CN118226254A

CN118226254A - Method for detecting motor based on self-encoder integrated learning model

Info

Publication number: CN118226254A
Application number: CN202410642058.XA
Authority: CN
Inventors: 陈兴达; 张胜; 顾正云; 卢明辉; 谢海圣
Original assignee: Suzhou Acoustic Industrial Technology Research Institute Co ltd
Current assignee: Suzhou Acoustic Industrial Technology Research Institute Co ltd
Priority date: 2024-05-23
Filing date: 2024-05-23
Publication date: 2024-06-21
Anticipated expiration: 2044-05-23
Also published as: CN118226254B

Abstract

The invention discloses a method for detecting a motor based on a self-encoder integrated learning model, and relates to the technical field of testing; the method comprises the steps of S101 extracting and obtaining normal audio characteristics of a motor from a training set, S102 obtaining a self-encoder integrated learning model, S103 training the self-encoder integrated learning model, S104 obtaining an integrated weight value, S2 obtaining an anomaly score through testing, and the like, wherein the method comprises the steps of calculating and obtaining the mean square error of training input characteristics and training output characteristics of each self-encoder and taking the mean square error as training errors in the training process, obtaining the reciprocal of each training error, normalizing and obtaining the integrated weight value, calculating and obtaining the mean square error of the testing input characteristics and the testing output characteristics of each self-encoder and taking the mean square error as a reconstruction error in the testing process, and carrying out weighted summation on the reconstruction errors to obtain the anomaly score, thereby improving the accuracy of motor anomaly detection.

Description

Method for detecting motor based on self-encoder integrated learning model

Technical Field

The invention relates to the technical field of testing, in particular to a method for detecting a motor based on a self-encoder integrated learning model.

Background

The motor is an important electromechanical device in modern production and life, and has important significance for smooth progress of industrial processes and safe operation of the device. But the produced motor is easy to have abnormal conditions, so that faults occur to generate potential safety hazards. The rapid discovery of abnormal equipment can reduce the number of defective products and prevent damage from continuing to propagate. For manual detection, detection of abnormal sounds and other characteristics is performed manually, so that resources of manpower and material resources are occupied greatly, but loss caused by manpower and economy can be reduced due to machine identification of abnormal equipment.

Anomaly detection of motor devices can be classified into two categories, supervised learning and unsupervised learning. Anomaly detection for supervised learning refers to normal data and abnormal data during training, while unsupervised learning refers to normal data only and no abnormal data during training. The anomaly detection of supervised learning is applicable to the situation that the anomaly data are more and easy to collect, and the anomaly detection of unsupervised learning is applicable to the situation that the anomaly data are less and difficult to collect.

The usual method of unsupervised learning is a gaussian model, a gaussian mixture model lof, local Outlier Factor, generating an antagonistic network GAN, GENERATIVE ADVERSARIAL Networks, and short-long term memory recurrent neural Networks Long Short Term Memory-Recurrent Neural Networks abbreviated as LSTM-RNN. However, the gaussian model, gaussian mixture model lof, local Outlier Factor, and the like have low capacity generalization capability, and it is difficult to process data of high-dimensional features. The GAN and LSTM-RNN models have large capacity and high generalization capability, but have many parameters and require a large amount of data. Therefore, the conventional self-encoder autoencoder abbreviated as AU solves this problem, and is thus widely used in the process of motor abnormality detection. However, the trained AU weights and thresholds have strong randomness and instability, thereby reducing accuracy.

The prior technical proposal is as follows:

the first prior art scheme: xu Zengpiao, guan Shuai, wang Yongjiang, et al A method for detecting abnormal sound faults of a loudspeaker based on vibration displacement [ J ]. University of Tianjin science and technology, 2012,27 (03): 64-67+78.DOI:10.13364/J. Issn.1672-6510.2012.03.009.

The second prior art scheme: li Mingchao motor fault diagnosis method based on abnormal sound detection [ D ]. University of five-cognac, 2014.

