Mechanical equipment fault diagnosis method and device based on integrated stack auto-encoder

Abstract

The invention discloses a mechanical equipment fault diagnosis method and device based on an integrated stack auto-encoder, and relates to the technical field of integrated learning. According to the method, the sample marginal theory is introduced into the UWA diversity index, the sample marginal theory and the UWA diversity index are combined to construct a new U-M diversity measurement index, and the classification condition of edge samples by a classifier is focused on; according to the invention, part of parameters in U-M measurement are optimized by using an improved particle swarm algorithm, and the invention further provides a sorting integration pruning method based on U-M indexes, which comprises the following steps: screening a classifier subset with both diversity and accuracy from an original classifier pool; in addition, the screened classifiers are combined by adopting a certain ensemble learning strategy (such as a majority voting method, a weighted voting method and the like) to obtain a final ensemble fault diagnosis result. According to the method, the accuracy of integrated learning classification can be improved, and the fault diagnosis requirements of the mechanical equipment can be met more comprehensively.

Description

Technical Field

[0001] The present invention belongs to the technical field of integrated learning, and in particular relates to a mechanical equipment fault diagnosis method and device based on an integrated stack autoencoder. Background Art

[0002] In actual industrial environments, mechanical equipment often operates under extremely harsh conditions of high temperature, high pressure, and high operating frequency for a long time, which seriously threatens the health and service life of the equipment. Once the equipment is abnormal, it will reduce industrial production efficiency and increase the maintenance and operation costs of the enterprise, or even cause serious production failures and threaten the life, health and safety of workers. Therefore, fault diagnosis of mechanical equipment is of great significance.

[0003] Fault diagnosis technology originated in the last century, and has successively produced fault diagnosis technology based on statistical analysis, fault diagnosis technology based on signal processing, and fault diagnosis technology based on data-driven. Research on fault diagnosis of mechanical equipment has always been the focus and hotspot of research. Vibration signals, as the most easily obtained data during the operation of mechanical equipment, often contain rich characteristics and can reflect the operating status and remaining life of the equipment. Therefore, they are widely used in mechanical fault diagnosis.

[0004] There are many mechanical fault diagnosis methods based on vibration signals. Among them, the fault diagnosis technology based on statistical analysis is difficult to handle nonlinear problems, and the fault diagnosis technology based on signal processing is too dependent on system models and requires certain professional knowledge. Therefore, both methods are not suitable for processing complex, changeable, and large numbers of nonlinear mechanical vibration data. The data-driven method can realize mechanical fault diagnosis without mathematical models, and has strong adaptability and can effectively process high-dimensional data, so it has been the most widely used. This method can be divided into two categories: machine learning methods and deep learning methods. Among them, machine learning is a typical pattern recognition method that can automatically discover pattern rules in data. However, as a shallow network, the learning ability of machine learning algorithms is limited. It is often necessary to extract fault features through signal decomposition methods first, and it relies on previous knowledge and experience. With the increasing intelligence level of mechanical equipment, the operating status data generated by the equipment will gradually increase, which makes it difficult for shallow machine learning models to learn highly abstract features of the data. The deep learning method with multiple hidden layers relies on its powerful feature extraction ability to extract the hidden feature information of mechanical equipment from massive data, so it has developed rapidly in recent years.

[0005] The deep learning method is a multi-layer deep neural network that can process high-dimensional redundant nonlinear data. The biggest advantage of the deep learning method is its powerful feature extraction capability. It can automatically learn the time domain features, frequency domain features, and time-frequency domain features in mechanical vibration data, and does not rely on domain experience and expert knowledge. It does not require manual feature extraction, which greatly improves the efficiency of mechanical fault diagnosis. The deep learning method does not require an in-depth understanding of the internal structure and complex mechanisms of the system, can process large-scale data sets, has strong adaptive capabilities, is suitable for most complex working conditions, and can be used in the field of mechanical fault diagnosis.

[0006] Ensemble learning is a general term for model fusion strategies, and its idea is to combine multiple basic classifiers using certain strategies to obtain decision results that are better than those of a single model. Ensemble learning methods are widely used in image classification, fault diagnosis, semantic analysis and other fields. In ensemble learning, it is first necessary to consider which machine learning or deep learning model should be selected as the base classifier. Among all algorithm models, autoencoder (AE), as an unsupervised nonlinear neural network, has strong feature extraction and data reconstruction capabilities, and can effectively process nonlinear and non-stationary mechanical vibration signals. In addition, the autoencoder has a simple principle and a large number of variants, which meets the requirements of large-scale integration and is suitable for ensemble learning. Therefore, this invention uses autoencoders to carry out related research in ensemble learning.

[0007] The autoencoder is a three-layer neural network, including an input layer, a hidden layer and an output layer. Its network structure is as follows: Figure 1 As shown. The autoencoder consists of two parts: an encoder and a decoder. The input data is first encoded into abstract features by the encoder, and then the decoder restores these feature representations to the original form of the data. That is, the encoder performs feature mapping from the input layer to the hidden layer, and the decoder performs data self-reconstruction from the hidden layer to the output layer. Assume that the number of nodes in the input layer and output layer of the autoencoder is m, and the number of nodes in the hidden layer is n. The input vector matrix X is transferred from the input layer to the hidden layer, and the hidden layer matrix h is calculated as: h = f(WX+b) (1)

[0008] Where X is the input vector matrix, W is the encoder weight matrix, b is the bias vector, and f is the activation function. The reconstructed vector is decoded by the decoder Calculated as:

[0009] In the formula, is the reconstructed input vector matrix, is the decoder weight matrix, is the reconstruction bias vector, and g is the activation function. For a given input data sequence, the model parameters of AE can be calculated by minimizing the reconstruction error J:

[0010]

[0011] The AE network (referring to the autoencoder network) only uses a single-layer encoder for feature processing and extraction. However, it is difficult for a single-layer neural network to deeply extract more effective features during training. The stacked autoencoder (SAE) is a deep autoencoder containing multiple hidden layers. It is constructed by stacking multiple autoencoders layer by layer. The network structure of SAE is as follows: Figure 2 As shown in the figure. SAE uses the hidden layer of the previous level autoencoder as the input layer of the next level autoencoder, and trains step by step until the network is built. The layer-by-layer pre-training method can help the model extract deeper features from the data. After the pre-training is completed, the SAE model is fine-tuned in a supervised manner using a labeled dataset. A Softmax fully connected layer is added as the output layer of the model and the model parameters are fine-tuned to form a SAE classifier that can perform classification tasks.

[0012] Although ensemble learning can improve the generalization ability of a single model in fault diagnosis, if the number of ensemble models is too large, the computational cost may be greatly increased while the improvement of the ensemble effect is not obvious. In order to effectively solve these problems, ensemble pruning came into being. Ensemble pruning, also known as selective ensemble, ensemble screening or ensemble refinement, refers to the selective deletion of classifiers in ensemble learning to reduce the complexity of the ensemble. For the subset of classifiers obtained by ensemble pruning, its ensemble performance is no less than that of the original classifier set.

[0013] In ensemble pruning, the first consideration is how to measure the importance of each classifier. Among them, classification accuracy is the most intuitive and easy to understand, and can reflect the classification performance of a single classifier. However, accuracy cannot reflect the impact of a single classifier on the classifier set, which is relatively one-sided and has obvious limitations in ensemble pruning. Diversity represents the differences between classifiers, and the size of diversity will directly affect the ensemble classification effect. Therefore, ensemble pruning often uses diversity indicators to measure the importance of each classifier.

[0014] How to quantitatively analyze the diversity between different classifiers has always been a hot topic and difficulty in the field of ensemble learning. According to different calculation methods, these metrics can be divided into two categories: paired and unpaired. Among them, the paired diversity metric first calculates the diversity value between each two classifiers, and then takes the average of all diversity values ​​as the diversity value of the classifier set, including indicators such as Q statistics, correlation coefficient ρ, inconsistency metric and double failure metric; unpaired diversity metric quantifies the diversity of the classifier set by direct overall calculation, which is simpler, including Kohavi-Wolpert variance, κ metric, difficulty metric and other indicators. Although these indicators can measure the diversity of the classifier set, they do not reflect the importance of a single model decision to the classifier as a whole, lack theoretical support, and are not conducive to ensemble pruning operations. As a new diversity metric, UWA (Uncertainty Weighted Accuracy) focuses on both the decision of the current classifier and the decision of the overall set, and focuses on samples with huge classification differences in the set, so it is suitable for ensemble pruning.

