Industrial internet of things channel modeling method and system based on interpretable random forest

Abstract

The application relates to the technical field of wireless communication and machine learning cross-fusion, in particular to an industrial internet of things channel modeling method and system based on an interpretable random forest. The method comprises the following steps: constructing a channel data set containing multiple industrial scenes, a receiving point line-of-sight state and a three-dimensional position; after joint environment-height balance sampling and standardization preprocessing, a multi-output regression model is established by using a random forest, system parameters are taken as inputs, and a root mean square delay spread, an azimuth spread and an elevation spread are jointly predicted; further, the model efficiency and precision are improved by using a hyperparameter optimization with a time penalty term, and the model decision mechanism is explained from a global and local level by using permutation feature importance and SHAP value analysis, and the consistency with a physical propagation law is verified. The application realizes high-precision, stable and interpretable joint prediction of key statistical characteristics of an industrial wireless channel, and can support channel modeling, network planning and optimized deployment of an industrial internet of things communication system.

Description

Technical Field

[0001] This invention relates to the technical field of cross-integration of wireless communication and machine learning, and in particular to an industrial Internet of Things channel modeling method and system based on interpretable random forest. Background Technology

[0002] The Industrial Internet of Things (IIoT) is a core enabling technology for smart manufacturing, relying on efficient, stable, and low-latency wireless communication. Wireless channel models are the theoretical foundation for communication system design, simulation, and performance evaluation, and their accuracy directly determines the overall system performance. However, typical industrial environments (such as automated workshops and warehouses) are complex, containing dense metal equipment, mobile machinery, and various obstacles, leading to strong multipath effects, predominance of non-line-of-sight propagation, and spatial non-stationarity in signal propagation. These characteristics pose a significant challenge to traditional channel modeling methods.

[0003] Existing industrial channel modeling methods can be mainly divided into three categories: 1) Deterministic methods (such as ray tracing): Although this method can achieve high modeling accuracy, it relies heavily on detailed environmental geometry and material information and has huge computational costs, making it difficult to apply to large-scale or dynamically changing industrial scenarios; 2) Semi-statistical semi-deterministic methods (such as geometric random channel models): This method compromises between model flexibility and physical rationality, but its modeling accuracy is often limited in industrial environments with complex metal reflection and scattering; 3) Pure statistical methods: Based on empirical formulas fitted from measured or simulated data, these methods are simple to implement, but the models usually lack physical interpretability and are difficult to accurately characterize the complex nonlinear coupling relationships between multiple key channel statistical parameters.

[0004] In recent years, machine learning (ML) methods have demonstrated their potential in handling high-dimensional nonlinear problems in channel modeling. However, existing ML research largely focuses on predicting single parameters such as path loss, paying insufficient attention to the core statistical parameters that determine critical communication system designs (such as equalizers, modulation, and antenna array configurations)—root mean square delay spread (RMS-DS) and root mean square angle spread (RMS-AS, including azimuth and elevation angles). Furthermore, there is a lack of a unified, joint prediction framework for these parameters. In addition, industrial applications demand extremely high stability and reliability from models, while machine learning models are often considered "black boxes," lacking transparency and interpretability in their decision-making mechanisms. This severely hinders the practical deployment of such models in critical industrial tasks and the trust and adoption by engineers.

[0005] In summary, channel modeling in the industrial IoT environment urgently requires a new method that can simultaneously meet the following requirements: (1) robust learning from heterogeneous, unbalanced measured or simulated data; (2) joint and accurate prediction of multiple interrelated key statistical parameters; and (3) prediction results that are not only accurate but also physically interpretable so that engineers can understand, trust, and use them in practical decision-making scenarios such as network diagnostics and planning. Existing technologies are all unable to simultaneously meet the above comprehensive requirements. Summary of the Invention

[0006] Therefore, this invention aims to solve the problems of insufficient modeling accuracy, difficulty in capturing multi-parameter coupling relationships, low engineering credibility due to lack of model interpretability, and sensitivity to the data imbalance problem that is common in real-world environments in existing industrial channel modeling technologies.

[0007] To address the aforementioned issues, this invention proposes an industrial IoT channel modeling method and system based on interpretable random forests. This technical solution uses deployment environment parameters and receiver spatial locations as inputs, and constructs a high-precision, highly generalizable channel model by uniformly modeling and jointly predicting multiple key statistical parameters in the channel. Furthermore, by introducing a model interpretability analysis mechanism, the consistency between the model's decision-making process and the physical propagation mechanism is verified, thereby enhancing the reliability and acceptability of the model in practical industrial applications.

[0008] The industrial IoT channel modeling method based on interpretable random forest includes the following steps:

[0009] S1: Acquire channel data in the industrial IoT environment, extract environment type, line-of-sight conditions of receiving point, and three-dimensional coordinates of receiving point as input features, extract root mean square delay spread, root mean square azimuth spread, and root mean square elevation spread as output targets, perform environment-altitude joint grouping balanced sampling on the acquired input features and the output targets, and then perform hierarchical partitioning and standardization processing to obtain training set, validation set and test set;

[0010] S2: Based on the training set, a multi-output random forest regression model is constructed. The model is trained using the recursive node splitting and weighted mean square error minimization criteria to jointly predict multiple channel statistical parameters. During the training process, the model hyperparameters are systematically tuned using the validation set and a grid search strategy that incorporates time penalty terms to balance model prediction accuracy and training computation efficiency.

[0011] S3: Based on the optimized model and the test set, the feature importance method is used to analyze the global feature contribution of the prediction results of each channel statistical parameter; and the Shapley additive interpretation method is used to analyze the local prediction contribution of each feature in a single sample to multiple channel statistical parameters; the analysis results are compared with the physical propagation law of wireless signals to verify the rationality of the model decision mechanism, and finally an interpretability analysis report containing global and local interpretations is output.

