Lightweight time sequence information fusion perception control method and system
Through the lightweight timing information fusion perception control method, the operating status of wireless sensor nodes is dynamically adjusted, the problem of excessive resource consumption is solved, efficient data processing and prediction is achieved, and large-scale deployment of IoT nodes is promoted.
Patent Information
- Application Number
- CN202510642655.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-05-19
- Publication Date
- 2025-07-22
AI Technical Summary
Existing wireless sensor nodes consume too much resources during task execution and data transmission, resulting in short battery life, data congestion and application limitations. Existing prediction models are difficult to deploy on low-power devices and increase energy consumption.
The lightweight timing information fusion perception control method is adopted to dynamically adjust the node operating status through state observation, data segmentation, curve feature extraction and trend prediction, and combine feasible space windows and LightGBM models to achieve efficient data segmentation and feature extraction, reducing calculation complexity and energy consumption.
Without affecting network performance, the overall energy consumption of nodes and network interaction load are significantly reduced, the scale of IoT perception deployment is optimized, and prediction accuracy and resource utilization efficiency are improved.
Smart Images

Figure CN120358211A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of the Internet of Things, and particularly to a lightweight time-series information fusion perception control method and system. Background Art
[0002] In recent years, driven by the development of digitalization, the scale of Internet of Things node deployment has increased exponentially. More than 70% of them are lightweight weak terminal nodes mainly for environmental perception functions, providing necessary data support for various digital applications and bringing huge changes to the informatization development in China.
[0003] However, the current wireless sensor node architecture still has the problem of excessive resource consumption in both task execution and data transmission, which limits the scale of Internet of Things perception deployment, as Figure 1 shown. On the one hand, when the sensing node is in the fully active state, it has good real-time performance and can capture the key information required by the application. However, it is usually accompanied by high power consumption and severely limited battery life, resulting in an increase in later maintenance costs. On the other hand, most existing sensing nodes use narrowband communication, and the data rate does not exceed 16 kbps to meet the communication requirements of low power consumption and long distance. However, with the continuous increase in the number of nodes, the upload of its massive sensing data may cause problems such as data congestion, frequent disconnection, and difficult access at the edge infrastructure. Since under the long-distance deployment requirement, the power consumption of data transmission accounts for more than 85% of the local power consumption of the node, the above two problems are interrelated and need to be solved simultaneously to effectively resolve the problems of high communication load, difficult optimization of power consumption and battery life, and limited applications of existing sensing nodes.
[0004] To solve the above problems, existing sensing nodes use dynamic power management (DPM) technology. By analyzing the future change trend of the target sensing information, the operating state of the node is dynamically adjusted, so that the node is only awakened and performs data acquisition and upload operations when necessary, thereby reducing the overall operating power consumption and network interaction load, and maximizing the overall energy efficiency, as Figure 2As shown in the figure. DPM needs to predict the future system load to determine when to switch to the low-power mode. However, existing fitting models that create underlying data structures and properties based on log data, such as computing load prediction and wireless network prediction, are difficult to deploy on resource-constrained low-power devices. Time series statistical models such as ARIMA and SARIMA have high computational efficiency, but are difficult to handle application scenarios with complex cycles. Machine learning models such as multi-layer perceptron (MLP), recurrent neural network (RNN), long short-term memory (LSTM), and convolutional neural network (CNN) provide more complex data fitting capabilities, but the hardware of low-power sensor nodes often cannot meet their requirements for memory and computing power. Furthermore, a series of processes in which DPM trains local models and adjusts the power state can be regarded as "additional tasks" of the terminal, which further consumes precious energy, increases the possibility of task delay or performance instability, and thus limits the wide application of passive Internet of Things terminal systems. Summary of the Invention
[0005] The purpose of the present invention is to overcome the shortcomings of the prior art, and provides a lightweight time series information fusion perception control method and system, which solves the deficiencies existing in the prior art.
[0006] The purpose of the present invention is achieved through the following technical solutions: a lightweight time series information fusion perception control method, the method includes:
[0007] State observation step: Obtain key perception data, and use an observer to continuously track the task load and communication load of the node locally;
[0008] Data segmentation step: Based on the feasible space window of multiple non-linear fitting functions, adaptively segment the input perception data and channel state;
[0009] Curve feature extraction step: Based on the curve features of time series data, preprocess the time series data, then extract features from the curve according to the characteristics of the time series data and node environment information, and use Lasso regression with L1 regularization to select important features;
[0010] Trend prediction step: Analyze and predict the compressed curve feature data based on the decision tree algorithm, classify the time series data into curves with different features through the LightGBM model, select the corresponding empirical function model for fitting according to the classification results, and finally adaptively adjust the algorithm to dynamically change the operating state of the node;
[0011] Dynamic control step: Based on the future trend of the data, implement a dynamic control strategy by autonomously adjusting node parameters, solve the contradiction between time series data processing and dynamic control in the edge environment, and balance real-time performance, lightweight and prediction accuracy.