[0015] The relevant theories of UWA are introduced as follows:

[0016] Let H = [{h 1 ,h 2 ,...,h L} is the total set of classifiers, and the sample set is X = [{x 1 ,x 2 ,...,x N}, h is a single classifier; L is the number of candidate classifiers, and N is the number of samples. Use the sample set X to classify faults for all classifiers in H. The classification result of H for the sample is obtained by voting by L classifiers. For any sample x in X, i For example, the classifier set H and the single classifier h have four different classification decision results:

[0017] e 00 (h,H,x i ,y i ):h(x i )≠y i &H(x i )≠y i ;

[0018] e 01 (h,H,x i ,y i ):h(x i )≠y i &H(x i )=y i ;

[0019] e10 (h,H,x i ,y i ):h(x i )=y i &H(x i )≠y i ;

[0020] e 11 (h,H,x i ,y i ):h(x i )=y i &H(x i )=y i ; (4)

[0021] Then the diversity metric UWA is calculated as:

[0022]

[0023] Where NT i Represents the sample (x i ,y i ) gives the proportion of classifiers that make correct classification decisions. The classification results of the classifier set H for the sample are obtained by voting by all classifiers in H. NF i =1-NT i Represents the sample (x i ,y i ) gives the proportion of classifiers with incorrect classification, I is the indicator function, and I(true) = 1, I(false) = 0. The construction mechanism of the UWA metric is as follows:

[0024] 1) For event e 00 For example, at this time, both h and H are positive to x i Classification error: When an error occurs in the ensemble, if the misclassified classifier is sent into the ensemble, the error of the ensemble will be further aggravated. Therefore, a negative value needs to be given to punish its impact on other correct classification models. The penalty ratio is proportional to the number of correctly classified classifiers in H.

[0025] 2) For event e 01 For example, at this time, h is i Classification error, H vs. x i The classification is correct; at this time, the integration result is correct. If a classifier with incorrect classification is added, it may lead to the reversal of the integration result. Therefore, a negative value should also be given for punishment, and the penalty ratio is proportional to the number of incorrectly classified classifiers in H.

[0026] 3) For event e 10 For example, at this time, h is i Correct classification, H for xi Classification error; at this time, the integration result is wrong. If a classifier with accurate classification is added, it is possible that the integration result will become correct. Therefore, a positive value needs to be given as a reward, and the reward ratio is proportional to the number of classifiers with correct classification in H.

[0027] 4) For event e 11 For example, at this time, both h and H are positive to x i The classification is correct; if a classifier with accurate classification is added at this time, the classification accuracy of the ensemble will be further enhanced, making the ensemble decision less likely to change. Therefore, a positive value should be given as a reward, and the reward ratio is proportional to the number of misclassified classifiers in H, in order to eliminate the negative impact of these classifiers on the ensemble and improve the accuracy of equipment fault diagnosis.

[0028] UWA focuses on the differences between a single classifier and a classifier set and quantifies them. The higher the UWA value, the higher the diversity between a single classifier and a classifier set.

[0029] The ensemble pruning method is one of the key technologies in ensemble learning. By selectively deleting classifiers, it can effectively reduce computing costs and improve operating efficiency. At the same time, it can optimize the ensemble structure and reduce the risk of overfitting. With the continuous deepening of research on ensemble pruning methods, the importance of ensemble pruning in ensemble learning has gradually become more prominent, and it has also been widely used in the field of fault diagnosis based on ensemble deep learning.

[0030] Sample Margin is a concept in the field of ensemble learning, which is mainly used to describe the importance of some samples with highly inconsistent classification results in the ensemble. The classification results of these samples often have a great impact on the overall effect of the ensemble. For multi-classification problems, let the total set of classifiers be H = [{h 1 ,h 2 ,...,h L}, L is the number of classifiers, N is the number of samples, and the sample set is X = [{x 1 ,x 2 ,...,x N}, use the majority voting method to integrate the results of all classifiers in H, then each sample x i The marginal value of margin(x i ) is defined as:

[0031] where y i For x i The real category, is the sample x i Remove y from all classifiers i The categories with the most votes, and The classification result in H is y i and The number of classifiers, and margin(x i )∈[-1,1].

[0032] When margin(x i )<0, it means that the sample is in the total x of the classifier i The label with the most votes in set H is not the true label. In this case, H is not a true label for sample x. i Classification error; when margin(x i )>0, indicating that the total set of classifiers H is effective for sample x i The classification is correct; when margin(x i )=0, it means that the sample is just on the margin and it is impossible to judge whether H has any effect on the sample x. i Generally speaking, we take the marginal value margin(x i ) Samples x close to 1 or -1 i These samples are called “hard” samples. The classification results of these samples in all classifiers are highly consistent, and their integrated classification results are not easily affected by a single classifier. i ) is close to 0 is called "marginal" samples. The classification results of these samples in all classifiers are highly inconsistent, and their ensemble classification results are easily affected by a single classifier. In ensemble learning, compared with "hard" samples, the classification of "marginal" samples often has a greater impact on the overall effect of the ensemble and needs to be given more attention.

[0033] Particle Swarm Optimization (PSO) is a swarm intelligence optimization algorithm that simulates bird behavior. Its characteristic is that it can search as many areas as possible in the solution space of the objective function. In the swarm intelligence algorithm, each individual with experience and wisdom interacts with each other, using the powerful group wisdom to solve complex optimization problems. The idea of ​​the particle swarm optimization algorithm is derived from the foraging behavior of bird flocks flying in the air. Individuals and groups cooperate with each other and share the collected information to find the optimal solution. The particle swarm algorithm is easy to implement, efficient, intuitive and easy to understand. It is widely used in fields such as function optimization, neural network training and fuzzy system control.

[0034] In the PSO algorithm, individuals in the population are called particles, which are defined as potential solutions to the problem to be optimized in the search space. Each particle adjusts its moving direction in a timely manner according to its own moving speed and flight experience. Assuming that there are N particles in the population in the search space, and the number of current iterations is t, the position and speed of the i-th particle can be expressed as x and i (t) = (xi1 (t),x i2 (t),…,x iD (t)) T and v i (t)=(v i1 (t),v i2 (t),…,v iD (t)) T , the optimal position of the i-th particle during the movement is pbest i (t) = (pbest i1 (t),pbest i2 (t),…,pbest iD (t)) T , the historical optimal position of all particles in the entire population is gbest(t)=(gbest 1 (t),gbest 2 (t),…,gbest D (t)) T , then after one iteration, the particle's velocity and position are updated as follows:

[0035] v i (t+1)=ωv i (t)+c 1 r 1 (pbest i (t)-x i (t))+c 2 r 2 (gbest i (t)-x i (t)) (7)

[0036] x i (t+1)=x i (t)+v i (t+1) (8)

[0037] Where ω is the inertia weight, which is responsible for adjusting the influence of the previous speed on the current speed; c 1 and c 2 is the learning factor, used to adjust the learning step size; r 1 and r 2 is a random number in the interval [0,1], which is used to increase the randomness of the algorithm. In the optimization process, each particle will consider the current speed, its own historical optimal position and the global optimal position of the population. These three conditions interact with each other and jointly determine the next movement state of each particle.

[0038] The standard PSO algorithm is efficient, simple and efficient, but the inertia weight coefficient ω and learning factor c of the algorithm 1、c 2 The fixed value of ω and c results in weak local search ability of particles, which is prone to early convergence and local optimal problems. To address these problems, the Improved Particle Swarm Optimization (IPSO) algorithm was born. The algorithm can dynamically adjust the parameters during the iteration process so that it can quickly optimize and find the optimal position in the global range. The improved ω and c 1 、c 2 for:

[0039]

[0040] In the formula, t max is the maximum number of iterations, c 1_max and c 1_min c 1 The maximum and minimum values ​​of c 2_max and c 2_min c 2 The maximum and minimum values ​​of ω max and ω min is the maximum and minimum value of the weight coefficient, and t is the current iteration number. 1 、c 2 Substituting the expression into equation (8), we can get the particle velocity update formula of the IPSO algorithm. Compared with the traditional PSO algorithm, the IPSO algorithm helps the system to slowly adjust the inertia weight ω in a linear decreasing manner, making the search process more reasonable, balancing the local search ability and global search ability of the algorithm, and making the algorithm's optimization ability better.

[0041] In the process of ensemble pruning, the existing methods often only refer to the model accuracy and ignore the diversity between models, which may lead to poor ensemble fault diagnosis effect of classifiers. In addition, in ensemble learning, the classification results of "marginal" samples in all classifiers are highly inconsistent. Therefore, compared with "hard" samples, the classification of "marginal" samples often has a greater impact on the overall effect of the ensemble and needs to be given more attention. However, the existing ensemble pruning methods often ignore such samples in the process of classifier selection, lack of attention to the classifiers that correctly classify these "marginal" samples, and have great limitations.

[0042] In the field of mechanical fault diagnosis, due to the nonlinear and complex characteristics of equipment failures, the use of a single fault diagnosis algorithm is difficult to adapt to various types of equipment failures and has certain limitations. Summary of the invention

[0043] In order to solve the above problems, the present invention proposes a mechanical equipment fault diagnosis method and device based on an integrated stack autoencoder, which uses integrated learning with model diversity to diagnose mechanical equipment faults, can improve the classification accuracy of "marginal" samples, and improve the classification accuracy of integrated learning of the entire sample, which can more comprehensively meet the fault diagnosis needs of complex and nonlinear mechanical equipment.

