Research On Anomaly Detection Algorithm Of Time Series Data In Intelligent Internet Of Things Based On Deep Learnin

With the development of communication and computer technology,a large amount of high-dimensional time-series data is generated in the smart Internet of Things.However,this data may contain anomalies and therefore anomaly detection algorithms need to be deployed.Limited by the computing resources of the device,anomaly detection algorithms needed to consider a balance of performance and model size.This paper designed an anomaly detection model for the high-dimensional time-series data generated by smart IoT.The lightweight algorithm of this model was further proposed in this paper to meet the needs of edge devices with limited computational resources.The specific research can be summarized as follows:(1)This paper initially presented the data characteristics and application environment of the smart IoT environment,with a focus on the description of the deep learning-based anomaly detection algorithm and lightweight algorithm process.Subsequently,a thorough analysis of the research characteristics of existing literature was conducted and summarized.(2)This paper presented Att-ADGAN,which is the anomaly detection model based on generative adversarial networks(GAN)with attention mechanism,for the high dimensionality and complexity of time-series data in the context of smart IoT.To effectively capture the high-dimensional complexity of smart IoT timeseries data,Att-ADGAN improved the model’s basic unit to an Enhanced LSTM and added a multi-channel attention mechanism.To address the generator inversion problem of traditional GAN anomaly detection models,Att-ADGAN’s generator architecture was improved to an encoder-decoder architecture that meets real-time detection needs.To improve detection performance,Att-ADGAN redesigned the reconstruction error in the anomaly score by quantifying it from both point-wise error and curve similarity aspects to better fit the definition of true error.Att-ADGAN achieved F1 scores of 0.9382,0.9521,and 0.9221 on the SWMRU,KDDCUP99,and HomeC datasets,respectively,demonstrated its effectiveness in experiments.In addition,comparative experiments were conducted with MAD-GAN,Tad-GAN,TAnoGAN,and AutoEncoder,based on the same datasets,which confirmed the superiority of Att-ADGAN’s detection performance.(3)This paper proposed a two-stage process-oriented knowledge distillation framework S-KDGAN,based on Att-ADGAN for edge devices with limited computing and storage resources.The framework employed Att-ADGAN as a teacher network to guide the training of a lightweight student network,utilizing both the final output and intermediate layer information of the teacher network to enhance distillation performance.Additionally,the framework explored the distillation effects of different structures and proposed a two-stage distillation method to further optimize the student network training based on initial distillation results.The S-KDGAN distillation framework was experimented on the S WMRU,KDDCUP99,and HomeC datasets with performance degradation rates of 1.38%,2.27%,and 1.61%,respectively.At the same time,the number of parameters was compressed by a factor of 2.37,48.85 and 15.93,the number of floating-point calculations was reduced by a factor of 18.9,294.16 and 101.46.Finally,this paper systematically summarized the main content and research results,and provided prospects for content that could be further optimized and improved.