[0012] In one embodiment of the present invention, step S1, which involves performing environment-height joint grouping balanced sampling on the acquired input features and the output target, followed by hierarchical partitioning and standardization to obtain the training set, validation set, and test set, specifically includes:

[0013] Based on environment type and receiver height The samples were jointly grouped to obtain the joint groups. ;

[0014] The sample size of each group is counted. Oversampling is performed on groups with a sample size below a preset threshold, and undersampling is performed on groups with a sample size above a preset threshold, in order to obtain a dataset with a balanced class distribution. ;

[0015] Based on the joint grouping For the dataset Perform stratified random partitioning, dividing the data into an initial training set, an initial validation set, and an initial test set according to a preset ratio, ensuring that the proportion of each group in the three subsets is consistent with the original distribution;

[0016] The input features and output targets of the initial training set, the initial validation set, and the initial test set are standardized respectively to obtain the final training set used for modeling. Validation set and test set .

[0017] In one embodiment of the present invention, the root mean square delay spread Root mean square azimuth extension Root mean square pitch angle extension The calculation expressions are as follows:

[0018] ,

[0019] ,

[0020] ,

[0021] Where K is the total number of multipaths. For the first The received power of each path, , , The first The path delay, azimuth angle, and elevation angle.

[0022] In one embodiment of the present invention, in step S2, a multi-output random forest regression model is constructed, and the model is trained using the recursive node splitting and weighted mean square error minimization criteria to jointly predict multiple channel statistical parameters as follows:

[0023] S21: Initialize a random forest containing T regression trees and set the initial set of hyperparameters. ,in The total number of regression trees, The maximum depth at which a single tree is allowed to grow. The minimum number of samples required for an internal node to continue splitting. Minimum number of samples required for leaf nodes This represents the maximum number of features randomly selected during node splitting.

[0024] S22: For each regression tree in the forest, perform the following training procedure:

[0025] S221: From the training set In this process, self-sampling is performed to generate a training subset of the regression tree;

[0026] S222: Starting from the root node, for each node to be split... Recursively perform the following operations:

[0027] Randomly select a feature of size p from all the input features. Feature subset traverse the feature subset Each feature in and all candidate split thresholds Based on the criterion of minimizing the weighted mean square error, the optimal splitting feature is selected. With split threshold ;

[0028] Based on the selected splitting characteristics With split threshold Split the current node into two child nodes, left and right;

[0029] S223: Repeat step S222 until the preset stopping condition is met, generating a complete regression tree;

[0030] S23: After training all T regression trees, the multi-output random forest regression model is formed; for new input feature vectors... The joint prediction output of the model The results are obtained by averaging the predictions from all regression trees:

[0031] ,in, is the predicted value of the t-th regression tree, which is the mean of the target vector output by all training samples in its corresponding leaf node.

[0032] In one embodiment of the invention, the optimal splitting feature is selected. With split threshold The method is as follows: the optimal splitting feature is obtained by minimizing the weighted mean square error of the left and right child nodes after the split. With split threshold The decision formula is as follows:

[0033] ,

[0034] in, It is based on the current node data, for features The generated set of all candidate split thresholds; For nodes Mean square error of the in-sample output target. , and For a node in the t-th regression tree The left and right child nodes generated after the split For nodes The mean vector of the output target for all samples.

[0035] In one embodiment of the present invention, during the training process, the method for optimizing the hyperparameters of the multi-output random forest regression model using the validation set and a grid search strategy incorporating time penalty terms is as follows:

[0036] S201: Set of hyperparameters Each parameter in the algorithm is assigned a set of candidate values, forming a discrete gridded search space.

[0037] S202: Traverse any k-th set of hyperparameter configurations within the gridded search space. Perform the following operations:

[0038] Using the hyperparameter configuration In the training set Retrain the multi-output random forest regression model and obtain the training time of this model. ;

[0039] The validation set was tested using a retrained model. Make predictions and calculate the overall standardized root mean square error on the validation set. Its comprehensive reflection of the model's joint prediction accuracy for multiple channel statistical parameters;

[0040] Based on the overall standardized root mean square error and the training time Construct the optimization objective function Calculate the current configuration The overall cost:

[0041] ,

[0042] in, These are preset tradeoff coefficients used to adjust the relative importance of prediction accuracy and model training efficiency in the optimization objective;

[0043] S203: Compare all hyperparameter configurations within the search space. Corresponding comprehensive cost , select to make The configuration with the smallest value is taken as the optimal hyperparameter configuration. ,Right now: ;

[0044] S204: Use the aforementioned optimal hyperparameter configuration Based on the training set The multi-output random forest regression model was retrained to obtain a model with optimized hyperparameters.

[0045] In one embodiment of the present invention, in step S3, the method for performing global feature contribution analysis on the prediction results of each channel statistical parameter using the permutation feature importance method is as follows:

[0046] The importance of permutation features is used in the test set. Upquantize the overall contribution of each input feature to the joint prediction of multiple channel statistical parameters; for the th Each input feature has an importance value. Defined as: applying this feature to the test set The increase in the model's prediction loss function after the values ​​in the range are randomly shuffled is calculated using the following formula:

[0047] ,in, For the model In the dataset The mean squared error loss is expressed as follows: , For dataset Total number of samples For the model to the first The input feature vector of each sample The output joint prediction vector of channel statistical parameters, For the first The vector of true channel statistics parameters corresponding to each sample For the test set The Middle A dataset in which the values ​​of each input feature are randomly shuffled;

[0048] By calculating all input features The values ​​are then sorted to obtain the global feature importance ranking result.