Third prior art scheme: yang Lei, weight, liliang, etc. Nuclear power electric equipment abnormal sound detection technique based on ITD-MFCC and convolutional neural network [ J ]. Noise and vibration control, 2023,43 (04): 122-128+207.

Fourth prior art solution: DCASE CHALLENGE 2020: unsupervised Detection of Anomalous Sounds for Machine Condition Monitoring.

Fifth prior art scheme: performing Anomaly Detection on Industrial Equipment Using Audio Signals.

Sixth prior art solution: sun Xuri, liu Mingfeng, cheng Hui, et cetera, combined with secondary feature extraction and LSTM-Autoencoder, network traffic anomaly detection method [ J ] university of Beijing traffic report, 2020,44 (02): 17-26.Doi:10.11860/J. Issn.1673-0291.20200005.

Seventh prior art solution: li Renjie Gear box abnormal sound detection research based on unsupervised challenge domain adaptation single classification [ D ]. Chongqing academy of technology, 2022.DOI:10.27854/d.cnki. Gcqkj.2022.000011.

Eighth prior art solution ：Y. Kawaguchi and T. Endo, "How can we detect anomalies from subsampled audio signals," 2017 IEEE 27th International Workshop on Machine Learning for Signal Processing (MLSP), Tokyo, Japan, 2017, pp. 1-6, doi: 10.1109/MLSP.2017.8168164. keywords: {Anomaly detection;Time-domain analysis;Hidden Markov models;Monitoring;Layout;Spectrogram;sub-Nyquist sampling;non-uniform sampling;end-to-end;long short-term memory (LSTM);autoencoder}.

Ninth prior art solution ：Marchi E, Vesperini F, Weninger F, et al. Non-linear prediction with LSTM recurrent neural networks for acoustic novelty detection[C]//2015 International Joint Conference on Neural Networks (IJCNN). IEEE, 2015: 1-7.

Tenth prior art solution ：Harsh Purohit, Ryo Tanabe, Kenji Ichige, Takashi Endo, Yuki Nikaido, Kaori Suefusa, and Yohei Kawaguchi, "MIMII Dataset: Sound Dataset for Malfunctioning Industrial Machine Investigation and Inspection," arXiv preprint arXiv:1909.09347, 2019.

Eleventh prior art solution ：Harsh Purohit, Ryo Tanabe, Kenji Ichige, Takashi Endo, Yuki Nikaido, Kaori Suefusa, and Yohei Kawaguchi, "MIMII Dataset: Sound Dataset for Malfunctioning Industrial Machine Investigation and Inspection," in Proc. 4th Workshop on Detection and Classification of Acoustic Scenes and Events (DCASE), 2019.

Disclosure of Invention

The invention provides a method for detecting a motor based on a self-encoder integrated learning model, which solves the technical problem of low detection precision of motor abnormality.

In order to solve the technical problems, the technical scheme adopted by the invention is as follows:

A method for detecting a motor based on a self-encoder integrated learning model comprises the following steps of: extracting and obtaining normal audio characteristics of the motor based on the normal audio data of the training current collector; step S102: obtaining a self-encoder integrated learning model, the self-encoder integrated learning model including a plurality of self-encoders; step S103: obtaining a weight and a threshold value of each self-encoder in the self-encoder integrated learning model, and obtaining a trained self-encoder integrated learning model; step S104: inputting the normal audio characteristics of the motor as training input characteristics into a trained self-encoder integrated learning model, obtaining the training output characteristics of each self-encoder in the self-encoder integrated learning model, calculating to obtain the mean square error of the training input characteristics and the training output characteristics of each self-encoder, taking the mean square error as training errors, obtaining the reciprocal of each training error, normalizing and obtaining the integrated weight; step S2: and extracting normal and abnormal audio characteristics of the motor based on the normal and abnormal audio data of the test collector, inputting the normal and abnormal audio characteristics of the motor as test input characteristics into a trained self-encoder integrated learning model, obtaining test output characteristics of each self-encoder in the self-encoder integrated learning model, namely the reconstructed characteristics, calculating to obtain the mean square error of the test input characteristics and the test output characteristics of each self-encoder, and taking the mean square error as reconstruction errors, and carrying out weighted summation on the reconstruction errors to obtain an abnormal score.