[0044] In a first aspect, the present invention provides a mechanical equipment fault diagnosis method based on an integrated stack autoencoder, comprising:

[0045] Obtaining fault data of target mechanical equipment;

[0046] Inputting the fault data into a trained and optimized fault diagnosis model to perform fault diagnosis and obtain a fault diagnosis result;

[0047] The fault diagnosis model is constructed based on several different SAE base classifiers screened by the UM diversity measurement method, and the UM diversity measurement method is a method of introducing the sample margin theory into the UWA index.

[0048] In a second aspect, the present invention provides a mechanical equipment fault diagnosis device based on an integrated stack autoencoder, comprising:

[0049] A data acquisition module, used to acquire fault data of target mechanical equipment;

[0050] A fault diagnosis module is used to input the fault data into a trained and optimized fault diagnosis model to perform fault diagnosis and obtain a fault diagnosis result;

[0051] The fault diagnosis model is constructed based on several SAE base classifiers screened by a UM diversity metric module, and the UM diversity metric module is constructed by introducing sample margin theory into the UWA index.

[0052] The beneficial effects of the technical solution provided by the embodiment of the present invention are as follows: the mechanical equipment fault diagnosis method and device based on the integrated stack autoencoder of the present invention, since the fault diagnosis model used is constructed based on several different SAE base classifiers screened by the UM diversity measurement method, the fault diagnosis model is constructed based on ensemble learning, and the several SAE base classifiers required for the construction are diverse; the present invention innovatively combines the ideas of ensemble learning and ensemble pruning with the deep learning model, fully utilizes the diversity between classifiers and combines the characteristics of multiple models, effectively improving the universality and robustness of the fault diagnosis algorithm; and since the UM diversity measurement method is a method of introducing sample marginal theory into the UWA indicator, it can effectively process "marginal" samples and minimize the impact of redundant classifiers on the integration, so the present invention can be applied to small sample fault diagnosis problems. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Figure 1 is the autoencoder network structure;

[0054] Figure 2 It is the SAE network structure;

[0055] Figure 3 A schematic diagram of a flow chart of a mechanical equipment fault diagnosis method based on an integrated stack autoencoder provided in an embodiment of the present invention;

[0056] Figure 4 Another schematic flow chart of a mechanical equipment fault diagnosis method based on an integrated stack autoencoder provided in an embodiment of the present invention;

[0057] Figure 5 A standard margin function image provided by an embodiment of the present invention;

[0058] Figure 6 The marginal margin function images corresponding to different μ values ​​provided by the embodiment of the present invention;

[0059] Figure 7 Margin function images corresponding to different γ values ​​provided in the embodiment of the present invention

[0060] Figure 8 The marginal margin function images corresponding to different ε values ​​provided in the embodiment of the present invention;

[0061] Fig. 9 An integrated pruning process based on IPSO-UM provided in an embodiment of the present invention;

[0062] Fig.10 The Case Western Reserve University rolling bearing experimental device provided for the embodiment of the present invention;

[0063] Fig.11The IPSO and PSO algorithm iteration process provided by the embodiment of the present invention;

[0064] Fig.12 The voting classification accuracy rates corresponding to the three integrated pruning methods provided in the embodiments of the present invention are intended to compare the classifier voting classification accuracy rate corresponding to the integrated pruning method based on the UM indicator with the accuracy rates of the other two methods;

[0065] Fig.13 A system schematic diagram of a mechanical equipment fault diagnosis device based on an integrated stack autoencoder provided in an embodiment of the present invention.

[0066] In the figure: 1 - induction motor; 2 - accelerometer; 3 - test bearing; 4 - loading motor; 10 - data acquisition module; 20 - fault diagnosis module. DETAILED DESCRIPTION

[0067] The present invention will be described in detail below with reference to the accompanying drawings and in conjunction with embodiments.

[0068] See also Figure 3 , a mechanical equipment fault diagnosis method based on an integrated stack autoencoder provided by an embodiment of the present invention includes the following steps S100 to S200.

[0069] Step S100, acquiring fault data of a target mechanical device.

[0070] Step S200, input the fault data into the trained and optimized fault diagnosis model to perform fault diagnosis and obtain the fault diagnosis result. The fault diagnosis model is constructed based on several different SAE base classifiers screened by the UM diversity measurement method, and the UM diversity measurement method is a method constructed by introducing the sample margin theory into the UWA indicator.

[0071] In some possible embodiments, the method further includes displaying the fault diagnosis result of the target mechanical equipment. The fault diagnosis model is trained and optimized according to the following steps (1-1) to (1-6).

[0072] Step (1-1), obtain a training sample set and construct multiple SAE initial classifiers.

[0073] Step (1-2), using the training sample set to train multiple SAE initial classifiers to obtain a set of trained SAE base classifiers.

[0074] Step (1-3), construct UM diversity measurement indicators.

[0075] Step (1-4), use the UM diversity metric to perform integrated pruning on the trained SAE base classifier set to obtain a SAE base classifier subset that takes both diversity and accuracy into account.

[0076] Step (1-5), integrate all SAE base classifiers in the SAE base classifier subset to obtain an integrated SAE classifier.

[0077] Step (1-6), construct a fault diagnosis model based on the integrated SAE classifier.

[0078] This embodiment uses the training sample set and the test sample set to train and optimize multiple SAE initial classifiers to obtain the final fault diagnosis model. This embodiment uses the bearing vibration data set provided by the Case Western Reserve University (CWRU) Bearing Data Center as a reference benchmark to discuss the process of establishing the sample set: CWRU rolling bearing experimental device such as Fig.10 As shown, the experimental device includes an induction motor 1, an accelerometer 2, a test bearing 3, a loading motor 4 and other components. Fig.10 As shown. According to the sampling frequency and the location of the bearing fault, these data sets can be divided into four categories: 48kHz fault-free benchmark data, 12kHz drive end fault data, 48kHz drive end fault data and 12kHz fan end fault data. Each set of fault data contains three fault types: rolling element, inner ring and outer ring fault. Each bearing is tested under four different loads, and the bearing introduces a single point fault. The fault diameter is 0.007, 0.014 and 0.021 inches, representing the three fault severity levels of mild, moderate and severe. Among them, the drive end sampling frequency is 12kHz and 48kHz, and the fan end bearing sampling frequency is 12kHz. Under each fault, the load is divided into 0 horsepower, 1 horsepower, 2 horsepower and 3 horsepower. This embodiment uses the vibration data at 1797rpm (0 horsepower) for fault classification and selects ten fault types, including normal state, slight inner ring fault, moderate inner ring fault, severe inner ring fault, slight outer ring fault, moderate outer ring fault, severe outer ring fault, slight ball fault, moderate ball fault and severe ball fault. Each fault mode contains 300 samples, and each sample is a vibration signal segment consisting of 400 sampling data points. The samples are divided into training sample set and test sample set according to the ratio of 2:1. The training sample set and test sample set contain 2000 and 1000 samples respectively. The detailed description information of these 10 faults is shown in Table 1.

[0079] Table 1 Description of Case Western Reserve bearing data set

[0080]

[0081]

[0082] In ensemble learning based on stacked autoencoder (SAE) models, the first issue to consider is how to construct a certain number of SAE initial classifiers. Currently, there are three ways to construct SAE initial classifiers: construct different SAE initial classifiers with different activation functions; construct different SAE initial classifiers with different training samples; and construct different SAE initial classifiers with different autoencoder variant models. The specific construction methods of these SAE models are as follows (1)-(3).

[0083] (1) Different activation functions.

[0084] The activation function (AF) is applied to the encoding and decoding process of the autoencoder, which has an important impact on the learning ability and training efficiency of the model. The activation function introduces nonlinear characteristics into the neural network, enabling it to learn complex nonlinear abstract features in the data and enhance the feature expression ability of the deep neural network. Different activation functions have different nonlinear properties. For the SAE model, there are 10 most commonly used activation functions. The specific formulas of these functions are shown in Table 2.

[0085] Table 2 Summary of commonly used activation functions for SAE

[0086]

[0087] These activation functions can be divided into two types: activation functions based on exponential operations and non-exponential operations. Among them, activation functions based on exponential operations (such as Sigmoid, Tanh, Selu, etc.) are the most commonly used activation functions in deep learning. Their function images are smooth and monotonous, and their output range is usually limited. The computational cost is low, which helps to maintain the stability of training and can learn more complex and abstract features. However, the main disadvantages are high computational cost and possible gradient vanishing problems. Activation functions based on non-exponential operations (such as ReLu, Leak_relu, etc.) have fast calculation speed, low computational cost, and can avoid gradient vanishing, but their output is not centered on zero, and they are difficult to train and have poor stability. In short, these two activation functions have different nonlinear mapping characteristics. In ensemble learning, these two activation functions can be used simultaneously to construct different SAE models.

[0088] (2) Different training samples.

[0089] In traditional ensemble learning methods, the Bagging algorithm performs multiple Bootstrap resampling on the training data set to obtain a certain number of different training sample sets. Each sample set trains a separate initial classifier (i.e., base classifier), and then integrates their predictions to obtain the final result. The Bagging algorithm is based on the Bootstrap sampling method and does not require excessive parameter adjustment. It has a clear principle and is simple to implement. It can effectively reduce the classification variance between models and reduce the impact of a single unstable model on the ensemble. Therefore, this embodiment can also use this method to construct a certain number of SAE models. Compared with other sampling methods, the Bootstrap sampling method can generate a large number of resampled sample sets, which helps to alleviate the problem of insufficient sample quantity. It is simple, practical, and has a wide range of applications.

