CN-121977827-A - Roller chain anomaly detection method and system based on sensorless multi-step deep decomposition network

CN121977827ACN 121977827 ACN121977827 ACN 121977827ACN-121977827-A

Abstract

The invention provides a roller chain abnormality detection method and system based on a sensorless multi-step deep decomposition network, and the method and the system comprise the steps of collecting multi-mode operation signals from a SEW driver of a roller chain driving motor and preprocessing the multi-mode operation signals, wherein the multi-mode operation signals comprise torque signals and motor encoder position signals, fusing position guide segmentation with the multi-step deep decomposition network to construct the sensorless multi-step deep decomposition network, obtaining health indexes of the roller chain based on the sensorless multi-step deep decomposition network and the preprocessed multi-mode operation signals, and adopting a WKNN uncertainty perception abnormality detection method to detect the abnormality of the roller chain based on the health indexes to obtain an abnormality detection result and an uncertainty interval. The invention solves the problems of difficult deployment of the sensor in the roller chain CM, poor trend of Health Index (HI), insufficient robustness of anomaly detection and the like.

Inventors

LIU QIANG
YU JINLONG
CHEN ZHUYUN
CAI YAN
LIN YIJIAN
LIN HONGQI
LI ZEHAO

Assignees

广东工业大学

Dates

Publication Date: 20260505
Application Date: 20260126

Claims (8)

1. The roller chain abnormality detection method based on the sensorless multi-step deep decomposition network is characterized by comprising the following steps of: collecting multi-mode operation signals from an SEW driver of a roller chain driving motor, and preprocessing the multi-mode operation signals, wherein the multi-mode operation signals comprise torque signals and motor encoder position signals; merging the position-guided segmentation and multi-step deep decomposition network to construct a sensorless multi-step deep decomposition network; Based on the sensorless multi-step deep decomposition network and the preprocessed multi-mode operation signals, health indexes of the roller chain are obtained; Based on the health index, adopting a WKNN uncertainty perception abnormality detection method to perform abnormality detection on the roller chain, and obtaining an abnormality detection result and an uncertainty section.
2. The method of claim 1, wherein the method of preprocessing the multi-modal operating signal comprises: identifying and eliminating abnormal values in the torque signals by adopting a box diagram method, and carrying out minimum-maximum normalization on the torque signals with the abnormal values eliminated to obtain preprocessed torque signals; calibrating the motor encoder position signal based on the motor encoder resolution to obtain a preprocessed motor encoder position signal; Based on the preprocessed torque signal and the preprocessed motor encoder position signal, a preprocessed multi-mode operation signal is obtained, and the preprocessed multi-mode operation signal is divided into a test set and a training set.
3. The method of claim 2, wherein the method of obtaining a health indicator of the roller chain comprises: based on the preprocessed motor encoder position signal, performing time domain segmentation on the preprocessed torque signal to obtain a torque subsequence; performing fast Fourier transform on the torque subsequence to obtain a frequency domain torque characteristic; decomposing the frequency domain torque characteristic into a trend component and a season component based on a multi-step deep decomposition network; The trend component and the season component are respectively and independently predicted through a linear projection layer, so that a prediction result is obtained; And normalizing the prediction result to obtain the health index of the roller chain.
4. The method of claim 2, wherein the method of anomaly detection of the roller chain comprises: Based on a KNN algorithm, aiming at the health index of each test sample, calculating the Euclidean distance between the test sample and the health index of each health sample in the training set to form a corresponding distance set; Based on the Euclidean distance, carrying out initial abnormality judgment by using the constructed three judgment conditions to obtain an initial abnormality judgment result; And carrying out weighted aggregation on the initial abnormality discrimination result to obtain a final abnormality detection result and an uncertainty interval.
5. The method of claim 4, wherein the three criteria include: the condition C1 is that when the average distance of the distance set is larger than a statistical threshold value determined by the distance set corresponding to the healthy sample in the training set, the distance set is judged to be abnormal; A condition C2, wherein when the proportion of the distances larger than the statistical threshold in the distance set is larger than or equal to 80%, the distance set is judged to be abnormal; And C3, judging that the device is abnormal when the minimum distance in the distance set is larger than the maximum distance in the distance set corresponding to the healthy sample in the training set.
6. The method of claim 4, wherein the uncertainty interval comprises a lower detection bound and an upper detection bound, wherein the lower detection bound corresponds to an early warning point and the upper detection bound corresponds to a determination of an outlier.
7. A sensorless multi-step deep decomposition network-based roller chain anomaly detection system for implementing the method of any one of claims 1-6, comprising: the system comprises a signal acquisition module, a motor encoder and a motor encoder, wherein the signal acquisition module is used for acquiring multi-mode operation signals from an SEW (sequence-oriented) driver of a roller chain driving motor and preprocessing the multi-mode operation signals, wherein the multi-mode operation signals comprise torque signals and motor encoder position signals; The network model construction module is used for fusing the position guide segmentation and multi-step deep decomposition network to construct a sensorless multi-step deep decomposition network; The health index acquisition module is used for acquiring the health index of the roller chain based on the sensorless multi-step deep decomposition network and the preprocessed multi-mode operation signals; the abnormality detection module is used for carrying out abnormality detection on the roller chain by adopting a WKNN uncertainty perception abnormality detection method based on the health index to obtain an abnormality detection result and an uncertainty section.
8. The system of claim 7, wherein the health indicator acquisition module comprises: the torque signal segmentation unit is used for carrying out time domain segmentation on the preprocessed torque signal based on the preprocessed motor encoder position signal to obtain a torque subsequence; the Fourier transform unit is used for carrying out fast Fourier transform on the torque subsequence to obtain a frequency domain torque characteristic; A feature decomposition unit for decomposing the frequency domain torque feature into a trend component and a season component based on a multi-step deep decomposition network; The independent prediction unit is used for independently predicting the trend component and the season component through the linear projection layer respectively to obtain a prediction result; And the normalization unit is used for normalizing the prediction result to obtain the health index of the roller chain.