The further technical proposal is that: the step S101, the step S102, the step S103, and the step S104 form a step of training in the step S1, and the step S101 specifically includes the following steps: based on the normal audio data of the training collector, framing, windowing and Fourier transforming the time domain signal to obtain a spectrogram; step S1012: weighting and summing the spectrograms through a mel triangular filter, and compressing characteristic dimensions to obtain a mel spectrogram; step S1013: and converting the energy of the mel spectrogram into db to obtain the normal audio characteristics of the motor.

The further technical proposal is that: in the step S102, the self-encoder includes an encoder and a decoder encoder, the number of layers, the number of neurons and the activation function of the encoder in the self-encoder integrated learning model are set, and the number of layers, the number of neurons and the activation function of the decoder encoder in the self-encoder integrated learning model are set, so that the architecture of the self-encoder is determined, and the self-encoder integrated learning model is obtained.

The further technical proposal is that: in the step S103, the steps of obtaining the weight and the threshold value of each self-encoder in the self-encoder integrated learning model include a step S1031 and a step S1032, the step of obtaining the trained self-encoder integrated learning model is a step S1033,

Step S1031: determining a loss function;

(1)

In the formula (1): as a loss function; m is batch size scale; q is the dimension of the input feature; /(I) An ith feature of the jth input feature vector in batch; /(I)The ith feature of the jth output feature vector in the batch; j is E [1, m ], i is E [1, q ];

step S1032: calculating the weight and the threshold value of each self-encoder in the self-encoder integrated learning model based on a loss function, an optimization function Adam, a gradient descent method and a back propagation method;

Step S1033: repeating the step S1032 to obtain a trained self-encoder integrated learning model, namely obtaining a plurality of trained self-encoders.

The further technical proposal is that: in said step S104, a training error is calculated according to equation (2), an integrated weight is calculated according to equation (3),

(2)

In the formula (2): Is a training error function; k is the training sample size; q is the dimension of the input feature; /(I) An ith feature of a jth input feature vector in the training set; /(I)The ith feature of the jth output feature vector in the training set; j e [1, k ], i e [1, q ];

(3)

In the formula (3): The integrated weight of the ith self-encoder, i epsilon [1, n ]; n is the number of trained self-encoders.

The further technical proposal is that: in the step S2, an anomaly score is calculated according to equation (4),

(4)

In the formula (4): score is an anomaly score; n is the number of the self-encoders in the trained self-encoder integrated learning model; The weight of the self-encoder in the j-th trained self-encoder integrated learning model is j epsilon [1, n ]; /(I) An ith feature that is an input feature vector; /(I)An ith feature which is an output feature vector; q is the dimension of the feature.

The beneficial effects of adopting above-mentioned technical scheme to produce lie in:

A method for detecting a motor based on a self-encoder integrated learning model comprises the following steps of: extracting and obtaining normal audio characteristics of the motor based on the normal audio data of the training current collector; step S102: obtaining a self-encoder integrated learning model, the self-encoder integrated learning model including a plurality of self-encoders; step S103: obtaining a weight and a threshold value of each self-encoder in the self-encoder integrated learning model, and obtaining a trained self-encoder integrated learning model; step S104: inputting the normal audio characteristics of the motor as training input characteristics into a trained self-encoder integrated learning model, obtaining the training output characteristics of each self-encoder in the self-encoder integrated learning model, calculating to obtain the mean square error of the training input characteristics and the training output characteristics of each self-encoder, taking the mean square error as training errors, obtaining the reciprocal of each training error, normalizing and obtaining the integrated weight; step S2: and extracting normal and abnormal audio characteristics of the motor based on the normal and abnormal audio data of the test collector, inputting the normal and abnormal audio characteristics of the motor as test input characteristics into a trained self-encoder integrated learning model, obtaining test output characteristics of each self-encoder in the self-encoder integrated learning model, namely the reconstructed characteristics, calculating to obtain the mean square error of the test input characteristics and the test output characteristics of each self-encoder, and taking the mean square error as reconstruction errors, and carrying out weighted summation on the reconstruction errors to obtain an abnormal score. The method comprises the steps of S104, S2 and the like, wherein the mean square error of the training input characteristic and the training output characteristic of each self-encoder is calculated and obtained in the training process and is used as a training error, the reciprocal of each training error is obtained, the weight value is normalized and obtained in an integrated manner, the mean square error of the test input characteristic and the test output characteristic of each self-encoder is calculated and obtained in the testing process and is used as a reconstruction error, the weighted summation is carried out on the reconstruction error to obtain an anomaly score, and the motor anomaly detection precision is improved.

