Research On Gearbox Abnormal Sound Detection Based On Unsupervised Adversarial Domain Adaptation One Class Classification

Large industrial rotating equipment is an important pillar of complex industrial system and the cornerstone of promoting the development of industrial production.The gearbox is the most critical functional component of rotating equipment,but the gearbox often fails due to its small structure,long-term overload and other reasons.The failure of the gearbox may cause mechanical shutdown and economic loss of the factory,or personal injury accidents.Therefore,it is particularly important to carry out gearbox abnormal detection based on data-driven.With the development of deep learning,the research on gearbox fault diagnosis under the framework of deep learning based on massive working data is more and more in-depth.For complex industrial systems,the requirement of extensive representation of training set data limits the health monitoring of gearbox.The reasons are as follows: first,the equipment failure rate of complex industrial systems is low,but the possible failure types cannot be described in detail.Second,the operation status of complex industrial system equipment will change with the change of specific production demand.The previous research on Gearbox Fault Diagnosis Based on fault classification hardly considered the above two problems at the same time.Therefore,this paper studies the method of gearbox abnormal detection from the perspective of unsupervised abnormal detection.It mainly includes the following contents:1.Aiming at the problem that the training data contains only one class of samples(normal sound data)and the test data contains unknown class anomalies,an unsupervised one classification network based on autoencoder is proposed as the baseline system.By analyzing the characteristics of sound signal in time domain,frequency domain and time-frequency domain,it is proposed to use the time-frequency domain characteristics that not only contain time series information,but also represent the amplitude and frequency information of sound signal as the abnormal detection input data.At the same time,compare MFCC and log-Mel spectrum,and select the log-Mel spectrum which is more representative of signal characteristic information as the time-frequency domain feature representation scheme.Based on the idea of feature reconstruction of autoencoder,an unsupervised abnormal detection model based on normal samples is established.2.Aiming at the problem of limited feature extraction ability of single-layer self encoder and time-consuming back propagation of stacked autoencoder,an unsupervised multi-layer extrem learning machine autoencoder network is proposed.Based on the autoencoder,combined with the extrem learning machine that does not rely on back propagation,a single-layer feedforward extrem learning machine autoencoder(ELM-AE)is formed.With regard to the defect that ELM-AE is only suitable for shallow feature extraction and the weak ability of complex feature extraction of high-dimensional data,the network stacking is further deepened,and OCC-ELM is used as the abnormal detection scheme at the last layer of the stacking network,forming a multi-layer extrem learning machine autoencoder one class classification network(MELM-AE-OCC)which can adapt to deep feature extraction.3.Aiming at the problem that MELM-AE-OCC network can not deal with the domain shift caused by different data distribution(caused by gearbox operating conditions),but the abnormal detection task is the same,one class classification network(UADA-OCC)is proposed.The MELM-AE-OCC network is further improved,a new inter domain point pair distance conservation loss suitable for unsupervised domain adaptation is proposed for the feature alignment task.At the same time,the adversarial training based on Wasserstein discriminator is used for the feature extraction layer to minimize the classification loss of the sample classifier and maximize the discrimination loss of the domain discriminator,making the inter domain feature representation more domain invariant.