[0049] In one embodiment of the present invention, in step S3, the Shapley additive interpretation method is used to analyze the local prediction contribution of each feature in a single sample to multiple channel statistical parameters as follows:

[0050] For the test set any sample in Its model prediction value Additive decomposition is the sum of the contributions of each input feature:

[0051] ,in, The baseline value predicted by the model is the average predicted output of the model across all training samples. For the first The Shapley additive explanatory vector of a feature for the sample is obtained by calculating the weighted average of the differences in the model output before and after adding the feature to all possible feature subsets; The total number of input features; The number of channel statistical parameters for joint prediction;

[0052] Each Shapley additive interpretation value vector The corresponding feature was quantified to predict the final value of the current sample. The direction and extent of its contribution.

[0053] Based on the same inventive concept, this invention also provides an industrial IoT channel modeling system based on interpretable random forest, including: a data construction and preprocessing module, a model training and optimization module, and an interpretability analysis module;

[0054] The data construction and preprocessing module is used to acquire channel data in the industrial Internet of Things environment, extract environment type, line-of-sight conditions of receiving point, and three-dimensional coordinates of receiving point as input features, and extract root mean square delay spread, root mean square azimuth spread, and root mean square elevation spread as output targets. The acquired input features and output targets are subjected to environment-altitude joint grouping balanced sampling, and then subjected to hierarchical partitioning and standardization processing to obtain training set, validation set and test set.

[0055] The model training and optimization module is used to construct a multi-output random forest regression model based on the training set. The model is trained using recursive node splitting and weighted mean square error minimization criteria to jointly predict multiple channel statistical parameters. During the training process, the model hyperparameters are systematically tuned using the validation set and a grid search strategy that incorporates time penalty terms to balance model prediction accuracy and training computation efficiency.

[0056] The interpretability analysis module is used to perform global feature contribution analysis on the prediction results of each channel statistical parameter based on the optimized model and the test set, using the permutation feature importance method; and to analyze the local prediction contribution of each feature in a single sample to multiple channel statistical parameters using the Shapley additive interpretation value method; the analysis results are compared with the physical propagation law of wireless signals to verify the rationality of the model decision mechanism, and finally output an interpretability analysis report containing global and local interpretations.

[0057] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the industrial Internet of Things channel modeling method based on interpretable random forest.

[0058] Compared with the prior art, the above-described technical solution of the present invention has the following advantages:

[0059] First, it boasts high prediction accuracy, strong stability, and physical interpretability: by jointly predicting multiple key channel parameters using a multi-output random forest framework, it effectively captures their inherent coupling relationships, demonstrating superior and stable prediction accuracy in various industrial environments. More importantly, through integrated permutation feature importance (PFI) and Shapley additive interpretation (SHAP) value analysis, the model's decision-making process becomes transparent. The revealed feature-response relationship is highly consistent with the physical mechanism of wireless propagation, fundamentally solving the core pain point of lack of trust in "black box" models in industrial applications. This upgrades the prediction results from "numerical references" to "decision-making basis" that engineers can understand and accept.

[0060] Secondly, robust modeling capability for industrial data characteristics: The environment-height joint grouping balanced sampling strategy effectively overcomes the inherent bias in data distribution caused by differences in simulation or measurement ranges in different scenarios, freeing model training from being dominated by a few scenarios and significantly improving generalization ability and modeling fairness under diverse working conditions.

[0061] Third, it offers the advantages of systems engineering that balance performance and efficiency: the proposed hyperparameter optimization strategy, based on Pareto front analysis, achieves an optimal trade-off between model prediction accuracy and computational training costs, ensuring the method's practicality and deployability in resource-constrained industrial environments. The entire process does not rely on precise environmental geometric priors, lowering the implementation threshold.

[0062] Fourth, it provides an end-to-end decision support toolchain: the resulting methods and systems output not only high-precision predictions, but also quantitative reports on model performance and physical interpretations of the prediction results. This enables its direct application to practical industrial tasks such as network planning simulation, online root cause diagnosis of faults, and channel-aware dynamic system optimization, transforming academic modeling capabilities into actionable engineering productivity. Attached Figure Description

[0063] To make the content of this invention easier to understand, the invention will be further described in detail below with reference to specific embodiments and accompanying drawings.

[0064] Figure 1 This is a flowchart of the industrial IoT channel modeling method based on interpretable random forest provided in an embodiment of the present invention;

[0065] Figure 2 This is a schematic diagram of the channel dataset construction and preprocessing process provided in an embodiment of the present invention;

[0066] Figure 3 This is a schematic diagram of the multi-output random forest regression model architecture provided in an embodiment of the present invention;

[0067] Figure 4 This is a Pareto front diagram of the hyperparameter optimization process of the random forest provided in this embodiment of the invention;

[0068] Figure 5 This is a schematic diagram of the global permutation feature importance analysis results provided in the embodiments of the present invention, wherein (a) represents the root mean square time delay spread of each input feature. The arrangement of features is ranked by importance value (PFI), and (b) represents the extension of each input feature to the root mean square azimuth angle. The input features are sorted by their feature importance values ​​(PFI), and (c) represents the expansion of each input feature with respect to the root mean square pitch angle. Ranking by importance of features (PFI);

[0069] Figure 6 This is a waterfall combination plot of Shapley Additive Interpretation (SHAP) values ​​for a single test sample provided in an embodiment of the present invention, where (a) represents the root mean square delay spread. The SHAP value waterfall plot, (b) represents the root mean square azimuth spread. SHAP value waterfall plot, root mean square pitch angle extension SHAP value waterfall chart;

[0070] Figure 7 This is a module framework diagram of an industrial IoT channel modeling system based on interpretable random forest provided in an embodiment of the present invention;

[0071] Explanation of reference numerals in the accompanying drawings: 100, Data construction and preprocessing module; 200, Model training and optimization module; 300, Interpretability analysis module. Detailed Implementation

[0072] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, so that those skilled in the art can better understand and implement the present invention. However, the embodiments described are not intended to limit the present invention.

