Research On Fault Prediction Algorithm Of Hydraulic System Based On Multi-task Deep Learning

As an important part of construction machinery,the hydraulic system has a direct impact on the working condition and performance of the equipment.Therefore,it is of great significance to monitor the condition of each key component of the hydraulic system.In recent years,deep learning algorithms have been widely used in the field of state monitoring due to their powerful feature extraction ability and self-adaptability.However,deeping learning only focuses on a single component in hydraulic system fault prediction,which cuts off the correlation between different components,making it difficult to completely describe a wide range of equipment states such as type and degree of component faults,and to accurately monitor similar fault states and complex faults in monitored components.Considering that there will be balance and imbalance between the data collected in the real world,it is necessary to study different algorithms for monitoring.Therefore,based on the multi-task network,this paper proposes different multi-task network algorithms for simultaneous monitoring of the fault state of multiple components such as valves and accumulators of the hydraulic system in view of the balance and imbalance of data.The research contents and innovations are as follows:(1)Considering that there is a lot of redundant information in the signal data collected by multi-sensor when carrying out multi-task network learning,it is easy to overload the network information and result in poor network effect.In addition,the difficulty of each task is different,which also results in different learning speeds.Often,simple tasks have been fitted and difficult tasks have just begun to work.In order to solve these problems,a attentionmultitask adaptive network model(CS-ADMLT+)is proposed for fault prediction under data balance.This method uses the attention mechanism to assign different weights to the input sensor information,so that the network pays attention to more useful sensor information,thus reducing the overload of network information.For each task learning balance problem,the Entropy Weight Method is introduced and the calculation method is improved,thus a new adaptive task weight allocation method is proposed,which adjusts the weight of each task according to the degree of dispersion of the loss value of the task.The test case shows that the multi-task network achieves much better results than the single-task method,with an average accuracy increase of 1.7%,in which the accumulator monitoring accuracy is improved by more than 7%,and the accumulator accuracy is increased by 2% again by adding attention mechanism and improved Entropy Weight Method.(2)In order to solve the imbalance of samples in various states of hydraulic system data in industrial environment,an adaptive cost sensitive matrix multi-task learning algorithm(CS-MLT)is proposed in this paper.The new loss function is constructed by adding the adaptive cost sensitive matrix to the loss function to deal with this imbalance.The adaptive cost sensitive matrix is composed of sample cost matrix and misclassification cost matrix.The sample cost matrix is used to balance the number of samples in each state,and the misclassification cost matrix is used to give different penalty coefficients when misclassification occurs between different states.The misclassification cost matrix will change with the number of iterations,thus realizing the adaptive adjustment of the cost sensitive method.Test cases show that,with the increase of data set imbalance,CS-MLT can still guarantee a high recall rate(recognition rate for a few classes)on the premise of better overall accuracy.(3)Combined with ATT-ADMLT+ and CS-MLT algorithm,the fault prediction software of hydraulic system is designed,in which the monitoring module and test module are mainly designed,and each function button of the two modules is designed according to actual requirements.