[0012] The specific steps of the data segmentation are as follows:
[0013] When the collected data reaches the weak terminal node, the function needs to segment the time series data according to the trend to enhance the accuracy of subsequent feature extraction and prediction. The function realizes the segmentation of time series data by dynamically maintaining a feasible space window. Each data segment starts from a fixed starting point and continuously absorbs subsequent data points until it exceeds the preset error range.
[0014] An initial feasible space window is established with the current segmentation starting point and the first data point, and the slope range from the starting point to the upper and lower boundaries of each data point (based on the error threshold δ) is calculated as the window boundary. For each newly arrived data point, calculate its actual slope with the starting point. If this value always lies between the current upper and lower boundaries of the window, it is allowed to be added to the current segment and the window boundary is updated to a more stringent slope limit. When the new data point causes the feasible space window to be unsatisfied (i.e., there is a contradiction between the upper and lower boundaries), the current segment is terminated and a new segment starting point is generated. Repeat the above process until all data points are processed, and finally output the multi-segment data division formed by the continuous feasible slope range constraints. This mechanism realizes the adaptive identification of data trend change points on the premise of ensuring that the fitting error is controllable.
[0015] The specific steps of the curve feature extraction are as follows:
[0016] Native data acquisition and preprocessing: Record the micro-watt to milli-watt level of electrical energy generated from a high-output impedance DC source through an oscilloscope, so as to collect multiple original time series data containing more than the set time-voltage points, and segment the original time series data through a feasible space window;
[0017] Training set curve feature extraction: Sequentially extract statistical features, time domain features, and frequency domain features from the preprocessed time series data segments. Each feature is calculated by calling the standard C library, and the overall changes, discreteness, distributions, and response features of the curve are quantified through various feature indicators to construct a time series data curve feature training set;
[0018] Training set curve feature selection: Delete the features with little influence on the prediction results, retain the features with great influence on the prediction results for training, visualize the importance of multiple columns of continuous features after feature extraction through the built-in feature importance function in sklearn, and debug the machine learning classifier to calculate the importance of the displayed features.
[0019] The specific steps of the trend prediction are as follows:
[0020] A function classification model for time-series data curves based on a multi-class decision tree is proposed. The model is trained by processing the historical time-series data set of sensor nodes and generates a callable model file to quickly classify the current segmented curve, match the fitting empirical formula with the smallest average residual value. LightGBM is introduced into the function classification of time-series data curves, and the feature binning technology is used to preprocess the input data to generate a feature histogram with a hierarchical index structure, where each histogram unit stores the statistical distribution parameters of the feature values. A dynamic split point selection algorithm is designed to calculate the information gain values corresponding to each split point by traversing the discretized indexes of the feature histogram, and the split threshold corresponding to the maximum gain value is selected as the decision tree split node. A residual optimization matching mechanism is established to perform the fitting calculations of multiple preset empirical formulas on the input data segment synchronously, and the fitting function with the smallest average residual value is selected as the classification output result.
[0021] Deploy a variable-frequency task scheduling controller that dynamically adjusts the working frequency of the node according to the real-time collected data fluctuation intensity. When the data change rate exceeds the first threshold, the execution frequency of data collection and communication operations is increased, and when the data change rate is lower than the second threshold, the execution frequency is decreased. Build a data trend prediction module to establish a change trend prediction model by analyzing the statistical characteristics of historical time-series data and adjust the working mode of the node in advance based on the prediction results of the current data window. Design an energy consumption-aware weight calculation algorithm that inputs the data change rate, network load intensity, and historical working mode parameters into the decision matrix to generate an optimal task scheduling sequence.
[0022] A lightweight time-series information fusion perception control system, the system includes a state observation module, a data segmentation module, a curve feature extraction module, a trend prediction module, and a dynamic control module;
[0023] The state observation module is configured to obtain key perception data and track the task load and communication load of the node locally in real time through an observer;
[0024] The data segmentation module is configured to perform adaptive segmentation on the input perception data and channel state based on the feasible space window of multiple nonlinear fitting functions;
[0025] The curve feature extraction module is configured to preprocess the time-series data based on the curve features of the time-series data, extract the features of the curve based on the characteristics of the time-series data and the node environment information, and use Lasso regression with L1 regularization to select important features;
[0026] The trend prediction module is configured to analyze and predict the compressed curve feature data based on the decision tree algorithm, classify the time series data into curves with different features through the LightGBM model, select the corresponding empirical function model for fitting according to the classification results, and finally adaptively adjust the algorithm to dynamically change the operating state of the nodes;
[0027] The dynamic control module is configured to implement a dynamic control strategy by autonomously adjusting node parameters based on the future trend of the data, solve the contradiction between time series data processing and dynamic control in the edge environment, and balance real-time performance, lightweight and prediction accuracy.