Description

Roller chain anomaly detection method and system based on sensorless multi-step deep decomposition network Technical Field The invention belongs to the technical field of roller chain state monitoring (CM) and predictive maintenance (PdM), and particularly relates to a roller chain abnormality detection method and system based on a sensorless multi-step deep decomposition network. Background In the context of intelligent manufacturing and industry 4.0, the reliability and operating efficiency of the device directly determine the overall equipment performance, and state monitoring (CM) and predictive maintenance (PdM) are core technologies for guaranteeing safe operation of the device and reducing unplanned shutdown cost, which have become research hotspots in the industry field. The rotary machine is used as a key carrier for intelligent manufacture, and the CM technology of the rotary machine is widely applied to parts such as bearings, gears, machine tools and the like, but the research of the CM technology of a roller chain serving as an industrial transmission core component is still in a hysteresis state due to the structural characteristics and the limitation of the running environment. The roller chain is widely applied to the fields of conveying machinery, agricultural equipment, mining machinery and the like by virtue of the advantages of no slip, wide transmission distance, severe environment resistance and the like, but is easy to cause progressive degradation due to abrasion, dislocation, insufficient lubrication and the like in long-term operation, and if the roller chain is not detected in time, faults such as chain breakage and the like can be caused, so that serious economic loss is caused. Existing CM technology for roller chains can be divided into two categories, traditional sensor driving and data driving. The traditional sensor driving method relies on vibration and acoustic signals, extracts characteristics through time domain (RMS, kurtosis), frequency domain (FFT), time-frequency domain (STFT, wavelet transformation), but is limited by sensor deployment and noise, and the data driving method, such as a LSTM, transformer-based deep learning model, can capture time sequence dependence, but needs a large amount of labeling data, and aims at the defect of a special model of a roller chain, and meanwhile omits physical association of 'position-torque', so that smoothness and trend of health indexes are insufficient. Three major problems to be solved in the existing roller chain CM technology are: (1) Physical sensors rely on and deploy limitations traditional CM technology relies on physical devices such as contact accelerometers, acoustic sensors, torque sensors, etc. For roller chains, on the one hand, the dynamic motion characteristics of the roller chains lead to the difficulty in stably installing contact sensors, and on the other hand, the roller chain system with long-distance transmission needs a large amount of sensor coverage, and is high in cost and complex in maintenance. In addition, the sensor signal is susceptible to industrial environmental noise, such as dust and vibration interference, resulting in reduced data quality and increased subsequent analysis errors. (2) The gap of the roller chain CM research is remarkable, the existing CM research mainly focuses on parts such as bearings, gears and the like, for example, the number of the bearings and gear CM related publications in 2000-2024 is far more than that of the roller chain, and the special technology for the roller chain is very few. The polygonal effect of the roller chain further increases the complexity of vibration and torque signals, and the traditional health indexes (such as RMS and kurtosis) aiming at the bearing have poor tendency and large fluctuation on the roller chain, so that the degradation process is difficult to effectively characterize. (3) The robustness of health assessment and anomaly detection is insufficient, the selection of the existing health indexes depends on experience, the trend consistency of the health indexes under different working conditions is poor, the existing anomaly detection method, such as 3STD, SVM and isolated forest, is lack of quantification of uncertainty, is easy to be subjected to noise interference to cause false alarm or missing report, and has weak model generalization capability, so that the method is difficult to adapt to different running conditions of roller chain speed change, load change and the like. Disclosure of Invention The invention provides a roller chain abnormality detection method and system based on a sensorless multi-step deep decomposition network, and aims to solve the problems of difficult sensor deployment, poor Health Index (HI) trend, insufficient abnormality detection robustness and the like in a roller chain CM. In order to achieve the above object, the present invention provides the following solutions: the roller chain abnormality detec