[0090] (3) Different variant models.

[0091] At present, most studies will optimize and improve the autoencoder model, and build more effective SAE variant models by adding relevant constraints or modifying the network structure. The feature extraction capabilities of these variant models are often better than those of ordinary SAE networks. Common SAE variant models include stacked sparse autoencoders (SSAE), stacked contracted autoencoders (SCAE), and stacked denoising autoencoders (SDAE). These models are as follows ①-③.

[0092] ① Stacked sparse autoencoder.

[0093] The Stacked Sparse Auto Encoder (SSAE) adds a KL divergence (Kullback-Leibler divergence) regularization sparse term to the loss function, which makes the neural network sparse and enables the model to learn more robust features with higher discriminative power. The loss function of the SSAE network is:

[0094]

[0095] In the formula, β is the KL regularization coefficient, which is used to control the sparse effect; ρ is the sparse coefficient, which represents the average activation of the hidden layer nodes, and is usually a small value; is the nth feature representation of the ith sample, is the average activation. By minimizing the above loss function, SSAE can extract features with stronger sparsity.

[0096] ② Stacked shrinking autoencoder.

[0097] The Stacked Contractive Auto Encoder (SCAE) achieves the effect of local spatial contraction by adding a contraction regularization term to the objective function of the traditional autoencoder. Its purpose is to suppress the perturbation of the training samples in all directions and improve the robustness of the model to small perturbations in the input samples. The regularization term is usually the Frobenius norm of the Jacobian matrix of the encoder parameters. The loss function of SCAE is calculated as:

[0098]

[0099] Where δ is a hyperparameter that controls the strength of the penalty term; J f (x i ) is the Jacobian matrix of the feature representation relative to the input; Represents the Frobenius norm of the Jacobian matrix. The shrinkage regularization term can make the function have a smaller input gradient and is usually used to measure the local sensitivity in the input space. Therefore, compared with other autoencoders, the shrinkage autoencoder can learn a more stable data representation, which helps to avoid overfitting and reduce the negative impact of disturbances on the model.

[0100] ③Stacked denoising autoencoder.

[0101] The Stacked Denoise Auto Encoder (SDAE) performs random zeroing (random inactivation) on the input samples based on the traditional autoencoder, causing local damage to the clean signal, and then uses the damaged samples to reconstruct the original input, thereby enhancing the generalization ability of the model. By damaging the input signal, the SDAE network can learn more robust data features, but the greater the proportion of zeroing, the less effective information hidden in the input data, which is not conducive to the extraction of abstract features.

[0102] In summary, the multiple SAE initial classifiers of this embodiment are constructed based on multiple different activation functions, multiple different training sample subsets and multiple different SAE variant models. Among them, the multiple different activation functions include multiple activation functions based on exponential operations and multiple activation functions based on non-exponential function operations; the multiple different training sample subsets are obtained by sampling the training sample set using the Bootstrap sampling method, and one training sample subset trains one SAE initial classifier. Assuming that there are a activation functions, b training data subsets, and c SAE variant models, a total of (a×b×c) SAE initial classifiers can be constructed. In this way, the efficiency of classifier construction can be increased, and SAE models with different characteristics can be constructed. When constructing a certain number of SAE classifiers, this embodiment combines the three classifier construction methods mentioned above to increase the number of classifiers.

[0103] The principles for constructing the UM diversity metric in steps (1-3) are as follows (1)-(2).

[0104] (1) Sample margin function design.

[0105] From the sample margin theory, we know that compared with "hard" samples, the classification of "marginal" samples often has a greater impact on the overall effect of the ensemble and needs to be paid special attention to. In ensemble learning, if a classifier A can correctly classify all "marginal" samples, while another classifier B can only correctly classify all "hard" samples, assuming that the classification accuracy of the two is consistent, then classifier A can effectively correct the classification bias in the ensemble and turn the ensemble classification results of some samples from wrong classification to correct classification; while classifier B can only enhance the confidence level of correctly classified samples, cannot further improve the ensemble accuracy, and may also lead to redundant classifiers. Therefore, in the ensemble pruning process, classifier A is more important than classifier B.

[0106] Since the contribution of “marginal” samples to the ensemble is higher than that of “hard” samples, we need to attach importance to classifiers that can correctly classify “marginal” samples and quantify the contribution of these classifiers to the ensemble by using a corresponding metric, which we call marginal margin. i For example, the sample margin is set to margin(x i ), and the marginal value of the "margin" sample is margin(x i ) approaches 0, so the designed margin function value should be within margin(x i) is close to 0 and obtains a higher value. In addition, “marginal” samples and “hard” samples have different importance in the ensemble, and the marginal margin function cannot be a simple linear function. Therefore, an exponential-based marginal margin function is considered. Let the total set of classifiers be H = {h 1 ,h 2 ,…,h L}, then the sample x i The sample margin function in H is defined as:

[0107]

[0108] In the formula, γ, μ, and ε are parameters. When γ=μ=1, ε=0, the image of the standard margin function is as follows: Figure 5 As shown. This function has the highest function value when the sample margin value margin is 0, the lowest function value when the margin is -1 or 1, and when the margin approaches 0, the slope of the function will gradually increase, reflecting the characteristics of a nonlinear function. This margin function can be explained as follows: 1) If a classifier can successfully classify those "marginal" samples with margin values ​​around 0, a higher function value is given to reflect its importance; 2) For those "hard" samples with margin values ​​very close to -1 and 1, the classification results of these samples in the ensemble are highly consistent, and the classification results of a single classifier have little effect on the overall ensemble. Therefore, if a classifier can successfully classify the "hard" samples, a lower function value is given, which is in line with the relevant laws of the sample margin theory.

[0109] The standard margin function is not flexible enough, so the new margin function There are also three parameters, γ, μ, and ε, which are used to specifically adjust the importance between "marginal" samples and "hard" samples. These three parameters have different meanings in integrated pruning.

[0110] ①The μ parameter is used to adjust the relative size of the margin value between the “hard” sample and the “marginal” sample to indicate the different importance of these two samples in the integration. When γ = 1 and ε = 0, the marginal margin function corresponding to some μ values ​​is as follows: Figure 6 As shown in the figure, it can be seen that the larger the μ value, the smaller the function value of the margin function at the margin value of -1 or 1, and the more it can highlight the difference between the two different samples. In addition, this figure also shows the image of the linear function for comparison. It can be seen that the slope of the linear function is 0 everywhere, which cannot highlight the importance of the "marginal" samples. For the exponential function, this problem can be avoided well.

[0111] ② The γ parameter is used to adjust the absolute size of the margin difference between the “hard” sample and the “margin” sample. Compared with the μ parameter, the γ parameter is to further expand the difference between the two different samples to highlight the importance of the “margin” sample. When μ = 1 and ε = 0, the margin function corresponding to some γ values ​​is as follows: Figure 7 As shown, it can be seen that the larger the γ value is, the larger the function value of the margin function at the margin value of 0 is.

[0112] ③The ε parameter controls the left and right movement of the margin function. When μ=γ=1, the marginal margin function is an even function, and different samples may have the same marginal margin value. For example, the marginal margin values ​​of two different samples with margin values ​​of 0.3 and -0.3 are the same. However, margin(x i )>0 indicates that the total set of classifiers H is good for sample x i The classification is correct, margin(x i )<0 indicates that the total set of classifiers H is effective for sample x i Classification error. For some samples with the same margin value but different ensemble classification results, in order to reflect their different importance in the ensemble pruning process, the parameter ε is introduced to adjust the function value. The margin function corresponding to some ε values ​​is as follows: Figure 8 shown.

[0113] (2) UM diversity indicator design.

[0114] Although the traditional UWA diversity index can represent the impact of a single classifier on the diversity of the classifier set, it lacks attention to the "marginal" samples. The models screened out by the ensemble pruning process are highly homogenized and cannot exert the powerful generalization ability of the ensemble. It is necessary to improve the existing diversity index. It can also be used to balance the size between the marginal margin value and the UWA value to avoid a large difference in the dimensions of the two. The proposed marginal margin function The UWA diversity metric is introduced to form the UM (Uncertainty Weighted Accuracy-Margin) diversity metric. The UM diversity metric is constructed according to the following formula (17).

