Time series prediction feature selection method and system based on partial autocorrelation function (pacf) and mutual information (mi) dynamic collaboration

Abstract

1. Invention name: Time series prediction feature selection method and system based on partial autocorrelation function (PACF) and mutual information (MI) dynamic cooperation 2. Technical field: The present invention belongs to the technical field of time series prediction, and relates to a method and system for dynamically screening linear and nonlinear features, which are suitable for industrial equipment predictive maintenance, energy scheduling, financial risk control and other scenes. 3. Technical scheme: By detecting the non-stationarity (ADF test) and memory (Hurst index) of time series data, the weight parameter α of PACF and MI is dynamically adjusted (non-stationary data α∈[0.2, 0.4], strong memory data α∈[0.6, 0.8]), the mixed score \text{Score}(k)=\alpha\cdot|\text{PACF}(k)|+(1‑\alpha)\cdot\text{MI}(k) is calculated, and the feature window is optimized by combining forward selection and backward pruning (correlation coefficient threshold 0.8). 4. Technical effect: The fault warning accuracy is improved by 17.7% (wind power data); the feature redundancy is reduced by 35%-53%; the data processing efficiency is improved by 50%.

Description

A Time-Series Prediction Feature Selection Method and System Based on Dynamic Coordination of Partial Autocorrelation Function (PACF) and Mutual Information (MI) Technical Field

[0001] This invention belongs to the field of time series prediction technology, specifically involving a feature selection method and system based on the dynamic synergy of partial autocorrelation function (PACF) and mutual information (MI). It is applicable to predictive maintenance of industrial equipment, energy dispatching, financial risk control and IoT sensor monitoring, supporting intelligent manufacturing and smart management. Background Technology

[0002] Existing technical problems

[0003] Traditional time series forecasting methods have the following drawbacks:

[0004] 1. Limitations of fixed weights: Using a fixed weight combination of PACF and MI cannot adapt to the non-stationarity of the data, noise interference, and differences in memory.

[0005] 2. Severe feature redundancy: High redundancy in candidate features leads to model overfitting;

[0006] 3. Insufficient coverage by a single indicator: Existing methods only target linear or nonlinear features (e.g., Zhao et al. only used conditional mutual information (CMI) to screen nonlinear features).

[0007] Comparison of existing technologies

[0008] Compared with traditional methods, Zhao et al. (2018) proposed a weight adjustment strategy based on fixed weights, dynamic adjustment based on CMI, dynamic collaborative feature coverage based on ADF+Hurst, linear + nonlinear, only nonlinear, linear + nonlinear collaborative screening, redundancy control method, no optimization strategy, dynamic window partitioning, forward selection + backward pruning. surface

[0009] *Note: The research by Zhao et al. (2018) (title: Feature Selection for Nonstationary Data Using Conditional Mutual Information, published in IEEE Access, Vol. 6, pp. 58211-58221) and IEEE Xplore proposed a dynamic weight adjustment strategy based on conditional mutual information (CMI). However, their method is only applicable to nonlinear feature selection and does not address the following issues: it does not incorporate the selection of linear features (such as PACF lag terms); and it does not introduce redundancy control algorithms (such as backward pruning), resulting in high feature redundancy.*

[0010] Traditional time series forecasting methods also have the following specific drawbacks:

[0011] 1. Limitations of fixed weights: Existing technologies use fixed weights to combine the partial autocorrelation function (PACF) and mutual information (MI), which cannot adaptively adjust according to data nonstationarity, noise interference, and differences in memory.

[0012] 2. Insufficient coverage by a single indicator: (Literature title: Feature Selection for Nonstationary Data Using Conditional Mutual Information, published in IEEE Access Vol. 6) proposes a dynamic weight adjustment strategy based on conditional mutual information (CMI), but its method is **only applicable to nonlinear feature selection** and has the following problems:

[0013] - Collaborative screening without incorporating linear features (such as PACF lag terms);

[0014] - The lack of redundancy control algorithms (such as backward pruning) results in high feature redundancy.

[0015] The improvement direction of this invention is to solve the problems of linear / nonlinear feature imbalance and redundancy by using a dynamic weight adaptation mechanism (combining ADF test and Hurst exponent) and window optimization algorithm (forward selection and backward pruning). Summary of the Invention

[0016] Technical solution

[0017] 1. Data Characteristic Detection

[0018] - Stationarity test: The ADF test was used, and the system was considered stationary when the p-value was <0.05;

[0019] - Memory test: Hurst index is used, and a value > 0.7 is considered to indicate strong memory.

[0020] 2. Dynamic weight adjustment rules

[0021] - Stationary data: α∈[0.2,0.4];

[0022] - Strong memory data: α∈[0.6,0.8];

[0023] - Priority: Stability > Strong memory.

[0024] 3. Mixed scoring calculation

[0025] Score(k) =α·|PACF(k)| +(1−α)·MI(k)

[0026] 4. Feature Filtering and Window Optimization

[0027] - Dynamic threshold: determined by the top 20% quantile or cross-validation;

[0028] - Window optimization: forward selection (in descending order of score) and backward pruning (correlation coefficient threshold 0.8).

