The Deep Learning-Based Muscle Fatigue Detection Model And Its Interpretability Research

After long-term repetitive movements of human muscles,they gradually enter a state of fatigue and affect their working state,which can pose significant safety hazards in specific scenarios,such as operational fatigue,intelligent prosthetic wear fatigue,and training fatigue.Therefore,accurate identification and evaluation of muscle state have important applications and socio-economic value in fields such as rehabilitation medicine,wearable devices,and auxiliary training.As a safe and non-invasive physiological signal acquisition method,surface electromyography(sEMG)signal detection technology is widely used in muscle fatigue detection tasks.The use of statistical methods and traditional machine learning methods to complete fatigue state detection requires a large amount of manual operation,lacking real-time and accuracy.Moreover,most researchers only focus on a specific type of muscle for fatigue detection,and the data processing method is limited to extracting or improving several time-frequency domain features from sEMG data,which does not have universal applicability.However,deep learning models have shown outstanding performance in the field of intelligent analysis and evaluation,and can be well transferred and applied to the field of muscle fatigue assessment.However,due to their black box characteristics,they lack certain reliability.Therefore,in response to the above issues,this article conducted research on fatigue assessment models and their interpretability methods:The experiment collected the sEMG signals of 8 different muscles from 30 volunteers without cervical and lumbar diseases during simulated flight training,and generated a data set with a total of 280000 samples including non fatigue period,transition period and fatigue period after preprocessing.Design multiple data processing methods and corresponding deep learning models,and conduct in-depth research on muscle fatigue assessment technology based on deep learning sEMG signals from multiple perspectives.At the data structure level,three types of data structures were designed:time series,feature series,and time-frequency maps.At the classification model level,a universal model TFS_ResNe was designed that is suitable for the three data structurest and a specific task model applicable to a single data structure were used to compare and study the classification performance of different data structures using the same depth model and a specific task model.Further analyze the decision-making logic of the evaluation model,conduct interpretability research on specific deep learning models,quantify the correlation distribution between data features and decision results,and increase credibility.The experimental results show that the corresponding model can achieve an accuracy of 97.2%～99.2%in evaluating muscle fatigue status using different data structures.The TFS designed in this paper_ResNet(T-time,F-feature,S-Spectrum)universal architecture learns the common relationships between different muscles,has strong universality,and is superior to traditional methods.And when using feature sequences as a data processing method,this evaluation model not only achieves an accuracy of 97.38%,but also has the advantages of less time consumption and memory consumption,and has high comprehensive performance of real-time operations,suitable for small and portable device applications.Furthermore,based on the principle of gradient integration,this article delves into the interpretability of the model and obtains the correlation distribution between different features and classification results in algorithm decision-making,enhancing the reliability of its decision-making mechanism in muscle fatigue detection tasks and providing possibilities for algorithm optimization.