Design And Application Of Adaptive Multi-task Learning Algorithm For Scenarios With Unknown Task Relationship

In the current age of data,machine learning algorithms have become a bridge between data and decision making.In practical applications,different machine learning tasks have relatively stable algorithms for solving.However,traditional machine learning only uses a single task for its own training and prediction,ignoring the shared benefits that other similar tasks may bring.Multi-task learning is proposed to efficiently utilize the shared information between multi-tasks to improve the overall prediction performance of the overall learning system.By providing additional effective training information for each individual machine learning task,the actual prediction performance of a single machine learning task is ultimately improved.??This paper mainly studies adaptive multi-task learning in a priori scenario of unknown task relationship,and the application of the algorithm in the prediction of the remaining life of multi-condition machining cutters.The existing multi-task learning algorithms can be divided into the following categories: multi-task learning methods based on feature transformation and feature selection,and low-rank based multi-task learning approach,multitask learning methods based on task clustering and task relationship learning,and multi-task deep learning method based on shared network structure.Compared with the single-task learning method,the performance of the above multi-task learning algorithm has been verified in multiple real-world scenarios.Clearly,deep structure based MTL can perfectly find features that are shared between tasks.But with the limit of the black-box character of the neural network and the small amount of the training data,it is hard to figure out the precise physical meaning of network parameters in the task level,leaving poor performance in the face of a large-scale number of fewshot tasks.Another one is the feature selection based approach,where by selecting features that are meaningful to the optimizing goal for multi-task learning,feature selection based MTL is born to be analyzed under more complex applications.But it can only work under the premise of similar tasks.Thus,when encountered with high-dimensional or large-scale applications unknown about the relationship between multiple tasks,either with their tremendous calculation overhead or difficulty to optimal clustering of different tasks,existing methods may not be adaptive enough for getting a precise solution.This paper first introduces the concept of the robust estimator,and for the problem that the robust estimator is difficult to optimize,a conversion method between the robust regularization term and the outlier process is proposed.Using continuous optimization adjusts the penalty function in the outlier process to achieve the convex approximation of the original robust estimator.Next,this paper summarizes the research status of multi-task learning,and its applicable scenarios are analyzed in detail.In order to solve problems such as difficulty in measuring task relationships and mutual influence between dissimilar tasks in large-scale unknown prior multi-task learning scenarios,a new robust multi-task learning algorithm based on robust estimators is proposed,which is named Robust Feature Selection and Continuous Clustering Multi-Task Learning(RFCMTL).This method adaptively cluster a massive number of tasks with approximately linear time complexity and simultaneously carry on selecting valuable features of tasks in the same cluster to acquire better regression performance.For the non-convex and non-smooth case of the objective function containing a robust estimator,it is converted into the form of an outlier process,and it is efficiently optimized by using the continuous alternating least square method.The algorithm proposed in this paper achieves the best regression prediction performance on multiple sets of domain standard datasets.In order to solve the problems of multiple complicated working conditions and difficulty to predict the status of a new cutter in the field of condition monitoring and forecasting scenarios,this paper introduces multi-task learning into the cutter remaining life prediction scenario.For the milling process of CNC machine tools,the inconsistency of the remaining life prediction function is more obvious under different working conditions.If the task information is used as a feature for single task learning,the task information only plays the role of sample space segmentation or function parameter accumulation in the parameter optimization process,and the useful information contained in the operation conditions is not deeply integrated.Robust estimator based adaptive multi-task learning can make the parameters of the prediction function converge more quickly and accurately under similar conditions,and the parameters between dissimilar conditions do not affect each other during the convergence process,thereby improving the overall rregression accuracy of life prediction tasks of milling cutters.In this study,multiple sets of cutter life-cycle data were used,and data noise reduction,feature extraction,and operating condition identification were performed to build a complete remaining useful life prediction system.Results of multiple controlled experiments prove that compared with the single-task learning model,the time complexity of the overall calculation is much lower than other learning models.Based on the robust estimator,the adaptive multi-task learning algorithm can provide a more accurate prediction of remaining useful life based on the real-time operating conditions and condition monitoring data of the cutter.Real-time operating conditions and condition monitoring data help the system to propose more accurate predictions of remaining life.In the scenario of cold start prediction of new cutters,the multi-task learning system can utilize useful information between similar operating conditions,and to some extent,solve the problems such as the lack of labeling and historical data of the new cutter.