See the description of the detailed description section.

Drawings

FIG. 1 is a flow chart of the present invention;

FIG. 2 is a data flow diagram of the present invention;

FIG. 3 is a data flow diagram of a self-encoder ensemble learning model during training;

FIG. 4 is a data flow diagram of a self-encoder ensemble learning model during testing.

Detailed Description

According to the technical problem of low motor abnormality detection precision, the application provides an improved industrial motor equipment abnormality detection method based on a self-encoder integrated learning model, namely self-encoder autoencoder integrated learning, for detecting motor abnormality sounds of industrial equipment.

The self-encoder AU is a common unsupervised learning model and can be used in a plurality of fields such as anomaly detection, image generation and the like. The self-encoder AU maps the characteristics according to the weight value, the threshold value and the activation function of the neuron, and has stronger fitting and generalization capabilities. Since the distribution of the abnormal data is different from that of the normal data, the self-encoder AU is trained by the normal data, so that the self-encoder AU has smaller reconstruction errors for the normal data and larger reconstruction errors for the abnormal data.

However, since the weights and the thresholds are randomly initialized during the training of the self-encoder AU, the training result is unstable, and the model has a large variance. The ensemble learning can effectively reduce the variance of the model, and thus employs the integration of a plurality of self-encoders AU to reduce the variance of the model. Meanwhile, due to instability of self-encoder AU training, different self-encoder AU training has merits, so that weight is set according to training errors of the self-encoder AU, and a plurality of self-encoders AU are weighted and summed, thereby providing an improved self-encoder integrated learning model.

The main steps of obtaining the improved self-encoder integrated learning model comprise feature extraction, training of a plurality of AU models, obtaining of integrated weights, and obtaining of anomaly scores through weighted summation.

During training, the normal data of the training set are subjected to feature extraction by the feature extraction module, then are input into a plurality of AUs for training, and the trained AUs are stored. And determining the integrated weight according to the training errors of the AU models.

During testing, the normal and abnormal data of the test set are subjected to feature extraction by the feature extraction module, then are input into the trained AU models to obtain the reconstruction errors of the test set of the AU models, and finally the abnormal score of the test set is obtained by weighting and summing according to the integrated weight.

The following description of the embodiments of the present application will be made clearly and completely with reference to the accompanying drawings, in which it is apparent that the embodiments described are only some embodiments of the present application, but not all embodiments. The following description of at least one exemplary embodiment is merely exemplary in nature and is in no way intended to limit the application, its application, or uses. All other embodiments, which can be made by those skilled in the art based on the embodiments of the application without making any inventive effort, are intended to be within the scope of the application.

In the following description, numerous specific details are set forth in order to provide a thorough understanding of the present application, but the present application may be practiced in other ways other than those described herein, and persons skilled in the art will readily appreciate that the present application is not limited to the specific embodiments disclosed below.

As shown in fig. 1 to 4, the invention discloses a method for detecting a motor based on a self-encoder integrated learning model, which comprises the steps of training in step S1 and testing in step S2 to obtain an anomaly score.

As shown in fig. 1, is a flow chart of the present invention.

