Research On Abnormal Detection Algorithm Of Power Battery Based On Deeplearning

With the development of deep learning,the field of power battery anomaly detection is also trying to use data-based deep learning algorithms to make up for the shortcomings of traditional methods.The application of deep learning algorithms can effectively extract information from power battery data,improve detection efficiency,and reduce detection costs.In this context,this paper mainly studies the power battery anomaly detection algorithm based on deep learning,analyze the charging data of the power battery and judge whether the battery is abnormal.Finally,we use the actual shared electric bicycle data in operation to verify the feasibility of the algorithm and combine the algorithm with actual needs.The following are the main research contents and contributions of this paper:First,this paper designs a temporal alignment algorithm for irregular battery data.Traditional time series data alignment algorithms cannot handle data with different lengths and different sampling frequencies.This paper proposes a temporal embedding algorithm based on the attention mechanism,which uses the attention mechanism to complete the alignment of irregular time series data.In addition,a correction module based on cyclic neural network is added to the algorithm to estimate the error.This paper also proposes a self-supervised method to train the network,which guarantees that the training of the network can be completed without the use of annotation.Experimental results show that this algorithm does not require additional data labels compared to other time series alignment algorithms.Moreover,the results obtained by this algorithm have a higher accuracy rate when used for subsequent anomaly detection.Second,this paper proposes a detection method for batteries with abnormal capacity fading.The method is divided into two parts:capacity estimation and multi-feature clustering.Compared with other capacity estimation algorithms,this paper splits the traditional two-dimensional convolution into feature convolution and time convolution.This approach allows for more accurate predictions with fewer parameters.At the same time,the feature difference branch is added to the capacity prediction model,which supplements the information of data changes,so that the error of the model is further reduced.The second part of the algorithm uses the estimated capacity to perform density clustering in combination with the physical environment such as the operating temperature of the battery to detect batteries with abnormal capacity.Experimental results show that the proposed capacity anomaly detection method is far superior to other methods in terms of various indicators and scope of application.Third,this paper designs a battery thermal runaway prediction method.This method encodes the features of the historical charging data with a autoencoder.The encoded results are fed into a classification network to infer whether the battery is at risk of thermal runaway.In this paper,a prototype mechanism is proposed to solve the problem that the network is difficult to train due to too little thermal runaway battery data in the dataset.This method can make full use of all thermal runaway battery data samples and use the distance between the input data and the prototype for abnormal judgment.Experimental results show that the method can accurately identify thermally runaway batteries in multiple battery datasets.