[0073] Example 1:

[0074] Reference Figure 1 As shown, this invention provides an industrial IoT channel modeling method based on interpretable random forests, comprising the following steps:

[0075] S1: Acquire channel data in the industrial IoT environment, extract environment type, line-of-sight conditions of receiving point, and three-dimensional coordinates of receiving point as input features, extract root mean square delay spread, root mean square azimuth spread, and root mean square elevation spread as output targets, perform environment-altitude joint grouping balanced sampling on the acquired input features and the output targets, and then perform hierarchical partitioning and standardization processing to obtain training set, validation set and test set;

[0076] S2: Based on the training set, a multi-output random forest regression model is constructed. The model is trained using the recursive node splitting and weighted mean square error minimization criteria to jointly predict multiple channel statistical parameters. During the training process, the model hyperparameters are systematically tuned using the validation set and a grid search strategy that incorporates time penalty terms to balance model prediction accuracy and training computation efficiency.

[0077] S3: Based on the optimized model and the test set, the feature importance method is used to analyze the global feature contribution of the prediction results of each channel statistical parameter; and the Shapley additive interpretation method is used to analyze the local prediction contribution of each feature in a single sample to multiple channel statistical parameters; the analysis results are compared with the physical propagation law of wireless signals to verify the rationality of the model decision mechanism, and finally an interpretability analysis report containing global and local interpretations is output.

[0078] Furthermore, such as Figure 2 As shown, step S1 aims to construct a high-quality, unbiased channel dataset suitable for model training and evaluation. The construction and preprocessing process of the dataset mainly includes two stages: data acquisition and feature extraction, and data preprocessing.

[0079] S11. Data Acquisition and Feature Definition

[0080] In this embodiment, channel data is acquired through high-fidelity ray tracing simulation. Simulation software (such as AltairWinProp) is used to perform 3D geometric and electromagnetic material modeling of various typical industrial environments (e.g., automated workshops, laboratories, warehouses, and semi-automated workshops). Simulation calculations generate channel impulse response data containing multipath parameters. Based on this simulation data, the model's input features and output targets are defined:

[0081] Input feature vector This includes: (1) Environment type E: Industrial scene category (such as workshop, warehouse, etc.) represented by classification codes. (2) Receiver point line-of-sight condition (LOS): Used to identify the propagation path status, where line-of-sight propagation is coded as 1 and non-line-of-sight propagation is coded as 2. (3) Receiver point three-dimensional coordinates : These represent the horizontal position and vertical height of the receiving point in the scene, respectively.

[0082] Output target vector The key statistical parameters of the channel to be jointly predicted include: root mean square delay spread. Root mean square azimuth extension Root mean square pitch angle extension Each parameter is calculated based on the multipath components, using the following formula:

[0083] ,

[0084] ,

[0085] ,

[0086] Where K is the total number of multipaths. For the first The received power of each path, , , The first The path delay, azimuth angle, and elevation angle.

[0087] S12, Data Preprocessing

[0088] To improve the robustness and generalization performance of the model training, the following preprocessing operations are performed on the extracted raw data:

[0089] To address the imbalance in environmental and spatial distribution caused by simulation settings in the original data, a joint balancing strategy is proposed. Each sample is categorized according to its environmental type. and receiver height The samples were jointly grouped to obtain the joint groups. ;

[0090] The sample size of each group is counted. Oversampling is performed on groups with a sample size below a preset threshold, and undersampling is performed on groups with a sample size above a preset threshold, in order to obtain a dataset with a balanced class distribution. ;

[0091] To maintain data distribution consistency among the training, validation, and test sets, based on the joint grouping... For the dataset Perform stratified random partitioning, dividing the data into an initial training set, an initial validation set, and an initial test set according to a preset ratio (e.g., 70%:15%:15%), ensuring that the proportion of each joint group in the three subsets is consistent with the original distribution;

[0092] To eliminate differences in feature units and accelerate model convergence, numerical features and output targets are standardized. The mean of each feature is calculated based on the initial training set. and standard deviation The input features and output targets of the initial training set, the initial validation set, and the initial test set are respectively subjected to Z-score transformation standardization: ;

[0093] Finally, the training set used for modeling is obtained. Validation set and test set .

[0094] Further, in step S2, construct as follows Figure 3 The multi-output random forest regression model shown employs recursive node splitting and a weighted mean square error minimization criterion for model training, and the method for jointly predicting multiple channel statistical parameters is as follows:

[0095] S21: Initialize a random forest containing T regression trees and set the initial set of hyperparameters. ,in This represents the total number of regression trees; This represents the maximum depth to which a single tree is allowed to grow. The minimum number of samples required for an internal node to continue splitting. Minimum number of samples required for inner leaf nodes; This is the maximum number of features to randomly select when splitting a node; the default is all features.

[0096] S22: For each regression tree in the forest, perform the following training procedure:

[0097] S221: From the training set The training subset of the regression tree is generated by sampling with replacement using the bootstrap method.