[0028] The data segmentation module specifically includes:
[0029] When the collected data reaches the weak terminal node, the function needs to segment the time series data according to the trend to enhance the accuracy of subsequent feature extraction and prediction. The function realizes the segmentation of time series data by dynamically maintaining a feasible space window. Each data segment starts from a fixed starting point and continuously absorbs subsequent data points until it exceeds the preset error range;
[0030] An initial feasible space window is established with the current segmentation starting point and the first data point. The slope range from the starting point to the upper and lower boundaries of each data point is calculated as the window boundary. For each newly arrived data point, its actual slope with the starting point is calculated. If this value is always between the current upper and lower boundaries of the window, it is allowed to be added to the current segment and the window boundary is updated to a more stringent slope limit. When the new data point causes the feasible space window to be unable to be satisfied, the current segment is terminated and a new segment starting point is generated. Repeat the above process until all data points are processed. Finally, the multi-segment data division formed by the continuous feasible slope range constraint is output. This mechanism realizes the adaptive identification of data trend change points on the premise of ensuring that the fitting error is controllable.
[0031] The curve feature extraction module specifically includes:
[0032] Native data acquisition and preprocessing: Record the micro-watt to milli-watt level electrical energy generated from a high output impedance DC source through an oscilloscope, so as to collect multiple original time series data containing more than the set time-voltage points, and segment the original time series data through a feasible space window;
[0033] Training set curve feature extraction: Sequentially extract statistical features, time domain features and frequency domain features from the preprocessed time series data segments. Each feature is calculated by calling the standard C library. The overall change, discreteness, distribution and response features of the curve are quantified through each feature index, and a time series data curve feature training set is constructed;
[0034] Training set curve feature selection: Delete features with little impact on the prediction result, and retain features with great impact on the prediction result for training. Visualize the importance of multiple columns of continuous features after feature extraction through the built-in feature importance function in sklearn, and debug the machine learning classifier to calculate the importance of the displayed features.
[0035] The trend prediction module specifically includes:
[0036] Propose a function classification model for time-series data curves based on multi-class decision trees. Complete training by processing the historical time-series data set of sensor nodes and generate a callable model file. Quickly classify the current segmented curve, match the fitting empirical formula with the smallest average residual value. Introduce LightGBM into the function classification of time-series data curves, and use the feature binning technique to preprocess the input data discretely to generate a feature histogram with a hierarchical index structure, where each histogram unit stores the statistical distribution parameters of the feature values; design a dynamic split point selection algorithm, calculate the information gain value corresponding to each split point by traversing the discretized index of the feature histogram, and select the split threshold corresponding to the maximum gain value as the decision tree splitting node; establish a residual optimization matching mechanism, perform fitting calculations of multiple preset empirical formulas on the input data segment synchronously, and select the fitting function with the smallest average residual value as the classification output result.
[0037] Establish an adaptive duty cycle scheduling strategy for wireless sensor nodes, dynamically adjust high-energy-consuming operations such as data collection and inter-node communication, increase the scheduling frequency when the data change monitored on the node is obvious, and decrease the scheduling frequency when the data change is relatively gentle. The expression of the real-time scheduling frequency is: ,
[0038] where, and are the upper and lower limits of the scheduling frequency set for this node, represents the coefficient of variation (CV) of future time, which is used to describe the degree of dispersion of the data. and represent the maximum and minimum values of the dispersion coefficient in historical data.
[0039] The duty cycle scheduling algorithm introduces the change rate of node tracking data and its historical situation, so as to achieve a more complete tracking of the changes in environmental data and network load with less total sensing, which serves as the basis for flexibly arranging tasks for sensor nodes. Through the prediction of local time series data based on decision trees, the changes and trends of data related to the current node can be obtained more accurately, so as to adjust the sensing and communication frequencies of the node before the environment fluctuates; conversely, when the environment is relatively stable, the sensing and communication frequencies are reduced to minimize energy waste.
[0040] The present invention has the following advantages:
[0041] 1. With the feasible space window time series data segmentation algorithm, the input key sensing data is segmented according to the sub-trajectory mode, laying a foundation for further upstream time series analysis tasks (clustering, compression, and prediction). Compared with the traditional sliding window segmentation algorithm, the present invention does not need to calculate the error for all data points when new data arrives, and only needs to maintain the boundaries of the feasible window, thus greatly reducing the computational complexity without losing data integrity.
[0042] 2. Support the general time series data curve feature extraction algorithm, reducing the storage and transmission overhead of model training. Compared with the traditional time series model, the present invention compresses the continuous and multi-dimensional original time series data, retaining as much important information of the curve as possible while reducing the dimension of the time series data, thereby greatly reducing the computational complexity of the time series data model while improving the performance of the prediction model, making it easier to deploy in resource-constrained nodes.
