Research And Implementation Of Lightweight Real-time Fault Early Warning Method And System Based On Adversarial Learning For Vehicles

In recent years,with the continuous development of the Internet of Vehicles technology,real-time fault warning systems for commercial vehicles have become an important means to solve the safety issues that commercial vehicle customers are concerned about.Commercial vehicles usually operate for long periods of time and at high intensity in outdoor long-distance environments,which can easily result in faults.If commercial vehicle manufacturers can provide customers with real-time fault warning services,it is beneficial for drivers to take targeted preventive maintenance measures before a fault occurs,achieve prevention before it happens,realize safe operation of commercial vehicles,and greatly improve the brand competitiveness of commercial vehicle manufacturers.Therefore,major commercial vehicle manufacturers have established their own Internet of Vehicles platforms to collect real-time operating data of commercial vehicles and solve real-time fault early warning problems through in-depth data analysis.Real-time fault early warning systems for commercial vehicles typically need to collect,process and analyze vehicle sensor data,extract fault-related features,and use efficient fault early warning methods for fault early warning and diagnosis.Traditional fault early warning methods can be classified into those based on mechanism models,statistics and signal processing.These methods are more dependent on deep professional experience and domain knowledge,and cannot deal with complex nonlinear relationships.Data-driven fault early warning methods are more suitable for analyzing large-scale data generated by connected devices and systems,can identify patterns and anomalies that traditional methods cannot detect.These methods minimize the reliance on human professional knowledge and the risk of human errors,adapting to changing conditions and environments.However,the application of data-driven real-time fault early warning methods in Internet of Vehicles data still faces many challenges:high-precision fault warning methods require complex model structures,and the large number of calculation parameters will bring additional calculation delays and more resource consumption.Obtaining a sufficient number of fault samples is very difficult,and labeling also requires a lot of time and effort.Training models with unbalanced datasets of positive and negative samples is difficult to achieve good performance.In system development,fault early warning methods and stream data processing are difficult to efficiently combine,and real-time performance needs to be guaranteed while also being able to flexibly expand the computational capabilities of the model inference process.Therefore,it is necessary tosolve these problems in terms of algorithm design,model training,and system integration to realize a high-throughput,low-latency,and scalable real-time fault early warning system.In response to the above challenges,this article proposes a lightweight real-time fault early warning method based on adversarial learning for vehicles,using a large amount of highfrequency real-time operating data from a certain large commercial vehicle manufacturer.The method includes a lightweight fault early warning model,LAT-FEW(Lightweight Adversarial Transformer for Fault Early Warning),based on adversarial Transformer,and an unsupervised training scheme based on two-stage adversarial training.The LAT-FEW model is trained using this scheme and combined with advanced stream data real-time processing technology to design and develop a commercial vehicle real-time fault early warning system.The main work and contributions of this paper can be summarized in the following three aspects:1.A lightweight real-time fault early warning method based on adversarial learning for vehicles is proposed,which achieves good fault early warning effects using lightweight model structures under the condition of scarce fault samples.The method proposes the LAT-FEW model based on adversarial Transformer.The model only uses a single layer of Transformer Encoder-Decoder structure improved by causal convolution calculation in the time prediction layer for capturing time-dependent relationships of sequential data.This approach retains parallel computing capabilities while significantly reducing computational parameters of the model.The unsupervised training scheme based on two-stage adversarial training is proposed to optimize the model using only normal sample data,avoiding the serious impact of imbalanced data training on model performance,and introducing adversarial ideas to improve the fault early warning performance of the lightweight model.2.Using two real datasets,a large number of experiments are designed to verify the effectiveness of the lightweight real-time fault early warning method based on adversarial learning for vehicles.The accuracy of the evaluation method is verified using three evaluation metrics,i.e.,precision,recall,and F1 score,while the real-time performance of the method is verified using the total parameter calculation and total duration calculation of the model under the same data volume.The generalizability of the method is evaluated through comparative experiments on two datasets.The effectiveness and impact of the main design structures proposed in the lightweight vehicle real-time fault early warning method based on adversarial learning are verified through comparative experiments with the baseline models and a series of ablation experiments.The experimental results show that with the proposed training scheme,the LAT-FEW model achieves the best comprehensive performance in fault early warning.3.Based on the real Internet of Vehicles big data platform of a large commercial vehicle manufacturer,a commercial vehicle real-time fault early warning system is designed and implemented.The system architecture is designed using big data technologies such as Flink,Kafka,ClickHouse,and Grafana.Four functional modules are designed for data real-time acquisition,data real-time processing and analysis,data storage and management,and visualization display.The LAT-FEW model based on adversarial learning is integrated into the Flink processing operator for online processing and model analysis of real-time operating data of commercial vehicles,improving the computational efficiency and scalability of the fault early warning model,and achieving an industrial-level real-time fault early warning application.