[0098] S222: Starting from the root node, for each node to be split... Recursively perform the following operations:

[0099] Randomly select a feature of size p from all the input features. Feature subset traverse the feature subset Each feature in and all candidate split thresholds Based on the criterion of minimizing the weighted mean square error, the optimal splitting feature is selected. With split threshold Among them, the optimal splitting feature is selected. With split threshold The method is as follows: the optimal splitting feature is obtained by minimizing the weighted mean square error of the left and right child nodes after the split. With split threshold The decision formula is as follows:

[0100] ,

[0101] in, It is based on the current node data, for features The generated set of all candidate split thresholds; For nodes Mean square error of the in-sample output target. , and For a node in the t-th regression tree The left and right child nodes generated after the split For nodes The mean vector of the output target for all samples in the dataset;

[0102] Based on the selected splitting characteristics With split threshold The current node Split into two child nodes, left and right. and ;

[0103] S223: Repeat step S222 until the preset stopping condition is met (such as the node reaching its maximum depth). or the number of samples within a node is less than At this point, the node becomes a leaf node and no longer splits, eventually resulting in a complete regression tree;

[0104] S23: After training all T regression trees, the multi-output random forest regression model is formed; for a new input feature vector... The prediction process is as follows:

[0105] The t-th regression tree is based on features Traverse from the root node to a leaf node. The predicted value of the tree Defined as the mean of the output target vectors of all training samples within this leaf node:

[0106] ;

[0107] The joint prediction output of the model It is the average of the predictions from all T regression trees, achieving integration and smoothing of the prediction results:

[0108] ,in, It is a joint prediction vector of three channel statistical parameters: root mean square delay spread, azimuth spread, and elevation spread.

[0109] During model training, the validation set is used. The method for hyperparameter optimization of the multi-output random forest regression model by incorporating a grid search strategy with a time penalty term is as follows:

[0110] S201: Set of hyperparameters Each parameter in the algorithm is assigned a set of candidate values, forming a discrete gridded search space.

[0111] S202: Traverse any k-th set of hyperparameter configurations within the gridded search space. Perform the following operations:

[0112] Using the hyperparameter configuration In the training set Retrain the multi-output random forest regression model and obtain the training time of this model. ;

[0113] The validation set was tested using a retrained model. Make predictions and calculate the overall standardized root mean square error on the validation set. Its comprehensive reflection of the model's joint prediction accuracy for multiple channel statistical parameters;

[0114] Based on the overall standardized root mean square error and the training time Construct the optimization objective function Calculate the current configuration The overall cost:

[0115] ,

[0116] in, This is a preset tradeoff coefficient used to adjust the relative importance of prediction accuracy and model training efficiency in the optimization objective. Users can dynamically adjust this coefficient based on the model training speed requirements of their actual deployment scenario. value: The larger the value, the stronger the penalty for training time, and the more the optimization process tends to choose hyperparameter configurations that enable faster training.

[0117] S203: Compare all hyperparameter configurations within the search space. Corresponding comprehensive cost , select to make The configuration with the smallest value is taken as the optimal hyperparameter configuration. ,Right now: ;

[0118] S204: Use the aforementioned optimal hyperparameter configuration Based on the training set The multi-output random forest regression model was retrained to obtain a model with optimized hyperparameters.

[0119] Based on the performance-cost trade-off obtained during hyperparameter optimization (such as... Figure 4 As shown in the Pareto front, this invention employs the "elbow" selection principle, a common principle in engineering trade-off analysis, to determine the final optimal hyperparameter configuration. The "elbow" refers to the inflection point region on the Pareto front where performance improvement tends to plateau while costs begin to increase significantly. Selecting a configuration in this region can achieve near-optimal model performance with relatively low computational cost.

[0120] Based on the above analysis, the optimal hyperparameter configuration is finally determined as follows: Compared to the initial configuration before optimization, the optimized configuration achieved both lower prediction error and shorter training time on the validation set, verifying the effectiveness of the grid search strategy with fused time penalty terms proposed in this invention in balancing model accuracy and computational efficiency.

[0121] To comprehensively evaluate the final performance of the optimized model, the test set was used. Tests were conducted, and the root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination were used. As a core evaluation metric. For any channel statistical parameter (can represent) Its definition is as follows:

[0122] ,

[0123] ,

[0124] ,

[0125] Where N is the total number of samples in the test set; and They represent the first The true value and the model prediction value of each sample; This is the mean of the true values ​​on the test set.

[0126] To verify the superiority of this invention, the optimized random forest model (denoted as RF-Opt) was compared with several state-of-the-art baseline models (including XGBoost, Multilayer Perceptron MLP, Convolutional Neural Network CNN, and Long Short-Term Memory Network LSTM) under the same dataset, preprocessing procedures, and evaluation criteria. The prediction performance comparison of each model on the test set is shown in Table 1.

[0127] Table 1 Comparison of prediction performance of each model on the test set

[0128]

[0129] Performance comparison results show that RF-Opt achieves the best performance in the prediction tasks of the three key channel statistical parameters: its RMSE and MAE values ​​are the lowest, while... All values ​​were the highest. Specifically, compared to the second-best performing XGBoost model, RF-Opt... , and The RMSE for prediction tasks decreased by approximately 18.7%, 27.8%, and 37.2%, respectively; the MAE decreased by approximately 34.9%, 41.8%, and 51.7%, respectively. These represent improvements of approximately 4.2%, 5.6%, and 5.7%, respectively. Compared to baseline neural network models (MLP, CNN, LSTM), RF-Opt demonstrates a more significant performance advantage.

[0130] Based on the above quantitative results, it can be seen that the RF-Opt method proposed in this invention not only performs best in terms of prediction error, but also has comprehensive advantages in terms of model fit and prediction stability. It is especially suitable for complex and variable channel characteristic modeling scenarios with multiple parameters coupled in the Industrial Internet of Things.

[0131] To verify the rationality of the model's decisions and improve its engineering credibility, in step S3, the present invention performs interpretability analysis on the optimized model, including global feature importance analysis and local prediction mechanism analysis.