Step S1: training;

As shown in fig. 2, during the training process, the normal audio characteristics of the motor are obtained based on the extraction of the normal audio data of the training current collector; the method comprises the steps of obtaining a self-encoder integrated learning model, wherein the self-encoder integrated learning model comprises a plurality of self-encoders, namely first to nth self-encoders AU ₁ to AUn; obtaining a loss function, and calculating the weight and the threshold of each self-encoder in the self-encoder integrated learning model based on the loss function, the optimization function Adam, the gradient descent method and the back propagation method to obtain a trained self-encoder integrated learning model; and inputting the normal audio characteristics of the motor into a trained self-encoder integrated learning model as training input characteristics, obtaining the training output characteristics of each self-encoder in the self-encoder integrated learning model, calculating to obtain the mean square error of the training input characteristics and the training output characteristics of each self-encoder, taking the mean square error as training errors, obtaining the reciprocal of each training error, and normalizing to obtain the integrated weight.

The specific division of the step S1 comprises the steps of extracting and obtaining normal audio characteristics of the motor from a training set in the step S101, obtaining an integrated learning model of the self-encoder in the step S102, training the integrated learning model of the self-encoder in the step S103 and obtaining integrated weights in the step S104.

Step S101: extracting and obtaining normal audio characteristics of the motor from the training set;

The step S101 is formed by extracting the audio data based on the normal state of the training current collector to obtain the normal audio characteristics of the motor, which is described in detail below.

Because mel scale can well describe human perception of frequency, audio features of a motor often adopt mel spectrum functions to extract features, and the step of extracting mel spectrum, namely, the step S101, comprises the following steps: step S1011 short-time Fourier transform, step S1012mel triangular filter weighted summation, and step S1013 unit conversion to db, decibel.

Step S1011: a short-time Fourier transform;

Based on the normal audio data of the training collector, the time domain signal is subjected to framing, windowing and Fourier transformation to obtain a spectrogram.

Step S1012: weighted summation of mel triangular filters;

And (3) carrying out weighted summation on the spectrograms through a mel triangular filter, and further compressing characteristic dimensions to obtain the mel spectrograms, wherein the mel spectrograms are consistent with human hearing.

Step S1013: converting the units into db;

and converting the energy of the mel spectrogram into db to obtain the normal audio characteristics of the motor.

Step S102: obtaining an integrated learning model from the encoder;

the self-encoder integrated learning model includes a plurality of self-encoders, i.e., first through nth self-encoders AU ₁ through AUn, to form step S102. Detailed description is as follows.

Step S102 is a step of constructing an AU architecture.

As shown in fig. 4, the self-encoder includes an encoder and a decoder, the number of layers, the number of neurons and the activation function of the encoder in the self-encoder ensemble learning model are set, and the number of layers, the number of neurons and the activation function of the decoder in the self-encoder ensemble learning model are set, so that the architecture of the self-encoder AU is determined, and the self-encoder ensemble learning model is obtained.

Step S103: training a self-encoder integrated learning model;

As shown in fig. 3, the obtaining a loss function calculates a weight and a threshold of each self-encoder in the self-encoder integrated learning model based on the loss function, the optimization function Adam, the gradient descent method and the back propagation method, and obtains a trained self-encoder integrated learning model, thereby forming step S103. Detailed description is as follows.

Step S103 is a step of training a plurality of self-encoders AU, and step S103 includes determining a loss function in step S1031 and calculating parameters of the self-encoders AU in step S1032, which are described in detail below.

Step S1031: determining a loss function;

the loss function is expressed as:

(1)

In the formula (1): as a loss function; m is batch size scale; q is the dimension of the input feature; /(I) An ith feature of the jth input feature vector in batch; /(I)The ith feature of the jth output feature vector in the batch; j is E [1, m ], i is E [1, q ].

Step S1032: solving parameters of the self-encoder AU;

the weight and threshold of each self-encoder AU in the self-encoder ensemble learning model are obtained based on the loss function, the optimization function Adam, the gradient descent method and the back propagation method calculation (based on the loss function, gradient from the encoder AU is obtained by the gradient descent method, then gradient is optimized by the optimization function Adam, and finally back propagation).

