Research On Industrial Defect Detection Method Based On Improved Spectral Residual Algorithm

With the increasing integration of smart hardware products in today’s industry,many discrete manufacturing companies are faced with new challenges,particularly in the detection of workpiece anomalies during product production.Traditional methods of workpiece inspection rely solely on physical inspection methods such as sensors,which can result in inaccurate inspection results.For this reason,many companies have optimized their production processes,auxiliary tools,source materials and many other aspects to varying degrees,but the results have not improved significantly.The reason for this is that most physical inspection methods only use simple rules of judgement,which ignore the useful information inherent in the data and are not refined enough to attain the desired results.Nowadays,with the development and successful application of machine learning in various fields,this thesis attempts to use machine learning methods to compensate for the shortcomings of traditional detection methods for discrete manufacturing.This thesis firstly proposes a significant anomaly detection algorithm based on the partial Fourier transform,secondly introduces the algorithm into the Transformer-based anomaly detection framework to further enhance the anomaly detection effect,and finally designs an anomaly monitoring service system for discrete manufacturing enterprises based on this anomaly detection framework.The significant anomaly detection algorithm uses a Spectral residual algorithm and a phase spectral transform algorithm to extract anomalous features from both frequency domain amplitude and frequency domain phase,and uses a partial Fourier transform algorithm to improve the overall computational efficiency.On real discrete manufacturing production data,the computational efficiency of the algorithm can be improved by about 4.5 times.The significant anomaly detection algorithm is integrated into the Transformer-based anomaly detection framework as a feature extraction means to further representatively learn anomaly features.In addition,the framework introduces the ST-Norm technique,which can comprehensively and efficiently learn anomaly information from both high-frequency and local components of the raw data.Experimental results show that the anomaly detection network framework proposed in this thesis performs better than today’s mainstream anomaly detection algorithms,with an average F1-score improvement of about 1.0%.Finally,for the purpose of reliability and universality,this thesis integrates the anomaly detection framework into the enterprise anomaly monitoring service system.In a practical business environment,the anomaly detection framework is accurate,efficient and highly generalization,helping customers to identify potential defective products in a timely manner,effectively reducing production costs for enterprises and facilitating the digitalisation of production in the discrete manufacturing industry.