[0043] 3. Based on the lightweight time series data trend prediction algorithm and low-power hardware implementation, the future trends of task and communication loads are coupled with their operating states, and the change rate of node tracking data and its historical situation are combined to achieve adaptive adjustment of the node operating state. Compared with the traditional algorithm, the present invention achieves a more complete tracking of the changes in environmental data and network load with less total sensing, further reducing the overall operating power consumption and network interaction load. Description of the Drawings
[0044] Figure 1 It is a schematic diagram of the existing wireless sensor node architecture; Figure 2 It is a schematic diagram of the structure of the dynamic power management coefficient; Figure 3 It is a schematic diagram of the structure of the present invention; Figure 4 It is a schematic diagram of the time series data segmentation algorithm based on FSW; Figure 5 It is a schematic diagram of the geometric representation of the feasible space of linear fitting and polynomial fitting; Figure 6Schematic diagram of the process of feature extraction and feature selection; Figure 7 Schematic diagram of Feature importance feature weights; Figure 8 Schematic diagram of Eli5 feature weights; Figure 9 Confusion matrix diagram of the function classification model for time-series data curves. Detailed implementation manners
[0045] To make the objectives, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are only a part rather than all of the embodiments of the present application. The components of the embodiments of the present application usually described and illustrated in the accompanying drawings here can be arranged and designed in various different configurations. Therefore, the detailed description of the embodiments of the present application provided below with reference to the accompanying drawings is not intended to limit the protection scope of the claimed present application, but merely represents the selected embodiments of the present application. All other embodiments obtained by those skilled in the art based on the embodiments of the present application without creative efforts fall within the protection scope of the present application. The present invention will be further described below with reference to the accompanying drawings.
[0046] Aiming at the problem that it is difficult to coordinate the real-time perception and operating power consumption of existing Internet of Things nodes and it is difficult to optimize the communication load, resulting in limited practical applications, the present invention proposes a time-series information fusion lightweight sensing and control mechanism system. By constructing an efficient curve feature extraction algorithm and a time-series data trend prediction algorithm, it realizes high-precision lightweight prediction of time-series data trends, accurately captures the complex non-linear relationships of time series while ensuring low overhead of local resources. While realizing fine-grained energy state control of sensing nodes, the present invention significantly reduces the deployment overheads of prediction models such as network traffic and node load, maximizes the overall energy efficiency of nodes without affecting network performance, further reduces the cost and instability of Internet of Things application data acquisition, and promotes the large-scale deployment and application of Internet of Things nodes.
[0047] Aiming at the problem that it is difficult to extract the time and space where key sensing data is located. Based on the feasible space window of multiple non-linear fitting functions, the present invention maintains and updates multiple variables describing the boundaries of the feasible space when new data arrives, so as to realize parallel and efficient adaptive time-series data segmentation. Subsequently, based on the efficient curve feature extraction algorithm for terminal time-series data, the present invention compresses the information loss of sensor data, and through a lightweight time-series data trend prediction algorithm, realizes the dynamic control of the future operating state of nodes.
[0048] Such as Figure 3As shown in the figure, the overall system consists of a sensor control module, a state observer, a lightweight time-series data processing and prediction algorithm, a dynamic control, and a communication subsystem. Among them, the lightweight time-series data processing and prediction include three parts: data segmentation, feature extraction, and trend prediction. Specifically, the present invention uses an observer to continuously track the local task load and communication load of the node, and through a lightweight time-series data processing and prediction algorithm, accurately estimate future sensing data without consuming too many resources, and dynamically adjust the operating states of the sensor control module and the communication subsystem according to the prediction results, thereby reducing the overall operating power consumption and network interaction load. The specific time-series data processing and prediction algorithm includes the following content:
[0049] First, the original time-series data includes excessive redundancy and unstructured parts, and a data segmentation algorithm is required to preprocess the original data. Aiming at the problem that the traditional data segmentation algorithm has poor effect when the historical data fluctuates greatly, the present invention is based on the feasible space window of multiple nonlinear fitting functions to adaptively segment the time-series data such as the input sensing data and channel state, thereby improving the performance of the prediction algorithm for data streams with high randomness and uncertainty.
[0050] As Figure 4 shown, by introducing an adaptive segmentation algorithm before time-series prediction, the present invention further improves the accuracy and efficiency of function fitting, and realizes the accurate capture of the characteristics of the node energy state data. It should be noted that the improper selection of the data range will seriously affect the ability of the algorithm to master the true trend of the data. For the traditional online segmentation algorithm, when each new data point arrives, a whole error calculation is required, and its low efficiency cannot be applied to the dynamic control of high-frequency and low-power sensors. Aiming at this problem, the present invention is based on the Feasible Space Window (FSW) algorithm to optimize the traditional fixed-length segmentation method, so that the prediction algorithm can cope with data streams with high randomness and uncertainty.