Repeating the step S1032 to obtain a trained self-encoder integrated learning model, namely obtaining a plurality of trained self-encoders AU.

Step S104: obtaining an integrated weight;

And inputting the normal audio characteristics of the motor as training input characteristics to a trained self-encoder integrated learning model, obtaining training output characteristics of each self-encoder in the self-encoder integrated learning model, calculating and obtaining mean square error of the training input characteristics and the training output characteristics of each self-encoder and taking the mean square error as training errors, obtaining reciprocal of each training error, normalizing and obtaining integrated weights, and forming step S104. Detailed description is as follows.

And (3) inputting the normal audio characteristics of the motor obtained in the step (S101) as input characteristics into the trained self-encoder integrated learning model obtained in the step (S103), obtaining the output characteristics of each self-encoder AU in the self-encoder integrated learning model, calculating the mean square error of the input characteristics and the output characteristics of each self-encoder AU to obtain the inverse of each training error as training error, and normalizing and obtaining the integrated weight.

As shown in fig. 3, the training error is calculated from the mean square error of the input and output characteristics of the encoder AU in the training set, and the training error is shown in formula (2).

(2)

In the formula (2): Is a training error function; k is the training sample size; q is the dimension of the input feature; /(I) An ith feature of a jth input feature vector in the training set; /(I)The ith feature of the jth output feature vector in the training set; j is E [1, k ], i is E [1, q ].

Generally, the smaller the error of the training set, the better the self-encoder AU effect, so taking the inverse error of the training set and then normalizing to be the integrated weight, as shown in equation (3).

(3)

In the formula (3): i is the integrated weight of the ith self-encoder AU, i is [1, n ]; n is the number of trained self-encoders AU.

Step S2: testing to obtain an anomaly score;

As shown in fig. 2, during the test, the normal and abnormal audio features of the motor are obtained based on the extraction of the normal and abnormal audio data of the test collector, the normal and abnormal audio features of the motor are input into the trained self-encoder integrated learning model as the test input features, the test output features of each self-encoder in the self-encoder integrated learning model, namely the reconstructed features, are obtained, the mean square error of the test input features and the test output features of each self-encoder is obtained through calculation and is used as the reconstruction error, and the weighted summation is performed on the reconstruction errors to obtain the abnormal score. Detailed description is as follows.

As shown in fig. 4, for the input samples of the test set, the samples are firstly subjected to feature extraction, then are input into an AU model of the self-encoder to obtain reconstructed features, reconstruction errors are obtained according to the input features and the output features, and then the weighted summation is performed on the reconstruction errors to obtain an anomaly score, wherein the greater the anomaly score is, the more the samples are abnormal. The formula for anomaly score is as follows:

(4)

In the formula (4): score is an anomaly score; n is the number of self-encoders AU in the trained self-encoder integrated learning model; the weight of the self-encoder AU in the j-th trained self-encoder integrated learning model is j epsilon [1, n ]; an ith feature that is an input feature vector; /(I) An ith feature which is an output feature vector; q is the dimension of the feature.

The application adopts the public dataset MIMII DATASET to verify the superiority of the improved self-encoder integrated learning model, namely autoencoder integrated learning. MIMII DATASET is a reliable dataset for investigation and inspection of faulty industrial machines. It contains four sounds produced by industrial machines, namely valves, pumps, fans and skid rails. Each machine includes seven separate product models, but only four models are disclosed, id0, id2, id4, and id6, respectively. The data for each model contains normal sound from 5000 seconds to 10000 seconds and abnormal sound of about 1000 seconds. To simulate real life scenes, various abnormal sounds such as pollution, leakage, rotational unbalance, and track damage are recorded. In addition, the data set mixes with the background noise and machine sounds recorded by multiple real factories. Sound was recorded by an eight-channel microphone array at a sampling rate of 16 kHz, 16 bits per sample.