[0029] Technical effect

[0030] 1. Improved dynamic adaptability: The accuracy of wind power data fault early warning has increased by 17.7%;

[0031] 2. Reduced redundancy: Feature redundancy is reduced by 35%-53%;

[0032] 3. Optimized computational efficiency: Data processing efficiency improved by 50%. Attached Figure Description

[0033] 1. Figure 1: System architecture diagram of this application

[0034] 2. Figure 2: Flowchart of the system algorithm of this application.

[0035] 3. Figure 3: Dynamic weight logic diagram of this application. Detailed Implementation

[0036] Example 1: Wind Turbine Gearbox Fault Prediction

[0037] - Data source: Vibration data of a wind farm over one year (sampling frequency 100 Hz);

[0038] - Parameter settings: α=0.3 (stationary data), threshold Score=0.65;

[0039] - Comparison of effects:

[0040] Method Fault Early Warning Accuracy Feature Redundancy Calculation Time (seconds) Zhao et al. (2018, CMI) 85.6% 32% 18.2 This Invention (α=0.3) 93.5% 15% 11.8 surface

[0041] Textual description: Compared with the single CMI method of Zhao et al. (2018), this invention improves the accuracy by 7.9% and reduces the redundancy by 53% through PACF-MI dynamic collaboration, proving the necessity of collaboration between linear and nonlinear features.

[0042] Example 2: Stock Price Trend Prediction

[0043] - Dynamic adjustment: α=0.7 (strong memory data), threshold Score=0.72;

[0044] - Results: Accuracy improved by 12.3%, efficiency improved by 50%.

[0045] The method's accuracy and efficiency have been improved by 12.3% to 50%. surface

[0046] Example 3: IoT Sensor Data Prediction

[0047] - Threshold optimization: Grid search dynamically adjusts the score from 0.5 to 0.8;

[0048] - Result: F1 score 0.91, RMSE = 0.087.

[0049] The F1 score (RMSE) for this invention is 0.91 to 0.087. surface

[0050] Example 4: Validation with Extremely Non-stationary Data

[0051] - Simulation data: ARIMA(1,1,1) + Gaussian noise + impulse interference;

[0052] - Strategy validation: Prioritizing stationary response (α=0.3) improves accuracy by an additional 9.1% compared to response memory (α=0.7).

[0053] - Data Source and Comparative Experiment: The publicly available dataset "Nonstationary_TS_2023" was used. The comparative experimental data is as follows:

[0054] Original accuracy (%) Optimized accuracy (%) Accuracy improvement Priority response stationarity (α=0.3) 72.5 91.6 19.1% Response memory (α=0.7) 72.5 82.5 10.0% surface

Claims

1. A method for selecting features for time series prediction, characterized in that the method is applied to vibration analysis of industrial equipment or monitoring by IoT sensors, and includes the following steps: 1.1 Performing stationarity detection and memory detection on the input time series data, wherein: the stationarity detection uses the ADF test, and the data is determined to be stationary when the p-value of the ADF test is less than 0.05; the memory detection uses the Hurst exponent, and the data is determined to have strong memory when the Hurst exponent is greater than 0.7; 1.2 Dynamically set the weight parameter α in the mixed scoring based on the detection results of step 1.1, where: When the data is stationary, the value of α ranges from [0.2, 0.4]; when the data has strong memory, the value of α ranges from [0.6, 0.8]; when the data satisfies both stationarity and strong memory, the value of α ranges from [0.2, 0.4]. 1.3 Calculate the absolute value of the partial autocorrelation function |PACF(k)| and the mutual information value MI(k) of the candidate features respectively, and calculate the mixed score according to the following formula: Score(k) = α·|PACF(k)| + (1−α)·MI(k) 1.4 Sort the candidate features in descending order of Score(k), and select the features whose Score(k) is greater than the dynamic threshold, which is the top 20% quantile of the candidate feature Score or the optimal value determined by cross-validation; 1.5 Window optimization is performed on the filtered features, including: forward selection: features are added sequentially from high to low according to Score(k); Backward pruning: Remove redundant features from the candidate features whose pairwise correlation coefficients are greater than 0.

8.

2. The method according to claim 1, characterized in that: The dynamic threshold may be automatically selected within the range of Score = 0.5 to 0.8 using a grid search method.

3. A time series prediction system for implementing the time series prediction feature selection method as described in claim 1 or 2, characterized in that, include: 3.1 Data acquisition module, including a multi-channel sensor interface and signal conditioning circuit, for real-time acquisition of vibration signals from industrial equipment; 3.2 Feature selection module, including an ADF test unit, a Hurst exponent calculation unit, and a dynamic weight controller, wherein the dynamic weight controller dynamically adjusts the weight parameter α in the mixed score based on the ADF test results and the Hurst exponent; 3.3 Model prediction module, based on an LSTM neural network processor, inputs the filtered features output by the feature selection module to generate time-series prediction results; 3.4 Data output module, including RS-485 communication interface and visualization terminal driver circuit, is used to transmit prediction results to intelligent manufacturing system or display device.

Images