[0051] The FSW algorithm is an effective method for determining whether another data point can be added to the current segment. When each new data point arrives, it is not necessary to recalculate the error index (such as residual) for all data points, but only to update the existing data. FSW uses a metric called Feasible Space (FS) to find the farthest segmentation point that satisfies the error bound condition in each line segment. Specifically, for the time series and , where is the point that is first added to the segment , is the point that is newly added to the segment , is the maximum error tolerance. Suppose , and it is desired to add the newly arrived data point to the segment provided that the segmentation criterion is satisfied:
[0052]
[0053]
[0054]
[0055] represents the slope of the straight line passing through points a and b simultaneously, represents any real point between and in the dataset, and and represent the upper and lower boundary points centered on respectively and .
[0056] The present invention defines that can be added to the segment , then must be located between the lines and , as shown in Figure 5 . In addition, if the point is accepted by the line segment , the boundaries of FS (the dark gray area in Figure 5 ) will be updated to and . The above process is the forward update process of FSW in the time series data segmentation algorithm.
[0057] When FS becomes empty, the update process of the boundaries and of FS stops, which means that the algorithm cannot approximate any more subsequent data points (including the current data point) within the error range. Therefore, the time series data from to can be divided into segments , and the next segment will continue to traverse forward from the point and update FS.
[0058] According to the segmentation criterion, when new data points arrive, the terminal does not need to calculate all the vertical distances between the interpolation line and the real data, but only needs to compare and update the boundaries of the current FS. By integrating the above FSW algorithm into the adaptive segmentation algorithm for sequential data such as input perception data and channel state, the algorithm can maintain high efficiency and low overhead while structuring the original data stream at a relatively high rate. After measurement, the FSW algorithm reduces the computational complexity of the classical sliding window algorithm with an average segment length L from O(Ln) to O(n), thus meeting the requirements of parallel and efficient segmentation of the charging power data stream of multiple nodes.
[0059] Furthermore, in order to convert the original data into an input that is easier for the model to understand, this patent mines the curve features in the sequential data through a curve feature extraction algorithm. After preprocessing the original data such as missing value processing, normalization, and noise reduction, the algorithm will extract features based on the characteristics of the sequential data and the node environment information, and finally use the Lasso regression with L1 regularization to screen important features.
[0060] To implement a multi-class classification prediction model for the energy state of passive nodes, curve feature extraction is planned for historical data segments. In order to accurately and efficiently represent the central tendency and dispersion tendency of sequential data, as well as the response of data in the time domain, an efficient curve feature extraction algorithm is planned to further improve the prediction efficiency of the multi-class classification prediction model for duty cycle adjustment without reducing the information integrity.
[0061] Table 1. Explanation Table of Dataset Feature Fields
[0062]
[0063] Feature extraction and selection are carried out from three categories: statistical features, time domain features, and frequency domain features of the curve, as shown in Table 1. Statistical features include the central tendency and dispersion tendency of the data: indicators such as mean, median, mode, and midrange can reflect the central tendency of the curve, while indicators such as variance, range, quartile, interquartile range, and coefficient of variation reflect the dispersion tendency of the curve; time domain characteristics refer to the response of historical energy state data in the time domain, which are usually divided into dimensional parameters and dimensionless parameters. Dimensional parameters have a direct physical interpretation and are widely used as the most common feature indicators, such as maximum value, minimum value, peak value, average value, variance, standard deviation, and mean square value. However, since these size indicators are sensitive to energy changes and are easily affected by the environment, dimensionless parameters are becoming increasingly important. Dimensionless parameters include peak factor, kurtosis factor, waveform factor, etc., which can effectively eliminate the influence of the above environmental disturbances.
[0064] As Figure 6 shown, the acquisition of curve feature extraction with high efficiency and low information loss is achieved through the following process.
[0065] Original raw data acquisition and preprocessing: Common energy sources for Internet of Things applications such as micro wind turbines, piezoelectric ceramics, and solar silicon panels were selected. The electrical energy in the range of microwatts to milliwatts generated from a high-output impedance DC source was recorded by an oscilloscope, thereby collecting multiple original time series data containing more than 1500 time-voltage points. To ensure the fitting effect and efficiency, the original data was normalized and the FSW algorithm was used to gradually segment the data.
[0066] Training set curve feature extraction: For the preprocessed time series data segments, the present invention sequentially extracts statistical features, time domain features, and frequency domain features from them. Specifically, each curve feature can be calculated by calling the standard C library, and the calculation of some features (such as mean and variance, peak factor and waveform factor, etc.) shares intermediate results. The present invention optimizes the performance and reduces the number of instructions by writing efficient code, thereby efficiently quantifying the overall changes, discreteness, distribution, response and other features of the curve through various feature indicators, and constructing a training set for the curve features of time series data.