According to the application, motor sound of an id0 pump is selected as experimental data, a training set and a testing set are cut by normal data according to the proportion of 8:2, and abnormal data are placed in the testing set.

Extracting the characteristics of the audio: and inputting audio data, and obtaining a stft spectrogram, namely Short-Time Fourier Transform, by taking 1024 points as a length and carrying out fft, namely fast Fourier transform, and 512 points as step sizes. And (3) carrying out characteristic compression on the spectrogram by taking the number of the mels as 64 to further obtain a mel frequency spectrum, and converting the energy into decibel units. Since the acquired signal is a steady state signal, successive 5 frames of the mel spectrum are stretched into a eigenvector with dimensions 64 times 5 equal to 320.

Constructing an AU model architecture: since the feature vector is 320-dimensional, both the input and output of the AU are 320-dimensional. The dimensions of the intermediate hidden layer should be lower than the dimensions of the input and output, as compared to the dimensions of the input and output, to compress the embedded features. The neuron topological structure of the AU model adopted by the application is as follows: an input dimension 320 dimension; the middle 5 hidden layers have dimensions of 64,8, 64, 64 respectively; the output dimension 320 is dimension.

Training AU model: taking Adam as an optimization function and equation 1 as a loss function, batch size is 64, and taking early stopping as an iteration stopping condition to train a model. Repeating training for multiple times, and storing the trained 5 AU models.

Solving an anomaly score: and (3) obtaining the training set error of each AU model according to the formula (2), and obtaining the weight according to the formula (3) to obtain the weight vectors of the trained 5 AU models. And then according to the formula 4, obtaining the abnormal score of the audio.

And (3) experimental results show that:

The test is to compare the existing self-encoder, namely AU model, integrated AU model sense-AU abbreviated e-AU, and the improved self-encoder integrated learning model, namely integrated AU model improved-sense-AU abbreviated i-e-AU, and judge the quality of the model by taking the area under curve AUC, namely area under curve, as an index.

The motor sound data of mimii pumps can be divided into three types of 6db,0db and 6db according to the signal to noise ratio. In general, the higher the signal-to-noise ratio, the higher the accuracy of the model.

For data with 6db signal-to-noise ratio, the experiment is repeated for 5 times to obtain 5 e-AU and i-e-AU models, and each e-AU and i-e-AU integrates 5 independent AU models respectively. The experimental results are shown in table 1.

Table 1: experimental result table for snr=6db dataset

As can be seen from Table 1, the average AUC of AU, e-AU, i-e-AU is 0.84840, 0.85800,0.86200, respectively; AU, e-AU, i-e-AU has an AU variance of 0.001213,4.96E-04,7.36E-04. Therefore, the e-AU and i-e-AU of the integrated learning are higher than AU in precision and the variance is smaller than AU, so that the integrated learning can effectively improve the precision and stability of the model. The accuracy of the i-e-AU is slightly higher than that of the e-AU, and the variance is slightly higher than that of the e-AU.

For data with 0db signal to noise ratio, the experiment was repeated 5 times to obtain 5 e-AU and i-e-AU models, each of which integrates 5 independent AU models. See table 2.

Table 2: experimental result table for snr=0db dataset

As can be seen from Table 2, the average AUC of AU, e-AU, i-e-AU is 0.72640,0.73400,0.73600, respectively; AU, e-AU, i-e-AU has an AU variance of 0.000239,1.04E-04,1.84E-04. Therefore, the e-AU and i-e-AU of the integrated learning are higher than AU in precision and the variance is smaller than AU, so that the integrated learning can effectively improve the precision and stability of the model. The accuracy of the i-e-AU is slightly higher than that of the e-AU, and the variance is slightly higher than that of the e-AU.

For data with the signal to noise ratio of-6 db, the experiment is repeated for 5 times, so that 5 e-AU and i-e-AU models are obtained, and each e-AU and i-e-AU integrates 5 independent AU models respectively. The experimental results are shown in table 3.