[0115]

[0116] In formula (17), h represents a single SAE base classifier, H represents the set of trained SAE base classifiers, and X represents the training sample set; x i represents a single sample in the training sample set X, y i Represents x i The true category of 10 (h,H,x i ,y i ) represents h(xi )=y i And H(x i )≠y i , e 01 (h,H,x i ,y i ) represents h(x i )≠y i And H(x i )=y i , e 11 (h,H,x i ,y i ) represents h(x i )=y i And H(x i )=y i , e 00 (h,H,x i ,y i ) represents h(x i )≠y i And H(x i )≠y i NT i Denotes the sample (x) within H i ,y i ) gives the proportion of classifiers that make correct classification decisions, NF i =1-NT i Denotes the sample (x) within H i ,y i ) gives the proportion of classifiers that are misclassified, and I represents the indicator function; Represents the marginal margin value, which is used to characterize the importance of samples with highly inconsistent classification results in ensemble learning. margin(x i ,y i ) represents the sample (x i ,y i ) marginal value; γ represents the first margin coefficient, which is used to characterize the absolute size of the margin difference between hard samples and marginal samples; ε represents the second margin coefficient, which is used to characterize the left and right position of the UM diversity metric value; μ represents the third margin coefficient, which is used to characterize the relative size of the marginal margin value between hard samples and marginal samples. In addition, γ can also be used to balance the size between the marginal margin value and the UWA value to avoid a large difference in the dimensions of the two. The proposed UM indicator adds a marginal margin term on the basis of the original UWA indicator.

[0117] Step (1-1) to step (1-6) is the process of using the UM diversity metric to perform a single pruning on the SAE base classifier set to obtain a fault diagnosis model. In special cases, the model obtained by a single pruning can meet the requirements. However, in most cases, the SAE base classifier set needs to be pruned multiple times, that is, multiple iterations, to obtain a fault diagnosis model. The following steps (2-1) to step (2-4) illustrate the method of training and optimizing a fault diagnosis model by performing multiple pruning using multiple UM diversity metrics. Among them, the pruning principle of step (1-4) and the integration principle of step (1-5) are introduced through the following embodiment (combined with the particle swarm algorithm).

[0118] Step (2-1), obtain a training sample set and construct multiple SAE initial classifiers.

[0119] Step (2-2), using the training sample set to train multiple SAE initial classifiers to obtain a set of trained SAE base classifiers.

[0120] Step (2-3), construct UM diversity measurement indicators.

[0121] Step (2-4), iteratively execute the following steps until the preset maximum number of iterations threshold is reached:

[0122] Step (2-4-1), using the improved particle swarm algorithm to optimize the index parameters of the UM diversity metric index, and obtain the optimal value of the index parameters. The index parameters include the first margin coefficient, the second margin coefficient and the third margin coefficient.

[0123] Step (2-4-2), update the UM diversity metric according to the optimal value.

[0124] Step (2-4-3), use the updated UM diversity metric to perform integrated pruning on the trained SAE base classifier set to obtain a SAE base classifier subset that takes both diversity and accuracy into account.

[0125] Step (2-4-4), integrate all SAE base classifiers in the SAE base classifier subset to obtain an integrated SAE classifier.

[0126] Step (2-4-5), obtain a test sample set, use the test sample set to test the integrated SAE classifier, obtain a tested integrated SAE classifier, and determine the fitness value based on the test results, so as to substitute the fitness value into the improved particle swarm formula to update the parameters of the particle swarm.

[0127] Step (2-4-6) determines whether the preset maximum number of iterations has been reached. If so, a fault diagnosis model is constructed based on the integrated SAE classifier of all test certificates. Otherwise, the process returns to the step of optimizing the index parameters of the UM diversity metric using the improved particle swarm algorithm and obtaining the optimal values ​​of the index parameters to continue the process.

[0128] Step (2-1) to step (2-3) refer to the previous steps (1-1) to step (1-3). Step (2-4) is described in detail below. In this embodiment, the goal of integrated pruning is to select k classifiers from the set H according to the UM indicator and form a subset U. This process needs to consider two issues: 1) how to set the parameters in the UM indicator; 2) how to use the UM indicator to select the classifier subset U from the total set of classifiers H. This embodiment optimizes the parameters in the UM indicator through step (2-4) and also prunes the total set of classifiers H.

[0129] Step (2-4-1) is a parameter optimization step, which specifically includes the following steps: step (2-4-11) to step (2-4-14).

[0130] Step (2-4-11), determine the accuracy of the SAE base classifier set, and determine the fitness function value according to the accuracy of the SAE base classifier set.

[0131] Step (2-4-12), initialize the basic parameters of the improved particle swarm algorithm, and initialize the position and velocity of each particle; wherein the particle represents the candidate values ​​of the first margin coefficient, the second margin coefficient and the third margin coefficient.

[0132] Step (2-4-13), update the particle's position, velocity and basic parameters according to the fitness function value.

[0133] Step (2-4-14) determines whether the preset maximum number of iterations threshold is reached. If not, return to determine the accuracy of the SAE base classifier set, and continue to determine the fitness function value based on the accuracy of the SAE base classifier set. If yes, the indicator parameter corresponding to the global optimal value of the particle is used as the optimization value.

[0134] In order to more effectively reduce the ensemble bias, we can consider combining multiple voting results as the fitness function value: after obtaining the pruned classifier subset U, calculate the voting accuracy ACC of the k models k , randomly extract (k-1) models from U and calculate their voting accuracy ACC k-1 , then randomly select (k-2) models from U and calculate their voting accuracy ACC k-2 , and so on, until we get k voting accuracy sets A:{ACC1 ,ACC 2 ,...,ACC k-1 ,ACC k}, and finally the result obtained by adding up all the voting accuracy in A is used as the fitness value. This method can maximize the stability of the integration result. The specific calculation steps for determining the accuracy of the SAE base classifier set in step (2-4-11) are shown in steps 1 to 4 below.

[0135] Step 1: Use k classifiers in U to perform fault diagnosis on the sample set and obtain k classification results. For any sample x in X i , the fraction of the fault category c is calculated as:

[0136]

[0137] Where Pre(h i ,x j ) represents the predicted label of sample xj by classifier hj.

[0138] Step 2, sample x j The predicted label of can be calculated as:

[0139] Label(x i )=max c∈{1,2,…,C} {Pre(x i ,1),Pre(x i ,2),…,Pre(x i ,C)} (19)

[0140] In the formula, C is the number of sample labels.

[0141] Step 3: Submit the sample x j The predicted labels are compared with the actual labels, and the final accuracy result can be expressed as:

[0142]

[0143] In the formula, Used to compare the predicted labels and true labels of samples, where N is the number of samples.

[0144] Step 4: Randomly extract (k-1) models from U and calculate the voting accuracy ACC k-1 , then randomly extract (k-2) models from U and calculate the voting accuracy ACC k-2 , until {ACC 1 ,ACC 2 ,...,ACC k-1 ,ACC k}, then The fitness function value is used for IPSO parameter optimization. The higher the fitness function value, the higher the voting accuracy of the classifier set and the better the overall classification performance of the classifier.

[0145] Using the IPSO algorithm to optimize the three parameters in the UM index can greatly reduce the computational complexity of the algorithm and improve the computational efficiency. The IPSO algorithm can achieve a balance between global search and local search by adjusting the inertia weight and acceleration factor, find the optimal parameters, improve the screening efficiency of integrated pruning, and ensure good results of subsequent fault diagnosis.

[0146] The ensemble pruning method can be roughly divided into three methods: sorting, clustering, and optimization. The methods based on clustering and optimization are costly and computationally complex, while the ensemble pruning method based on sorting is computationally cheaper and simpler, and is suitable for situations with a large number of classifiers. Therefore, this embodiment will use the sorting method for ensemble pruning to screen out k classifiers that take into account both accuracy and diversity. Step (2-4-3) is an ensemble pruning method, and step (2-4-3) specifically includes the following steps (2-4-31) to (2-4-35).

[0147] Step (2-4-31), obtain the optimal value.

[0148] Step (2-4-32), calculate the UM diversity index value between each SAE base classifier in the SAE base classifier subset and the SAE base classifier subset, and obtain multiple UM diversity index values.

[0149] Step (2-4-33), arrange the multiple UM diversity index values ​​in descending order to filter out the diversity index values ​​corresponding to the multiple preset number thresholds, and use the multiple SAE base classifiers corresponding to the diversity index values ​​corresponding to the multiple preset number thresholds as the SAE base classifier subset.

[0150] Step (2-4-34), integrating multiple SAE base classifiers in the SAE base classifier subset;

[0151] Step (2-4-35), taking the integrated fault diagnosis result as the fitness value. In this embodiment, the majority voting method is used to integrate the SAE base classifiers in the SAE base classifier subset, and the voting result is taken as the fitness value.

[0152] When obtaining the optimal value, the UM diversity index has three parameters, γ, μ, and ε, which are used to adjust the marginal margin value between samples. However, manual adjustment is time-consuming and labor-intensive, costly, and may not necessarily obtain the optimal parameters. The Improved Particle Swarm Optimization (IPSO) algorithm has fast convergence speed, strong adaptability, and can better find the global optimal solution. Therefore, this embodiment will use the IPSO algorithm to optimize the γ, μ, and ε parameters. In addition, the fitness function of the IPSO algorithm is mainly used to evaluate the quality of the particle position in the search space, and is usually defined as the objective function of the problem to be optimized or an evaluation index related to the objective function. In integrated pruning, the choice of fitness function will affect the overall effect of integrated pruning and needs to be taken seriously.