[0067] Training set curve feature selection: The dynamic power management of nodes has strict speed requirements, so the training time of the data prediction model should be as short as possible. The present invention saves the time and resources for model training by deleting the features with less influence on the prediction results and only retaining the features with greater influence on the results for training. Further, the present invention visualizes the importance of the 13 columns of continuous features after feature extraction through the built-in feature importance function in sklearn, as Figure 7 shown. Due to the influence of environmental noise and feature collinearity, there are certain defects in the feature selection of feature importance. The present invention debugs the machine learning classifier through Eli5, calculates the importance of the features with display through permutation importance, and explains the prediction function of LGBMClassifier and LGBMRegressor, as Figure 8 shown. The present invention retains the features with a relatively high degree of correlation between the features and the target, and deletes the features with extremely low importance in the prediction of the model, so as to reduce the dimension of the model and thus reduce the training cost of the model.
[0068] Furthermore, the present invention analyzes and predicts the compressed curve feature data based on the decision tree algorithm. Specifically, the efficient and lightweight multi-class LightGBM model is used to classify the time series data into curves with different features, and the corresponding empirical function model is selected for fitting according to the classification results. Finally, the algorithm adaptively adjusts to dynamically change the operating state of the node, so as to release the power consumption constraint and communication constraint with lower power consumption.
[0069] The present invention introduces a novel time series data prediction algorithm, which can predict the sampling and communication state trends of low-power wireless sensor nodes without prior observation. Specifically, the present invention explores the common points of the sampling data curves in a heterogeneous environment and the features used to describe the curves. Thereby, an algorithm for classifying the empirical distribution function of the energy state curve based on multiple curve features extracted from time series data is designed. The algorithm matches the empirical distribution function of the minimum mean square error (MSE) with the time series data curve, so that the mathematical expression obtained by curve fitting can represent the future change trend of the target sensing information.
[0070] First, a function classification model for time series data curves is proposed. The model is based on the multi-class decision tree algorithm, trained through the time series data curve dataset of sensor nodes and saves the model, and quickly classifies the current segmented curve to match the fitting empirical formula with the smallest average residual value. The principle of decision tree analysis is to generate an empirical tree through data segmentation created by applying a set of simple rules. Decision tree analysis is to generate a set of rules through repeated splitting processes for predicting future energy levels.
[0071] Through refined parameter adjustment, LightGBM is introduced into the function classification for time series data curves, and the results obtained are as Figure 8 shown. Common tree methods include Classification and Regression Tree (CART), C4.5, XGBoost, and LightGBM, etc. Among them, LightGBM shows excellent classification effects in multiple application fields. LightGBM uses sorting to discretize the feature values, uses the discrete values in the histogram as indexes, calculates the cumulative statistics, and finds the best split point according to the discrete values of the histogram after the traversal is completed, thus realizing the overall optimization of the traditional algorithm - Gradient Boosting Decision Tree (GBDT). Compared with XGBoost and random forest models, the algorithm of this patent has obtained the best results in the four indicators of accuracy, precision, recall rate, and training time consumption, and while maintaining good performance, it has greatly reduced the time cost.
[0072] Based on the data trend obtained by the prediction algorithm, the present invention solves the conflict between the precise control of traditional low-power technologies and the minimization of computing resources. For this purpose, the present invention establishes an adaptive duty cycle scheduling strategy for wireless sensor nodes to dynamically adjust high-energy-consuming operations such as data collection and inter-node communication. The core idea is to increase the scheduling frequency when the data change on the node is obvious, and decrease the scheduling frequency when the data change is relatively gentle. The expression of the adjusted real-time scheduling frequency is:
[0073] ,
[0074] where, and are the upper and lower limits of the scheduling frequency set for this node. represents the coefficient of variation (CV) of future time, which is used to describe the degree of dispersion of data. and represent the maximum and minimum values of the discrete coefficient in historical data. Further, the present invention introduces the change rate of node tracking data and its historical situation into the duty cycle scheduling algorithm, so as to achieve a relatively complete tracking of the changes in environmental data and network load with less total sensing, as the basis for flexibly arranging tasks for sensor nodes. Through the prediction of local time series data based on decision trees, the present invention can accurately obtain the changes and trends of data related to the current node, so as to adjust the sensing and communication frequencies of the node before the environment fluctuates; conversely, when the environment is relatively stable, reduce the sensing and communication frequencies to minimize energy waste.
[0075] The above are only the preferred embodiments of the present invention. It should be understood that the present invention is not limited to the form disclosed herein, should not be regarded as excluding other embodiments, but can be used in various other combinations, modifications and improvements, and can be changed within the scope of the concept described herein through the above teachings or the technology or knowledge in related fields. And the changes and alterations made by those skilled in the art without departing from the spirit and scope of the present invention shall fall within the protection scope of the appended claims of the present invention.