[0132] Specifically, the global feature importance analysis employs a permutation feature importance method, and the method for analyzing the global feature contribution of the prediction results of each channel statistical parameter is as follows:

[0133] The importance of permutation features is used in the test set. Upquantize the overall contribution of each input feature to the joint prediction of multiple channel statistical parameters. For the ... Each input feature has an importance value. Defined as: applying this feature to the test set The increase in the model's prediction loss function after the values ​​in the matrix are randomly shuffled represents the magnitude of the decrease in the model's prediction performance. The specific calculation formula is as follows:

[0134] ,in, For the model In the dataset The mean squared error loss is calculated as follows: , For dataset Total number of samples For the model to the first The input feature vector of each sample The output joint prediction vector of channel statistical parameters, For the first The vector of true channel statistics parameters corresponding to each sample For the test set The Middle The dataset in which the values ​​of an input feature are randomly shuffled is an operation designed to disrupt the statistical association between the feature and the prediction target.

[0135] By calculating all input features The values ​​are then sorted to obtain a global feature importance ranking, thereby identifying the system parameters most critical to model prediction.

[0136] like Figure 5 As shown in (a) to (b), the global feature importance analysis reveals the differentiated impact of different input features on the prediction of various channel parameters: for the root mean square delay spread The environment type E and the horizontal coordinates (X, Y) of the receiving point are the most important features. For the root mean square azimuth extension... The horizontal coordinate (especially the X coordinate) plays a dominant role. For the root mean square pitch angle extension... The importance of the receiver height (Z coordinate) is significantly increased.

[0137] The ranking results are highly consistent with the physical laws governing the propagation of wireless signals in complex industrial environments: the environment type determines the distribution of macroscopic scatterers, the horizontal coordinate mainly affects the azimuth distribution of multipath propagation, and altitude is directly related to the propagation conditions in the pitch dimension. This indicates that the model has successfully learned the key geometric and environmental factors behind the channel characteristics, rather than blindly fitting data. This analysis not only enhances the engineering credibility of the model but also provides an interpretable theoretical basis for channel feature selection and ranging scheme design in practical communication systems.

[0138] Specifically, to further understand the model's decision-making logic at the sample level, the local prediction mechanism analysis employs the Shapley additive interpretation (SHAP) value method. The method for analyzing the local prediction contribution of each feature in a single sample to multiple channel statistical parameters is as follows:

[0139] For the test set any sample in Its model prediction value Additive decomposition is the sum of the contributions of each input feature:

[0140] ,in, The baseline value predicted by the model is the average predicted output of the model across all training samples. For the first The Shapley additive explanatory vector of a feature for the sample is obtained by calculating the weighted average of the differences in the model output before and after adding the feature to all possible feature subsets; The total number of input features; The number of channel statistical parameters for joint prediction;

[0141] Each Shapley additive interpretation value vector The corresponding feature was quantified to predict the final value of the current sample. The direction and magnitude of the contribution are as follows: a positive value indicates that the prediction is being pushed toward a higher value, while a negative value indicates that the prediction result is being suppressed.

[0142] like Figure 6 As shown in (a) to (b), taking a line-of-sight sample located at the boundary of a semi-automatic workshop as an example, its SHAP waterfall plot clearly shows the specific impact of each feature on the prediction of this sample: the effect of environment type E on the root mean square delay spread. It exhibits a positive contribution, consistent with the abundant multipath physical characteristics of this type of environment; the line-of-sight conditional LOS contributes to the root mean square delay spread. The negative contribution indicates that line-of-sight propagation helps reduce the root mean square delay spread. ; Extension of the horizontal coordinates of the receiving point to the root mean square azimuth angle With root mean square pitch angle extension All of these are the main positive driving factors, reflecting the key role of spatial location in the expansion of the angular domain.

[0143] The local interpretation results demonstrate that the model's micro-decision mechanism is highly consistent with the physical principles of wireless propagation, thus jointly verifying the rationality and interpretability of the constructed model at both the global feature importance and local prediction interpretation levels. The method proposed in this invention not only achieves high-precision joint prediction but also forms a reliable channel modeling framework with a transparent decision-making mechanism that conforms to physical laws, meeting the stringent requirements of the Industrial Internet of Things for model reliability and interpretability.

[0144] Example 2:

[0145] like Figure 7 As shown, based on the same inventive concept as Embodiment 1, the present invention also provides an industrial Internet of Things channel modeling system based on interpretable random forest, the system comprising: a data construction and preprocessing module 100, a model training and optimization module 200, and an interpretability analysis module 300;

[0146] The data construction and preprocessing module 100 is used to acquire channel data in the industrial Internet of Things environment, extract environment type, line-of-sight conditions of receiving point, and three-dimensional coordinates of receiving point as input features, and extract root mean square delay spread, root mean square azimuth spread, and root mean square elevation spread as output targets. The acquired input features and output targets are subjected to environment-altitude joint grouping balanced sampling, and then subjected to hierarchical partitioning and standardization processing to obtain training set, validation set and test set.

[0147] The model training and optimization module 200 is used to construct a multi-output random forest regression model based on the training set, and to train the model using the recursive node splitting and weighted mean square error minimization criteria to jointly predict multiple channel statistical parameters. During the training process, the model hyperparameters are systematically tuned using the validation set through a grid search strategy that incorporates time penalty terms to balance model prediction accuracy and training computation efficiency.

[0148] The interpretability analysis module 300 is used to perform global feature contribution analysis on the prediction results of each channel statistical parameter based on the optimized model and the test set, using the permutation feature importance method; and to analyze the local prediction contribution of each feature in a single sample to multiple channel statistical parameters using the Shapley additive interpretation value method; compare the analysis results with the physical propagation law of wireless signals to verify the rationality of the model decision mechanism, and finally output an interpretability analysis report containing global and local interpretations.

