Research And Application Of Sensor Prediction And Compensation Technology Based On Time Series Data

One of the keys of industrial Io T lies in the construction of data infrastructure,and sensor devices,regarded as the carrier of the perception layer in the Io T architecture,are the nerve endings of data collection.As an important scientific service resource,timeseries data commonly exists among the data types collected by sensors,featuring with large data volume,high data dimensionality,and complex dependency relationships.Sensor prediction and error compensation technology,as a key research area in time series analysis,aims at mining the spatio-temporal rules inherent in data series,analyzing the dependency laws between different sources and heterogeneous attributes,and then constructing data models to predict or correct the target data.Among them,the critical issues of such research are effectively handling the contextual information of multi-domain fused data series,exploring the change patterns of time series over history,and capturing the knowledge of future change states of data.Based on this,the research in this thesis focuses on the following three aspects.A prediction framework for time series data based on Encoder-Decoder architecture is proposed.At the data level,the time series is decomposed into multiple predictable basic pattern categories to highlight the inherent properties of the time series and strengthen the ability to learn the characteristics of historical lapse laws.At the model level,the traditional encoder-decoder structure is improved.And an encoder-decoder model(ED-LSTM)based on a long short-term memory neural network(LSTM)is proposed,which uses a long short-term memory neural(LSTM)in the encoder to extract essential information of data features and uses bidirectional LSTM(Bi-LSTM)in the decoder to capture the state information in two directions: historical data dependence and future changes.Finally,based on the above methods,a combined analysis of different sources and different types of related features is carried out to form a timeseries data forecasting framework.Experiments show that the proposed framework achieves better prediction results than classical methods.A data error compensation method based on a sparse attention mechanism is proposed.Firstly,based on the standard Transform architecture,a sparse attention mechanism is introduced to replace the self-attention mechanism.This mechanism retains highly relevant information and deletes irrelevant information through the top-k explicit selection method,solving the problem of the self-attention mechanism attention—the problem of failure to extract relevant information due to lack of concentration.Secondly,a sequence decomposition block is added inside the transform model to gradually extract seasonal trends from the hidden state information of the network during the inference process,in order to eliminate periodic fluctuations and emphasize the long-term trends of the sequence.Finally,the experimental comparative analysis shows that the proposed model can correct and compensate for the monitoring data error more effectively compared with the classical model.To verify the feasibility of the proposed time series error and compensation method,the involved business logic is split and designed based on the microservice architecture.Therefore,a set of time-series data prediction and error compensation analysis platforms is realized.The platform develops a unified data center to access,analyze,and store multi-source heterogeneous data in terms of data management.On the model service,a machine learning service based on data prediction and compensation is built,and the modelling process is abstracted into operator components to provide visual modeling.Finally,the practice of data prediction and compensation technology is carried out based on the uploaded measurement point data,which verifies the solution's effectiveness in this paper.