Claims
1. A lightweight temporal information fusion perception and control method, characterized in that: The method includes: State observation step: Obtain key perception data based on event triggering, and use a lightweight observer to track the task load and communication load of the node locally in real time; Data segmentation step: Construct a non-linear fitting function library, calculate the feasible space window of each function in parallel, and adaptively segment the granularity according to the input window overlap degree and node load dynamic options; Curve feature extraction step: Based on the curve features of time series data, extract the statistical features, time domain features and load coupling features from the preprocessed time series data curve, and use Lasso regression with L1 regularization for screening to reduce the dimension; Trend prediction step: Analyze and predict the compressed curve feature data based on the decision tree algorithm, divide the time series data into corresponding base classes through the LightGBM model, and call the corresponding empirical function library for fitting according to the classification results. Finally, adaptively adjust the algorithm to dynamically change the operating state of the node; Dynamic control step: Based on the future trend of the data, implement a dynamic control strategy by autonomously adjusting the node parameters, solve the contradiction between time series data processing and dynamic control in the edge environment, and balance real-time performance, lightweight and prediction accuracy.
2. The lightweight time-series information fusion perception control method according to claim 1, wherein: The specific data segmentation step includes: When the collected data reaches the weak terminal node, the function needs to segment the time series data according to the trend to enhance the accuracy of subsequent feature extraction and prediction. The function realizes the segmentation of time series data by dynamically maintaining the feasible space window. Each data segment starts from a fixed starting point and continuously absorbs subsequent data points until it exceeds the preset error range; Establish an initial feasible space window with the current segmentation starting point and the first data point, calculate the slope range from the starting point to the upper and lower boundaries of each data point as the window boundary. For each newly arrived data point, calculate its actual slope with the starting point. If this value is always between the current upper and lower boundaries of the window, it is allowed to be added to the current segment and the window boundary is updated to a more stringent slope limit. When the new data point causes the feasible space window to be unable to be satisfied, terminate the current segment and generate a new segment starting point, and repeat the above process until all data points are processed. Finally, output the multi-segment data division formed by the continuous feasible slope range constraint. This mechanism realizes the adaptive identification of data trend change points on the premise of ensuring that the fitting error is controllable.
3. The lightweight time-series information fusion perception control method according to claim 1, wherein: The specific curve feature extraction step includes: Native data collection and preprocessing: Record the electrical energy from micro-watts to milli-watts generated by a high-output impedance DC source through an oscilloscope, so as to collect multiple original time series data containing more than the set time-voltage points, and segment the original time series data through the feasible space window function; Training set curve feature extraction: Extract statistical features, time domain features and frequency domain features from the preprocessed time series data segments in turn. Each feature is calculated by calling the standard C library. Quantify the overall change, discreteness, distribution and response features of the curve through each feature index, and construct a time series data curve feature training set; Training set curve feature selection: Delete features with little impact on the prediction results, and retain features with great impact on the prediction results for training. Visualize the importance of multiple columns of continuous features after feature extraction through the built-in feature importance function in sklearn, and debug the machine learning classifier to calculate the importance of the displayed features.
4. A lightweight temporal information fusion perception control method according to claim 1, characterized in that: The specific steps of the trend prediction include: Propose a function classification model for time-series data curves based on a multi-class decision tree. Complete training by processing the historical time-series data set of sensor nodes and generate a callable model file. Quickly classify the current segmented curve, match the fitting empirical formula with the smallest average residual value. Introduce LightGBM into the function classification of time-series data curves, and use the feature binning technique to preprocess the input data to generate a feature histogram with a hierarchical index structure, where each histogram unit stores the statistical distribution parameters of the feature values; design a dynamic split point selection algorithm, calculate the information gain values corresponding to each split point by traversing the discretized index of the feature histogram, and select the split threshold corresponding to the maximum gain value as the decision tree splitting node; establish a residual optimization matching mechanism, perform fitting calculations of multiple preset empirical formulas on the input data segment synchronously, and select the fitting function with the smallest average residual value as the classification output result. Deploy a variable-frequency task scheduling controller, which dynamically adjusts the working frequency of the node according to the data fluctuation intensity collected in real time. When the data change rate exceeds the first threshold, increase the execution frequency of data collection and communication operations. When the data change rate is lower than the second threshold, reduce the execution frequency; construct a data trend prediction module, establish a change trend prediction model by analyzing the statistical characteristics of historical time-series data, and adjust the working mode of the node in advance based on the prediction results of the current data window; design an energy consumption-aware weight calculation algorithm, input the data change rate, network load intensity, and historical working mode parameters into the decision matrix to generate an optimal task scheduling sequence.