[0153] In this embodiment, each iteration will find the three parameters of γ, μ, and ε in the UM index, and then k SAE models can be screened out through the sorting ensemble pruning method. Therefore, this embodiment can consider the overall classification performance of these k models as the fitness function value. The voting method is a combination method of multiple classifiers. This method determines the integration result of the classifier through the mechanism of majority voting, which can represent the overall classification performance of the classifier set and can be used as the fitness function value. At this time, the higher the fitness function value, the better the integrated classification performance.

[0154] However, this embodiment also uses the UM diversity metric as a pruning reference indicator and uses the IPSO algorithm to optimize some of its parameters to avoid manual adjustment of parameters. Fig. 9 As shown, the specific calculation process is shown in the following steps 1 to 11.

[0155] Step 1: Use sensors to collect vibration data of mechanical equipment under different working conditions, and then divide the original data into training samples and test samples without preprocessing.

[0156] Step 2: Use different activation functions, different training data subsets, and different SAE variant models to build a certain number of SAE models with different characteristics to form the original classifier set H, which contains L classifiers.

[0157] Step 3: Use the sample set to conduct fault classification experiments on all classifiers in H, use the accuracy as the evaluation index, and use the diagnosis result of the test sample set as the classification result, and obtain a total of L fault diagnosis results.

[0158] Step 4: Determine the population size N and the maximum number of iterations t of the particle swarm max , fitness function, initial learning factor and other parameters, and then initialize the position and velocity of each particle in the population.

[0159] Step 5: Send the single classifier with the highest accuracy to the empty set U and remove it from H. Let the number of classifiers in U be g.

[0160] Step 6: Use the IPSO algorithm to optimize the parameters γ, μ, and ε, then input the obtained parameter solution into the integrated model, and use the UM indicator to calculate each classifier h in H i Diversity value between

[0161] Step 7: Arrange in descending order, and put the classifier h corresponding to the highest value max Send it to U and delete this classifier from H.

[0162] Step 8: Repeat the above steps until k classifiers in U are selected and the fitness function value is calculated.

[0163] Step 9: According to the fitness function value, use the IPSO calculation formula to update the particle's inertia weight, learning factor and other parameters, and calculate the two attributes of speed and position to obtain the individual optimal particle and the global optimal particle. If the fitness value of the individual optimal particle is higher than the global optimal value, it will be updated to the global optimal value, otherwise it will remain unchanged.

[0164] Step 10: If the number of iterations at this time is greater than or equal to t max , then this round of optimization ends, otherwise return to step 6 to continue iterative calculation until the number of iterations is reached.

[0165] Step 11, output the global optimal value and its corresponding particle, and output the corresponding three parameters γ, μ, ε and the k classifiers obtained by integrated pruning.

[0166] The above steps are the specific process of the proposed integrated pruning based on the UM indicator. This method mainly includes two parts: classifier screening based on the sorting method and parameter optimization based on the IPSO algorithm. First, the local optimal parameters obtained by each iteration of IPSO are substituted into the UM indicator, and then the sorting method is used to screen out k classifiers and obtain the corresponding integrated fault diagnosis results. If the number of iterations is less than the maximum number of iterations, the IPSO algorithm is continued to be used for parameter optimization, otherwise the three parameters of γ, μ, and ε are output at this time. Finally, the UM indicator after parameter optimization is used to sort the total set of classifiers H for integrated pruning, and k classifiers that take into account diversity and accuracy are screened out. These k models have good classification capabilities and have certain classification differences between each other, which can ensure the diversity of the integration and ensure that the subsequent fault diagnosis results have high accuracy and strong stability.

[0167] After obtaining the classifiers after ensemble pruning, the next step is to consider what strategy should be used to integrate these classifiers. Common classifier integration strategies include majority voting, weighted voting, averaging, stacking, etc.

[0168] The majority voting method is a simple ensemble method that uses multiple classifiers to independently classify samples, and then decides the final category based on the votes of the classification results. Each classifier has the same voting weight, which is simple and easy to implement and applicable to a variety of classification tasks. However, the majority voting method may be affected by the performance of some classifiers, especially when some classifiers have poor performance, the ensemble accuracy may be poor.

[0169] Weighted voting is an improved majority voting method that determines the final category by assigning different weights to each classifier. The weights can be assigned based on the performance or confidence of the classifier. This method can fully utilize the unique advantages of each classifier and improve prediction performance. However, this method requires determining the appropriate weight allocation, and the weight has a huge impact on the ensemble accuracy, which may require more computing resources.

[0170] The average method obtains the final prediction value by calculating the average of the prediction results of multiple models. This method is simple and effective, and can smooth the prediction results and reduce the impact of outliers. However, it has major defects when there are large differences between models.

[0171] The Stacking method uses a meta-model to combine the prediction results of the basic models. The meta-model is usually a simple model, such as linear regression or logistic regression, and the meta-model makes the final prediction based on the prediction results of the basic models. It can make full use of the unique advantages of each model. The meta-model can learn how to assign weights based on the prediction results of the basic models to achieve weighted combination. The Stacking method can greatly improve the performance of integrated classification and reduce the risk of overfitting. It is very flexible and changeable. However, since the Stacking method involves the combination of multiple models, the interpretability of its prediction results may be reduced, which makes it more difficult to understand the decision-making process of the model.

[0172] Among these methods, the majority voting method has low computational cost, high computational efficiency and is convenient to calculate. Therefore, this embodiment selects the majority voting method to integrate the classifiers obtained by integrated pruning, and obtains the final mechanical fault diagnosis result from the integrated result.

[0173] This embodiment also experimentally verifies the method of this embodiment. The data set uses the bearing vibration data set provided by the Case Western Reserve Bearing Data Center, which is the same as the aforementioned sample set. It is also an important reference benchmark for verifying the effectiveness of the algorithm. The division of the sample set in the experiment is the same as in the previous text. This embodiment uses the Bootstrap method to sample the training sample set and construct 4 training data subsets. Each subset trains an SAE model, and then uses the 10 activation functions listed in Table 2, 3 autoencoder variants (stack sparse autoencoder, stack shrinkage autoencoder, stack denoising autoencoder), and 4 training data subsets to construct a total of (10×4×3) SAE classifiers for subsequent integrated pruning experiments.

[0174] In the integrated pruning experiment, this embodiment uses the IPSO algorithm to optimize the three parameters of γ, μ, and ε in the UM diversity index. After each parameter iteration, the UM index and the selected model will also change. In order to verify the effectiveness of the IPSO algorithm, this experiment will use the standard PSO algorithm as a comparison method. The PSO algorithm parameters are set as follows: the population size is 15, the maximum number of iterations is 60, and the learning factor c is 1. 1 =c 2 =3, inertia weight ω is 0.4; IPSO algorithm parameters are set to: maximum learning factor c 1_max and c 2_max are 2 and 1 respectively, and the minimum value c 1_min and c 2_min The maximum and minimum values ​​of the weight coefficient are 0.9 and 1 respectively. The other parameters are the same as those of the PSO algorithm, and the k value is 12. In addition, the calculation process of the fitness function is shown in Section 3.4.1. The higher the fitness function value, the better the overall classification performance of the selected classifier set. The parameter optimization process of the IPSO and PSO algorithms is as follows: Fig.11 As shown in the figure. During the 60 iterations, the PSO algorithm fell into the local optimum 6 times, and found the global optimum at the 42nd iteration. The fitness function value at this time reached the maximum value of 11.29; while the IPSO algorithm only fell into the local optimum 4 times during the entire iteration process, and found the global optimum at the 26th iteration, which was earlier than the PSO algorithm. In addition, during the iteration process, the maximum value of the IPSO fitness function was 11.37, which was also higher than the maximum value of 11.29 of the PSO algorithm. Compared with the PSO algorithm, the IPSO algorithm can get rid of the local optimum earlier and perform better in terms of convergence speed and convergence accuracy.

[0175] In this integrated pruning experiment, k SAE models were selected from 120 SAE models. In order to verify the overall classification performance of these k models, the voting method will be used to integrate the classification results of these models and obtain the corresponding fault diagnosis accuracy. As the k value increases, the fault diagnosis results of the voting method will gradually improve until convergence. In order to verify the effectiveness of the sorted integrated pruning method, this experiment will also use two integrated pruning methods for comparison: the first is a pruning method based only on accuracy measurement. This method arranges the 120 SAE models in descending order according to the classification accuracy and selects the k SAE models with the highest accuracy; the second is a random screening method, which randomly selects k SAE models from 120 SAE. The voting results of these three integrated pruning methods under different k values ​​are shown in the figure. Fig.12 shown.

