new training set ; Determine whether the maximum number of cycles has been reached If not reached, return to standard self-learning, otherwise, based on the training set Perform standard self-learning until the number of additional training times is reached , get a complete data set and has the optimal parameter set Intelligent fault diagnosis model for strong non-stationary working conditions; The collected vibration signal to be diagnosed is input into the trained intelligent fault diagnosis model to obtain the bearing fault diagnosis result; Sample parameterization is to regard samples as model parameters, train the parameters that reduce the objective function through stochastic gradient descent, and then derive the parameters into generated samples; Model digitization is to treat the parameters of the fault diagnosis model as data and fix the parameter values ​​during the training process.

2. A fault diagnosis method based on standard self-learning data enhancement according to claim 1, characterized in that: The one-dimensional convolutional neural network includes multiple convolutional layers, pooling layers and fully connected layers; The convolution layer uses ReLU (Rectified Linear Unit) as the activation function, and the stride of the convolution operation is 1; All convolutional layers are connected to pooling layers to reduce the dimensionality of the output features of the convolutional layers; The features of the input samples after multi-layer convolution and pooling are flattened into a one-dimensional vector, and then fault diagnosis is performed through three layers of full connection.

3. A fault diagnosis method based on standard self-learning data enhancement as claimed in claim 1, characterized in that: The standard self-learning aims to learn classification knowledge, optimizes parameters in the fault diagnosis model by repeatedly inputting updated samples, and self-learns the evaluation criteria for determining whether a sample is a disturbance sample.

4. A fault diagnosis method based on standard self-learning data enhancement as claimed in claim 1, characterized in that: The data enhancement is guided by the output of the model itself and generates perturbation samples through sample parameterization and model dataization. Among them, the judgment criterion for the perturbation sample is whether it can interfere with the model judgment, specifically: after the sample is input into the model, it can cause a disturbance in the posterior probability of the model.

5. A fault diagnosis method based on standard self-learning data enhancement as claimed in claim 1, characterized in that: The output of the intelligent fault diagnosis model is the posterior probability that the vibration signal to be diagnosed belongs to each fault type. The probabilities are sorted, and the fault type with the highest probability is the final bearing fault diagnosis result.

6. A fault diagnosis system based on standard self-learning data enhancement, characterized in that: Includes model building module, model training module and fault diagnosis module; The model building module is configured to: build a fault diagnosis model based on a one-dimensional convolutional neural network; The model training module is configured to train the fault diagnosis model through a cross-adversarial training method of standard self-learning and data enhancement. The specific training process is as follows: Initialize the data set , and randomly initialize the model parameters , set the hyperparameters 、 、 、 、 、 and fighting cycles Additional training times for the model after ,initialization ;Based on the training data set Perform standard self-learning until the maximum number of iterations is reached ,make , and then get the training model parameters ;Use model data and sample parameterization methods to enhance data, and use data sets ,go through Iterations generate new perturbation sample data sets ;merge and Create a new training set ; Determine whether the maximum number of cycles has been reached If not reached, return to standard self-learning, otherwise, based on the training set Perform standard self-learning until the number of additional training times is reached , get a complete data set and has the optimal parameter set Intelligent fault diagnosis model for strong non-stationary working conditions; The fault diagnosis module is configured to: input the collected vibration signal to be diagnosed into the trained intelligent fault diagnosis model to obtain the bearing fault diagnosis result; Sample parameterization is to regard samples as model parameters, train the parameters that reduce the objective function through stochastic gradient descent, and then derive the parameters into generated samples; Model digitization is to treat the parameters of the fault diagnosis model as data and fix the parameter values ​​during the training process.

7. A computer-readable storage medium having a program stored thereon, characterized in that: When the program is executed by a processor, the steps of the fault diagnosis method based on standard self-learning data enhancement as described in any one of claims 1 to 5 are implemented.

8. An electronic device comprising a memory, a processor, and a program stored in the memory and executable on the processor, wherein: When the processor executes the program, the steps of the fault diagnosis method based on standard self-learning data enhancement as described in any one of claims 1 to 5 are implemented.

Images