5. A lightweight time-series information fusion perception control system, characterized in that: The system includes a state monitoring module, a data segmentation module, a curve feature extraction module, a trend prediction module, and a dynamic control module; The state observation module is configured to obtain key perception data and track the task load and communication load of the node locally in real time through an observer; The data segmentation module is configured to adaptively segment the input perception data and channel state based on the feasible space window of multiple non-linear fitting functions; The curve feature extraction module is configured to preprocess the time-series data based on the curve features of the time-series data, extract features from the curve according to the characteristics of the time-series data and the node environment information, and use Lasso regression with L1 regularization to select important features; The trend prediction module is configured to analyze and predict the compressed curve feature data based on the decision tree algorithm, classify the time-series data into curves with different features through the LightGBM model, select the corresponding empirical function model for fitting according to the classification results, and finally dynamically change the operating state of the node through the adaptive adjustment algorithm. The dynamic control module is configured to implement a dynamic control strategy by autonomously adjusting node parameters based on the future trend of data, solve the contradiction between time-series data processing and dynamic control in the edge environment, and balance real-time performance, lightweight and prediction accuracy.
6. The lightweight time-series information fusion perception control system according to claim 5, characterized in that: The data segmentation module specifically includes: When the collected data reaches the weak terminal node, the function needs to segment the time-series data according to the trend to enhance the accuracy of subsequent feature extraction and prediction. The function realizes the segmentation of time-series data by dynamically maintaining a feasible space window. Each data segment starts from a fixed starting point and continuously absorbs subsequent data points until it exceeds the preset error range; An initial feasible space window is established with the current segmentation starting point and the first data point. The slope range from the starting point to the upper and lower boundaries of each data point is calculated as the window boundary. For each newly arrived data point, its actual slope with the starting point is calculated. If this value is always between the current upper and lower boundaries of the window, it is allowed to be added to the current segment and the window boundary is updated to a more stringent slope limit. When the new data point causes the feasible space window to be unable to be satisfied, the current segment is terminated and a new segment starting point is generated. Repeat the above process until all data points are processed. Finally, the multi-segment data division formed by the continuous feasible slope range constraint is output. This mechanism realizes the adaptive recognition of data trend change points on the premise of ensuring that the fitting error is controllable.
7. The lightweight time-series information fusion perception control system according to claim 5, wherein: The curve feature extraction module specifically includes: Native data acquisition and preprocessing: Record the micro-watt to milli-watt level electrical energy generated from a high output impedance DC source through an oscilloscope, so as to collect multiple original time series data containing more than the set time-voltage points, and perform data segmentation on the original time series data through a feasible space window; Training set curve feature extraction: Sequentially extract statistical features, time domain features and frequency domain features from the preprocessed time series data segments. Each feature is calculated by calling the standard C library. The overall change, discreteness, distribution and response features of the curve are quantified through each feature index, and a time series data curve feature training set is constructed; Training set curve feature selection: Delete the features with little influence on the prediction result, and retain the features with great influence on the prediction result for training. The importance of multiple columns of continuous features after feature extraction is visualized through the feature importance function built in sklearn, and the machine learning classifier is debugged to calculate the importance of the displayed features.
8. A lightweight time-series information fusion perception control system according to claim 5, characterized in that: The trend prediction module specifically includes: A function classification model for time-series data curves based on a multi-class decision tree is proposed. The model is trained by processing the historical time-series data set of sensor nodes and generates a callable model file to quickly classify the current segmented curve, match the fitting empirical formula with the smallest average residual value, introduce LightGBM into the function classification of time-series data curves, and use the feature binning technology to discretize and preprocess the input data to generate a feature histogram with a hierarchical index structure, where each histogram unit stores the statistical distribution parameters of the feature values; design a dynamic split point selection algorithm, calculate the information gain values corresponding to each split point by traversing the discretized indexes of the feature histogram, and select the split threshold corresponding to the maximum gain value as the decision tree splitting node; establish a residual optimization matching mechanism, perform the fitting calculations of multiple preset empirical formulas on the input data segment synchronously, and select the fitting function with the smallest average residual value as the classification output result. An adaptive duty cycle scheduling strategy for wireless sensor nodes is established to dynamically adjust high-energy-consuming operations such as data collection and inter-node communication. When the data change monitored on the node is obvious, the scheduling frequency is increased, while when the data change is relatively gentle, the scheduling frequency is decreased. The expression of the real-time scheduling frequency is: , Among them, and are the upper and lower limits of the scheduling frequency set for this node, represents the coefficient of variation of future time and is used to describe the degree of dispersion of data, and represent the maximum and minimum values of the coefficient of dispersion in historical data; The duty cycle scheduling algorithm introduces the change rate of the node's tracked data and its historical situation, so as to achieve a relatively complete tracking of the changes in environmental data and network load with less total sensing. As the basis for flexibly arranging tasks for sensor nodes, through the local time-series data prediction based on the decision tree, the changes and trends of the data related to the current node can be obtained more accurately, so as to adjust the sensing and communication frequencies of the node before the environment fluctuates; on the contrary, when the environment is relatively stable, the sensing and communication frequencies are reduced to minimize energy waste.
Citation Information
Cited By
Gas sensor response fingerprint feature extraction method based on parameter fitting
CN120724116A
A method for extracting response fingerprint features of a gas sensor based on parameter fitting
CN120724116B