[0176] Depend on Fig.12 It can be seen that the voting accuracy based on the UM indicator is the highest in all cases and converges when the number of classifiers k reaches 12; for the integrated pruning method based on the accuracy indicator, when k increases to 16, its voting classification accuracy reaches 99.8%, and thereafter, when the number of k is increased, the voting results basically do not change; for the integrated pruning method of random screening, when k increases to 17, its classification accuracy reaches the convergence value of 99.9%. The latter two pruning methods are significantly inferior to the integrated pruning method based on the UM indicator in terms of convergence accuracy. In addition, it can be seen from Fig.12 As can be seen from the figure, with the increase of the number of classifiers k, only the voting accuracy based on the UM index shows an increasing trend, and the other two methods have varying degrees of fluctuations. For the random ensemble pruning method, the voting accuracy of this method varies greatly and is disordered because the random selection method will make the ensemble results uncontrollable. For the ensemble pruning method based on the accuracy index, its voting accuracy shows a trend of first rising, then falling, and then rising until convergence, which is contrary to the general ensemble law. The reason is that there is such a classifier: they only have more classification errors for a certain type of samples and fewer classification errors for other types of samples. Therefore, if there are multiple such classifiers in the ensemble, the impact of these classifiers will be higher than that of other base classifiers, resulting in a higher error rate in the classification results of some samples, which results in the lack of classifier diversity and reduced ensemble performance.

[0177] This embodiment mainly includes two parts: the establishment of UM diversity metric and the proposal of sorting ensemble pruning method. Among them, UM diversity metric introduces sample marginal correlation theory into UWA diversity index, focusing on those samples at the edge of the integration. After that, this embodiment combines the sorting method with the UM metric to screen out the most diverse SAE model set from the classifier pool. Finally, the CWRU bearing dataset is used for experimental verification, and the effectiveness of the proposed method in mechanical fault diagnosis is verified by comparing it with the method.

[0178] The method proposed in this embodiment combines the ideas of ensemble learning and ensemble pruning with the deep learning model, fully utilizes the diversity between classifiers and combines the characteristics of multiple models, effectively improving the universality and robustness of the fault diagnosis algorithm; in addition, since the algorithm of this embodiment can effectively process "marginal" samples and minimize the impact of redundant classifiers on the ensemble, it can be applied to small sample fault diagnosis problems.

[0179] Firstly, the sample margin theory is introduced into the UWA diversity index, and the sample margin theory is combined with the UWA diversity index. ① A new UM (Uncertainty Weighted Accuracy-Margin) diversity metric is constructed, focusing on the classification of "marginal" samples by the classifier; ② Then the improved particle swarm optimization (IPSO) is used to optimize some parameters in the UM metric; ③ Then, this embodiment proposes a UM index sorting ensemble pruning method to screen out a subset of classifiers that take into account diversity and accuracy from the original classifier pool; ④ Finally, the screened classifiers are combined using a certain ensemble learning strategy (such as majority voting, weighted voting, etc.) to obtain the final integrated fault diagnosis result.

[0180] This embodiment combines ensemble learning with deep learning ideas. In the field of mechanical fault diagnosis, by combining multiple autoencoder models and technologies, ensemble deep learning can improve the accuracy and robustness of mechanical fault diagnosis, and can more accurately identify and classify different fault types. By combining multiple models or technologies, ensemble deep learning can reduce the risk of overfitting, improve the generalization ability of the model, and make it more adaptable and stable when facing new tasks or new data.

[0181] In addition, in order to address the problem of increased computational costs and insignificant improvement in the accuracy of integrated fault diagnosis caused by too many classifiers in integrated deep learning, this embodiment introduces the idea of ​​integrated pruning into integrated learning, removes redundant classifiers, improves the generalization performance of classifiers, and constructs a more efficient and better-performing integrated classifier set to improve the computational efficiency of fault diagnosis. Specifically, in order to address the problem that the existing integrated pruning methods lack attention to "marginal" samples, resulting in insufficient diversity of the screened classifiers, this embodiment introduces the sample margin theory into the UWA indicator and constructs a new UM diversity metric, focusing on the classification of "marginal" samples by the classifier; then a sorted integrated pruning method based on the UM indicator is proposed to screen out the most diverse classifier set from the original classifier pool and delete redundant classifiers.

[0182] See also Fig.13 , a mechanical equipment fault diagnosis device based on an integrated stack autoencoder provided in this embodiment includes a data acquisition module 10 and a fault diagnosis module 20. The data acquisition module 10 is used to obtain the fault data of the target mechanical equipment. The fault diagnosis module 20 is used to input the fault data into a trained and optimized fault diagnosis model for fault diagnosis to obtain a fault diagnosis result. Among them, the fault diagnosis model is constructed based on several different SAE base classifiers screened by the UM diversity measurement method, and the UM diversity measurement method is a method of introducing the sample marginal theory into the UWA indicator.

[0183] In some possible embodiments, the fault diagnosis model is trained and optimized according to the following modules: a first acquisition module, used to acquire a training sample set and construct multiple SAE initial classifiers; a first training module, used to train multiple SAE initial classifiers using the training sample set to obtain a set of trained SAE base classifiers; a first indicator construction module, used to construct a UM diversity measurement indicator; a first integrated pruning module, used to integrate and prune the trained SAE base classifier set using the UM diversity measurement indicator to obtain a subset of SAE base classifiers that takes into account diversity and accuracy; a first integration module, used to integrate all SAE base classifiers in the SAE base classifier subset to obtain an integrated SAE classifier; a first construction module, used to construct a fault diagnosis model based on the integrated SAE classifier.

[0184] In some other possible embodiments, the fault diagnosis model is trained and optimized according to the following modules: a second acquisition module, used to acquire a training sample set and construct multiple SAE initial classifiers; a second training module, used to train multiple SAE initial classifiers using the training sample set to obtain a set of trained SAE base classifiers; a second indicator construction module, used to construct a UM diversity measurement indicator; an iteration module, iteratively executing the following iteration steps until a preset maximum number of iterations threshold is reached.

[0185] The iteration module includes: an optimization module, which is used to optimize the indicator parameters of the UM diversity metric using an improved particle swarm algorithm to obtain the optimal values ​​of the indicator parameters; the indicator parameters include a first margin coefficient, a second margin coefficient and a third margin coefficient; an update module, which is used to update the UM diversity metric according to the optimal value; a second integrated pruning module, which is used to integrate and prune the trained SAE base classifier set using the updated UM diversity metric to obtain a SAE base classifier subset that takes into account diversity and accuracy; a second integration module, which is used to integrate all SAE base classifiers in the SAE base classifier subset to obtain to the integrated SAE classifier; a testing module, used to obtain a test sample set, test the integrated SAE classifier using the test sample set, obtain the tested integrated SAE classifier, and determine the fitness value according to the test result, so as to substitute the fitness value into the improved particle swarm formula to update the parameters of the particle swarm; a first judgment module, used to judge whether the preset maximum number of iterations threshold is reached, if so, construct a fault diagnosis model according to the integrated SAE classifier of all test certificates; otherwise, return to the step of optimizing the indicator parameters of the UM diversity metric indicator using the improved particle swarm algorithm, and obtain the optimized value of the indicator parameter to continue execution.

[0186] In some possible embodiments, the UM diversity metric is constructed according to the following formula:

[0187]

[0188] In the formula, h represents a single SAE base classifier, H represents the set of trained SAE base classifiers, and X represents the training sample set; x i represents a single sample in the training sample set X, y i Represents x i The true category of 10 (h,H,x i ,y i ) represents h(x i )=y i And H(x i )≠y i , e 01 (h,H,x i ,yi ) represents h(x i )≠y i And H(x i )=y i , e 11 (h,H,x i ,y i ) represents h(x i )=y i And H(x i )=y i , e 00 (h,H,x i ,y i ) represents h(x i )≠y i And H(x i )≠y i NT i Denotes the sample (x) within H i ,y i ) gives the proportion of classifiers that make correct classification decisions, NF i =1-NT i Denotes the sample (x) within H i ,y i ) gives the proportion of classifiers that are misclassified, and I represents the indicator function; Represents the marginal margin value, which is used to characterize the importance of samples with highly inconsistent classification results in ensemble learning. margin(x i ,y i ) represents the sample (x i ,y i )’s marginal value; γ represents the first margin coefficient, which is used to characterize the absolute size of the margin difference between hard samples and marginal samples; ε represents the second margin coefficient, which is used to characterize the left and right positions of the UM diversity metric value; μ represents the third margin coefficient, which is used to characterize the relative size of the marginal margin value between hard samples and marginal samples.

[0189] In some possible embodiments, multiple SAE initial classifiers are constructed based on multiple different activation functions, multiple different training sample subsets, and multiple different SAE variant models; wherein the multiple different activation functions include multiple activation functions based on exponential operations and multiple activation functions based on non-exponential function operations; and the multiple different training sample subsets are obtained by sampling the training sample set using the Bootstrap sampling method.

[0190] In some possible embodiments, the optimization module includes: a determination module, which is used to determine the accuracy of the SAE base classifier set and determine the fitness function value according to the accuracy of the SAE base classifier set; an initialization module, which is used to initialize the basic parameters of the improved particle swarm algorithm and initialize the position and speed of each particle; wherein the particle represents the candidate values ​​of the first margin coefficient, the second margin coefficient and the third margin coefficient; a parameter updating module, which is used to update the position, speed and basic parameters of the particle according to the fitness function value; a second judgment module, which is used to judge whether the preset second number of iterations is reached. If not, it returns to determine the accuracy of the SAE base classifier set, and continues to execute the step of determining the fitness function value according to the accuracy of the SAE base classifier set. If yes, the indicator parameter corresponding to the global optimal value of the particle is used as the optimization value.

