Research On Temperature Anomaly Detection Of Low-Speed Blower Based On Infrared Images

With the continuous development of automation industrial technology,blower as an important industrial equipment is widely used and requires long-term and highly reliable operation.The existing methods for detecting temperature anomalies in blowers still rely mainly on manual inspections,which have certain limitations and dangers when dealing with large and complex equipment and cannot achieve real-time monitoring.In order to solve this problem,this paper takes infrared imags of blowers as the main research object,and proposes a low-speed blower temperature anomaly detection algorithm based on infrared images,which focuses on the detection of two common types of anomalies,namely temperature anomaly changes and single-point anomalies,and combines semantic segmentation technology of deep learning to quickly locate abnormal areas.The specific work and innovation points are as follows:(1)To address the issue of interference with infrared data from the blower in practical usage scenarios,this paper proposes a robust segmentation algorithm for infrared images of the blower based on an improved Point Rend network.Firstly,the blower data is collected using an infrared imager and filtered and annotated to obtain a self-constructed dataset for infrared image segmentation of the blower.Then,an attack layer containing various common image attacks is designed and inserted after the input layer of the segmentation network.During training,the attack is randomly added to the images to enhance the robustness of the network.Finally,the SENet(Squeeze-andExcitation Networks,SENet)channel attention mechanism is embedded in the residual blocks of the feature extraction network,which dynamically allocates the weights of feature channels to enhance the representation capability of the segmentation network and further improve the segmentation performance of the network.Experimental results show that the improved network has higher robustness and can accurately segment the components of infrared blower images under common attacks.The mean segmentation accuracy reaches over 97%,and the MIo U(Mean Intersection over Union,MIo U)is also maintained at over 83%.(2)To address the problem of difficultly solving personalized user data needs,this paper proposes a blower component temperature anomaly detection algorithm based on hierarchical strategy,which consists of four modules in total.The first module is the temperature data extraction and pre-processing module,which extracts the temperature change curve of the corresponding component based on the infrared video data of the blower machine and the improved semantic segmentation network in subsequent processing,making it easier to identify anomalies and locate problematic components.The second module is the multi-data analysis module,which solves the efficiency problem of processing a large amount of historical data at once by using the K-Means++algorithm for clustering.The third module is the single data global analysis module,which classifies samples based on the KNN(K-Nearest Neighbor,KNN)algorithm after sampling single data,and judges the abnormality of the data based on the classification results,achieving fine detection of single data.The fourth module is the single data local analysis module,which uses multi-order difference combined with sliding window to analyze the trend of local data changes and quickly give the degree of anomaly,solving the real-time detection problem.The experimental results show that the algorithm can effectively detect temperature change anomalies,and the four modules used in combination can process data of any scale and length,achieving analysis from coarse to fine,with an average detection accuracy of each module reaching 90% or more.(3)To address the problem of single-point anomalies and the scarcity of abnormal samples in the blower,this paper proposes a blower single-point anomaly detection algorithm based on unsupervised learning and feature design.Firstly,the image sequence set constructed by using infrared video data of the blower is used to train the unsupervised video anomaly detection network MPN(Meta Prototype Network,MPN),which learns the normal patterns in the data to remove the interference caused by normal dark spots,bright spots,and slight perspective changes,and provides the anomaly distribution map for each image.Then,combined with the improved semantic segmentation network,the positions with larger abnormal values in the anomaly distribution map are tracked,and the repeated positions are recorded to obtain the rough detection result.Finally,three types of single-point anomaly features are designed based on the pixel characteristics in the neighborhood of the abnormal point,which are the trend feature of local variance change,the feature of local edge four-direction distance,and the feature of local gradient field.The rough detection result is screened according to the three features to obtain the fine detection result and segment the abnormal points.The experimental results show that even in the case of using unlabeled small sample data,accurate detection of single-point anomalies can be achieved,with an average detection accuracy of over 94%.