Table 3: experiment result table of SNR= -6db data set

As can be seen from Table 3, the average AUC of the AU, e-AU, i-e-AU is 0.73080,0.73400,0.73600, respectively; AU, e-AU, i-e-AU with AU variance 0.000279,6.40E-05,2.40E-05. Therefore, the e-AU and i-e-AU of the integrated learning are higher than AU in precision and the variance is smaller than AU, so that the integrated learning can effectively improve the precision and stability of the model. The accuracy of the i-e-AU is slightly higher than the e-AU in terms of accuracy, and the variance is slightly lower than the e-AU.

Therefore, it can be concluded that the i-e-AU has improved accuracy compared to AU; the stability of the i-e-AU is improved compared with that of the AU, and the stability of the i-e-AU is consistent with that of the e-AU; so i-e-AU is the best model.

Claims

1. A method for detecting a motor based on a self-encoder integrated learning model is characterized by comprising the following steps of: the method comprises the following steps of: extracting and obtaining normal audio characteristics of the motor based on the normal audio data of the training current collector; step S102: obtaining a self-encoder integrated learning model, the self-encoder integrated learning model including a plurality of self-encoders; step S103: obtaining a weight and a threshold value of each self-encoder in the self-encoder integrated learning model, and obtaining a trained self-encoder integrated learning model; step S104: inputting the normal audio characteristics of the motor as training input characteristics into a trained self-encoder integrated learning model, obtaining the training output characteristics of each self-encoder in the self-encoder integrated learning model, calculating to obtain the mean square error of the training input characteristics and the training output characteristics of each self-encoder, taking the mean square error as training errors, obtaining the reciprocal of each training error, normalizing and obtaining the integrated weight; step S2: and extracting normal and abnormal audio characteristics of the motor based on the normal and abnormal audio data of the test collector, inputting the normal and abnormal audio characteristics of the motor as test input characteristics into a trained self-encoder integrated learning model, obtaining test output characteristics of each self-encoder in the self-encoder integrated learning model, namely the reconstructed characteristics, calculating to obtain the mean square error of the test input characteristics and the test output characteristics of each self-encoder, and taking the mean square error as reconstruction errors, and carrying out weighted summation on the reconstruction errors to obtain an abnormal score.

2. The method for detecting a motor based on a self-encoder integrated learning model according to claim 1, wherein: the step S101, the step S102, the step S103, and the step S104 form a step of training in the step S1, and the step S101 specifically includes the following steps: based on the normal audio data of the training collector, framing, windowing and Fourier transforming the time domain signal to obtain a spectrogram; step S1012: weighting and summing the spectrograms through a mel triangular filter, and compressing characteristic dimensions to obtain a mel spectrogram; step S1013: and converting the energy of the mel spectrogram into db to obtain the normal audio characteristics of the motor.

3. The method for detecting a motor based on a self-encoder integrated learning model according to claim 1, wherein: in the step S102, the self-encoder includes an encoder and a decoder encoder, the number of layers, the number of neurons and the activation function of the encoder in the self-encoder integrated learning model are set, and the number of layers, the number of neurons and the activation function of the decoder encoder in the self-encoder integrated learning model are set, so that the architecture of the self-encoder is determined, and the self-encoder integrated learning model is obtained.

4. The method for detecting a motor based on a self-encoder integrated learning model according to claim 1, wherein: in the step S103, the steps of obtaining the weight and the threshold value of each self-encoder in the self-encoder integrated learning model include a step S1031 and a step S1032, the step of obtaining the trained self-encoder integrated learning model is a step S1033,

Step S1031: determining a loss function;

(1)

5. The method for detecting a motor based on a self-encoder integrated learning model according to claim 1, wherein: in said step S104, a training error is calculated according to equation (2), an integrated weight is calculated according to equation (3),

(2)

(3)

6. The method for detecting a motor based on a self-encoder integrated learning model according to claim 1, wherein: in the step S2, an anomaly score is calculated according to equation (4),

(4)