[0149] This embodiment proposes an industrial IoT channel modeling system based on interpretable random forest, which is used to implement the aforementioned industrial IoT channel modeling method based on interpretable random forest. Therefore, the specific implementation of the industrial IoT channel modeling system based on interpretable random forest can be found in the embodiment section of the aforementioned industrial IoT channel modeling method based on interpretable random forest. For example, the data construction and preprocessing module 100, the model training and optimization module 200, and the interpretability analysis module 300 are respectively used to implement steps S1, S2, and S3 in the industrial IoT channel modeling method based on interpretable random forest in Embodiment 1. Therefore, its specific implementation can be referred to the description of the corresponding embodiments. To avoid redundancy, it will not be repeated here.

[0150] Example 3:

[0151] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, it implements the industrial Internet of Things channel modeling method based on interpretable random forest described in Embodiment 1.

[0152] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0153] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0154] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0155] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0156] Obviously, the above embodiments are merely illustrative examples for clear explanation and are not intended to limit the implementation. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively list all possible implementations here. However, obvious variations or modifications derived therefrom are still within the scope of protection of this invention.

Claims

1. A channel modeling method for industrial IoT based on interpretable random forest, characterized in that, Includes the following steps: S1: Acquire channel data in the industrial IoT environment, extract environment type, line-of-sight conditions of receiving point, and three-dimensional coordinates of receiving point as input features, extract root mean square delay spread, root mean square azimuth spread, and root mean square elevation spread as output targets, perform environment-altitude joint grouping balanced sampling on the acquired input features and the output targets, and then perform hierarchical partitioning and standardization processing to obtain training set, validation set and test set; S2: Based on the training set, a multi-output random forest regression model is constructed. The model is trained using the recursive node splitting and weighted mean square error minimization criteria to jointly predict multiple channel statistical parameters. During training, the model hyperparameters are systematically tuned using the validation set and a grid search strategy that incorporates time penalty terms, in order to balance model prediction accuracy and training computation efficiency. S3: Based on the optimized model and the test set, the feature importance method is used to analyze the global feature contribution of the prediction results of each channel statistical parameter; and the Shapley additive interpretation method is used to analyze the local prediction contribution of each feature in a single sample to multiple channel statistical parameters; the analysis results are compared with the physical propagation law of wireless signals to verify the rationality of the model decision mechanism, and finally an interpretability analysis report containing global and local interpretations is output. In step S2, a multi-output random forest regression model is constructed. The model is trained using recursive node splitting and weighted mean square error minimization criteria to jointly predict multiple channel statistical parameters, as follows: S21: Initialize a random forest containing T regression trees and set the initial set of hyperparameters. ,in The total number of regression trees, The maximum depth at which a single tree is allowed to grow. The minimum number of samples required for an internal node to continue splitting. Minimum number of samples required for leaf nodes This represents the maximum number of features randomly selected during node splitting. S22: For each regression tree in the forest, perform the following training procedure: S221: From the training set In this process, self-sampling is performed to generate a training subset of the regression tree; S222: Starting from the root node, for each node to be split... Recursively perform the following operations: Randomly select a feature of size p from all the input features. Feature subset traverse the feature subset Each feature in and all candidate split thresholds Based on the criterion of minimizing the weighted mean square error, the optimal splitting feature is selected. With split threshold ; Based on the selected splitting characteristics With split threshold Split the current node into two child nodes, left and right; S223: Repeat step S222 until the preset stopping condition is met, generating a complete regression tree; S23: After training all T regression trees, the multi-output random forest regression model is formed; for new input feature vectors... The joint prediction output of the model The results are obtained by averaging the predictions from all regression trees: ， in, is the predicted value of the t-th regression tree, which is the mean of the target vector output by all training samples in its corresponding leaf node.

2. The industrial IoT channel modeling method based on interpretable random forest according to claim 1, characterized in that: In step S1, the method for performing environment-height joint grouping balanced sampling on the acquired input features and the output target, followed by hierarchical partitioning and standardization to obtain the training set, validation set, and test set specifically includes: Based on environment type and receiver height The samples were jointly grouped to obtain the joint groups. ; The sample size of each group is counted. Oversampling is performed on groups with a sample size below a preset threshold, and undersampling is performed on groups with a sample size above a preset threshold, in order to obtain a dataset with a balanced class distribution. ; Based on the joint grouping For the dataset Perform stratified random partitioning, dividing the data into an initial training set, an initial validation set, and an initial test set according to a preset ratio, ensuring that the proportion of each group in the three subsets is consistent with the original distribution; The input features and output targets of the initial training set, the initial validation set, and the initial test set are standardized respectively to obtain the final training set used for modeling. Validation set and test set .

3. The industrial IoT channel modeling method based on interpretable random forest according to claim 1, characterized in that: The root mean square delay spread Root mean square azimuth extension Root mean square pitch angle extension The calculation expressions are as follows: ， ， ， Where K is the total number of multipaths. For the first The received power of each path, , , The first The path delay, azimuth angle, and elevation angle.

4. The industrial IoT channel modeling method based on interpretable random forest according to claim 1, characterized in that: Select the optimal splitting feature With split threshold The method is as follows: the optimal splitting feature is obtained by minimizing the weighted mean square error of the left and right child nodes after the split. With split threshold The decision formula is as follows: ， in, It is based on the current node data, for features The generated set of all candidate split thresholds; For nodes Mean square error of the in-sample output target. , and For a node in the t-th regression tree The left and right child nodes generated after the split For nodes The mean vector of the output target for all samples.