[0191] In some possible embodiments, the second integrated pruning module includes: a parameter acquisition module, used to obtain an optimization value; an index value module, used to calculate the UM diversity index value between each SAE base classifier in the SAE base classifier subset and the SAE base classifier subset, and obtain multiple UM diversity index values; a sorting pruning module, used to arrange the multiple UM diversity index values ​​in descending order to filter out the diversity index values ​​corresponding to multiple preset number thresholds, and use the multiple SAE base classifiers corresponding to the diversity index values ​​corresponding to the multiple preset number thresholds as the SAE base classifier subset; a third integration module, used to integrate the multiple SAE base classifiers in the SAE base classifier subset, and use the integrated fault diagnosis result as the fitness value.

[0192] In some possible embodiments, in the first integration module and the second integration module, the SAE base classifiers in the SAE base classifier subset are integrated using majority voting method respectively, and in the second integration module, the voting result is used as the fitness value.

[0193] In some possible embodiments, each training sample subset trains an SAE initial classifier to obtain a SAE initial classifier trained with the training sample subset.

[0194] The device provided in the embodiment of the present application is adopted. Since the device adopts the same inventive concept as the above-mentioned method provided in the embodiment of the present application, on the premise that the method can solve the technical problem, the device can also solve the technical problem.

[0195] It is known from common technical knowledge that the present invention can be implemented by other embodiments that do not deviate from its spirit or essential features. Therefore, the above disclosed embodiments are only illustrative in all respects and are not exclusive. All changes within the scope of the present invention or within the scope equivalent to the present invention are included in the present invention.

Claims

1. A mechanical equipment fault diagnosis method based on an integrated stack autoencoder, characterized in that: include: Obtaining fault data of target mechanical equipment; Inputting the fault data into a trained and optimized fault diagnosis model to perform fault diagnosis and obtain a fault diagnosis result; The fault diagnosis model is constructed based on several different SAE base classifiers screened by the UM diversity measurement method, and the UM diversity measurement method is a method of introducing the sample margin theory into the UWA index.

2. The mechanical equipment fault diagnosis method based on integrated stack autoencoder according to claim 1 is characterized in that: The fault diagnosis model is trained and optimized according to the following method: Obtain training sample sets and construct multiple SAE initial classifiers; Using the training sample set to train the multiple SAE initial classifiers to obtain a trained SAE base classifier set; Constructing UM diversity metrics; The UM diversity metric is used to perform integrated pruning on the trained SAE base classifier set to obtain a subset of SAE base classifiers that takes both diversity and accuracy into consideration. Integrating all SAE base classifiers in the SAE base classifier subset to obtain an integrated SAE classifier; The fault diagnosis model is constructed according to the integrated SAE classifier.

3. The mechanical equipment fault diagnosis method based on integrated stack autoencoder according to claim 1 is characterized in that: The fault diagnosis model is trained and optimized according to the following method: Obtain training sample sets and construct multiple SAE initial classifiers; Using the training sample set to train the multiple SAE initial classifiers to obtain a trained SAE base classifier set; Constructing UM diversity metrics; The iterations are performed as follows until the preset maximum number of iterations is reached: Using an improved particle swarm algorithm to optimize the indicator parameters of the UM diversity metric indicator, and obtain the optimal value of the indicator parameter; the indicator parameter includes a first margin coefficient, a second margin coefficient and a third margin coefficient; Update the UM diversity metric according to the optimization value; Using the updated UM diversity metric to perform integrated pruning on the trained SAE base classifier set, to obtain a SAE base classifier subset that takes both diversity and accuracy into account; Integrating all SAE base classifiers in the SAE base classifier subset to obtain an integrated SAE classifier; Acquire a test sample set, test the integrated SAE classifier using the test sample set to obtain a tested integrated SAE classifier, and determine a fitness value according to the test result, so as to substitute the fitness value into the formula of the improved particle swarm to update the parameters of the particle swarm; Determine whether the preset maximum number of iterations threshold is reached. If so, construct the fault diagnosis model according to all tested integrated SAE classifiers; otherwise, return to the step of optimizing the index parameters of the UM diversity metric index using the improved particle swarm algorithm to obtain the optimized value of the index parameter and continue to execute.

4. The mechanical equipment fault diagnosis method based on integrated stack autoencoder according to claim 2 or 3, characterized in that: The UM diversity metric is constructed according to the following formula: In the formula, h represents a single SAE base classifier, H represents the set of trained SAE base classifiers, and X represents the training sample set; x i represents a single sample in the training sample set X, y i Represents x i The true category of 10 (h,H,x i ,y i ) represents h(x i )=y i And H(x i )≠y i , e 01 (h,H,x i ,y i ) represents h(x i )≠y i And H(x i )=y i , e 11 (h,H,x i ,y i ) represents h(x i )=y i And H(x i )=y i , e 00 (h,H,x i ,y i ) represents h(x i )≠y i And H(x i )≠y i NT i Denotes the sample (x) within H i ,y i ) gives the proportion of classifiers that make correct classification decisions, NF i =1-NT i Denotes the sample (x) within H i ,y i ) gives the proportion of classifiers that are misclassified, and I represents the indicator function; represents the marginal margin value, which is used to characterize the importance of samples with highly inconsistent classification results in ensemble learning. i ,y i ) represents the sample (x i ,y i )’s marginal value; γ represents the first margin coefficient, which is used to characterize the absolute size of the margin difference between hard samples and edge samples; ε represents the second margin coefficient, which is used to characterize the left and right positions of the UM diversity metric value; μ represents the third margin coefficient, which is used to characterize the relative size of the marginal margin value between hard samples and edge samples.

5. The mechanical equipment fault diagnosis method based on integrated stack autoencoder according to claim 2 is characterized in that: The multiple SAE initial classifiers are constructed based on multiple different activation functions, multiple different training sample subsets and multiple different SAE variant models; wherein the multiple different activation functions include multiple activation functions based on exponential operations and multiple activation functions based on non-exponential function operations; the multiple different training sample subsets are obtained by sampling the training sample set using the Bootstrap sampling method.

6. The mechanical equipment fault diagnosis method based on integrated stack autoencoder according to claim 3 is characterized in that: The method of optimizing the index parameters of the UM diversity metric index by using the improved particle swarm algorithm to obtain the optimized values ​​of the index parameters includes: Determine the accuracy of the SAE base classifier set, and determine the fitness function value according to the accuracy of the SAE base classifier set; Initializing basic parameters of the improved particle swarm algorithm, and initializing the position and velocity of each particle; wherein the particle represents the candidate values ​​of the first margin coefficient, the second margin coefficient, and the third margin coefficient; Update the position, speed and basic parameters of the particle according to the fitness function value; Determine whether the preset iteration number threshold is reached. If not, return to determine the accuracy of the SAE base classifier set, and continue to execute the step of determining the fitness function value according to the accuracy of the SAE base classifier set. If yes, use the indicator parameter corresponding to the global optimal value of the particle as the optimization value.

7. The mechanical equipment fault diagnosis method based on integrated stack autoencoder according to claim 3 is characterized in that: The updated UM diversity metric is used to perform integrated pruning on the trained SAE base classifier set to obtain a SAE base classifier subset that takes both diversity and accuracy into consideration, including: Obtaining the optimal value; Calculating a UM diversity index value between each SAE base classifier in the SAE base classifier subset and the SAE base classifier subset to obtain a plurality of UM diversity index values; Arrange the multiple UM diversity index values ​​in descending order to filter out the diversity index values ​​corresponding to the multiple preset number thresholds, and use the multiple SAE base classifiers corresponding to the multiple preset number thresholds as the SAE base classifier subset; The multiple SAE base classifiers in the SAE base classifier subset are integrated, and the integrated fault diagnosis result is used as the fitness value.

8. The mechanical equipment fault diagnosis method based on integrated stack autoencoder according to claim 3 is characterized in that: The step of integrating all SAE base classifiers in the SAE base classifier subset and taking the integrated fault diagnosis result as the fitness value includes: The SAE base classifiers in the SAE base classifier subset are integrated using a majority voting method, and the voting result is used as the fitness value.

9. The mechanical equipment fault diagnosis method based on integrated stack autoencoder according to claim 5, characterized in that: Each training sample subset is used to train an SAE initial classifier, and a SAE initial classifier trained with the training sample subset is obtained.

10. A mechanical equipment fault diagnosis device based on an integrated stack autoencoder, characterized in that: include: A data acquisition module, used to acquire fault data of target mechanical equipment; A fault diagnosis module is used to input the fault data into a trained and optimized fault diagnosis model to perform fault diagnosis and obtain a fault diagnosis result; The fault diagnosis model is constructed based on several SAE base classifiers screened by a UM diversity metric module, and the UM diversity metric module is constructed by introducing sample margin theory into the UWA index.

Images