5. The industrial IoT channel modeling method based on interpretable random forest according to claim 4, characterized in that: During training, the hyperparameters of the multi-output random forest regression model are optimized using the validation set and a grid search strategy incorporating time penalty terms, as follows: S201: Set of hyperparameters Each parameter in the algorithm is assigned a set of candidate values, forming a discrete gridded search space. S202: Traverse any k-th set of hyperparameter configurations within the gridded search space. Perform the following operations: Using the hyperparameter configuration In the training set Retrain the multi-output random forest regression model and obtain the training time of this model. ; The validation set was tested using a retrained model. Make predictions and calculate the overall standardized root mean square error on the validation set. Its comprehensive reflection of the model's joint prediction accuracy for multiple channel statistical parameters; Based on the overall standardized root mean square error and the training time Construct the optimization objective function Calculate the current configuration The overall cost: ， in, These are preset tradeoff coefficients used to adjust the relative importance of prediction accuracy and model training efficiency in the optimization objective; S203: Compare all hyperparameter configurations within the search space. Corresponding comprehensive cost , select to make The configuration with the smallest value is taken as the optimal hyperparameter configuration. ,Right now: ; S204: Use the aforementioned optimal hyperparameter configuration Based on the training set The multi-output random forest regression model was retrained to obtain a model with optimized hyperparameters.

6. The industrial IoT channel modeling method based on interpretable random forest according to claim 1, characterized in that: In step S3, the following method is used to analyze the global feature contribution of the prediction results of each channel statistical parameter using the permutation feature importance method: The importance of permutation features is used in the test set. Upquantize the overall contribution of each input feature to the joint prediction of multiple channel statistical parameters; for the th Each input feature has an importance value. Defined as: applying this feature to the test set The increase in the model's prediction loss function after the values ​​in the range are randomly shuffled is calculated using the following formula: ， in, For the model In the dataset The mean squared error loss is expressed as follows: , For dataset Total number of samples For the model to the first The input feature vector of each sample The output joint prediction vector of channel statistical parameters, For the first The vector of true channel statistics parameters corresponding to each sample For the test set The Middle A dataset in which the values ​​of each input feature are randomly shuffled; By calculating all input features The values ​​are then sorted to obtain the global feature importance ranking result.

7. The industrial IoT channel modeling method based on interpretable random forest according to claim 6, characterized in that: In step S3, the Shapley additive interpretation method is used to analyze the local prediction contribution of each feature in a single sample to multiple channel statistical parameters, as follows: For the test set any sample in Its model prediction value Additive decomposition is the sum of the contributions of each input feature: ， in, The baseline value predicted by the model is the average predicted output of the model across all training samples. For the first The Shapley additive explanatory vector of a feature for the sample is obtained by calculating the weighted average of the differences in the model output before and after adding the feature to all possible feature subsets; The total number of input features; The number of channel statistical parameters for joint prediction; Each Shapley additive interpretation value vector The corresponding feature was quantified to predict the final value of the current sample. The direction and extent of its contribution.

8. An industrial IoT channel modeling system based on interpretable random forest, characterized in that, Includes the following modules: The data construction and preprocessing module is used to acquire channel data in the industrial Internet of Things environment, extract environment type, line-of-sight conditions of receiving point, and three-dimensional coordinates of receiving point as input features, and extract root mean square delay spread, root mean square azimuth spread, and root mean square elevation spread as output targets. The acquired input features and output targets are subjected to environment-altitude joint grouping balanced sampling, and then subjected to hierarchical partitioning and standardization processing to obtain training set, validation set and test set. The model training and optimization module is used to construct a multi-output random forest regression model based on the training set, and to train the model using the recursive node splitting and weighted mean square error minimization criteria to jointly predict multiple channel statistical parameters. During training, the model hyperparameters are systematically tuned using the validation set and a grid search strategy that incorporates time penalty terms, in order to balance model prediction accuracy and training computation efficiency. A multi-output random forest regression model was constructed, and recursive node splitting and weighted mean square error minimization criteria were used for model training to jointly predict multiple channel statistical parameters. Initialize a random forest containing T regression trees and set the initial set of hyperparameters. ,in The total number of regression trees, The maximum depth at which a single tree is allowed to grow. The minimum number of samples required for an internal node to continue splitting. Minimum number of samples required for leaf nodes This represents the maximum number of features randomly selected during node splitting. For each regressing tree in the forest, perform the following training process: From the training set In this process, self-sampling is performed to generate a training subset of the regression tree; Starting from the root node, for each node to be split... Recursively perform the following operations: Randomly select a feature of size p from all the input features. Feature subset traverse the feature subset Each feature in and all candidate split thresholds Based on the criterion of minimizing the weighted mean square error, the optimal splitting feature is selected. With split threshold ; Based on the selected splitting characteristics With split threshold Split the current node into two child nodes, left and right; Repeat the previous step until the preset stopping condition is met, generating a complete regression tree; After training all T regression trees, the multi-output random forest regression model is formed; for a new input feature vector... The joint prediction output of the model The results are obtained by averaging the predictions from all regression trees: ， in, The predicted value of the t-th regression tree is the mean of the target vector output by all training samples in its corresponding leaf node. The interpretability analysis module is used to perform global feature contribution analysis on the prediction results of each channel statistical parameter based on the optimized model and the test set, using the permutation feature importance method; and to analyze the local prediction contribution of each feature in a single sample to multiple channel statistical parameters using the Shapley additive interpretation value method; the analysis results are compared with the physical propagation law of wireless signals to verify the rationality of the model decision mechanism, and finally output an interpretability analysis report containing global and local interpretations.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the industrial IoT channel modeling method based on interpretable random forest as described in any one of claims 1 to 7.

Images