CNN high-dimensional hyperparameter lightweight adaptive optimization method for non-stationary time series classification

Abstract

This invention discloses a lightweight adaptive optimization method for high-dimensional hyperparameters of convolutional neural networks (CNNs) for non-stationary time-series signal classification. It aims to address the technical challenges of performance degradation in time-series signal classification models and the reliance on expensive real-world evaluations for hyperparameter configuration under non-stationary perturbation scenarios. This method uses a deep convolutional neural network as the core classification carrier, treating the hyperparameter combinations within the deep convolutional neural network as decision variables to be optimized. With robust classification error rate, computational complexity, and training time as core optimization objectives, it constructs a closed-loop collaborative optimization mechanism of "perception-evaluation-decision" and utilizes a meta-learning dual-branch convolutional polynomial surrogate-assisted evolutionary algorithm (MetaDCP-SAEA) to achieve efficient configuration. This method requires no manual intervention; the convolutional neural network used for classifying non-stationary time-series signals can automatically search for the optimal hyperparameter combination. In simulated non-stationary noise environments, the reduction in classification accuracy can be controlled within 9.17%. It is suitable for robust classification scenarios of non-stationary time-series signals such as industrial IoT monitoring and medical signal diagnosis. It helps to lower the engineering threshold of artificial intelligence technology, promotes the large-scale application of automatic machine learning in complex environments, and has broad market prospects and application value.

Description

Technical Field

[0001] This invention belongs to the field of deep learning and automated machine learning technology, specifically involving the intersection of model hyperparameter optimization and non-stationary temporal signal processing. It is particularly suitable for robust classification tasks of non-stationary temporal signals in scenarios such as industrial IoT monitoring and medical signal diagnosis. It provides an efficient and low-cost lightweight adaptive optimization method for high-dimensional hyperparameters of Convolutional Neural Network (CNN) deep classification networks in non-stationary environments. Background Technology

[0002] With the rapid development of deep learning technology, deep convolutional neural networks, with their superior feature extraction and pattern recognition capabilities, have become a core technological support for time-series signal classification tasks, and have been widely used in key areas such as industrial weld defect detection, medical heart sound / ECG signal diagnosis, and equipment vibration monitoring. Especially in complex scenarios, deep networks can capture subtle changes in time-series signals through multi-level feature fusion, providing a technical guarantee for accurate classification. However, the classification performance of deep convolutional neural networks is highly dependent on hyperparameter configuration. As the complexity of the network architecture increases, the dimensionality of the hyperparameters to be optimized grows exponentially, forming a high-dimensional, nonlinear, and strongly coupled search space. Traditional methods of configuring hyperparameters based on human experience are not only inefficient but also struggle to find the globally optimal solution in the high-dimensional space, thus limiting model performance. Furthermore, time-series signals in real-world applications generally exhibit non-stationary characteristics, such as sensor drift, environmental noise fluctuations, and changes in equipment operating conditions. This makes hyperparameters optimized based on static data prone to degradation in classification performance after deployment, severely impacting the model's generalization ability and reliability. For example, spectral drift caused by bearing wear in industrial production increases the false alarm rate in defect detection, while equipment differences and individual physiological differences in medical diagnosis reduce the accuracy of heart sound signal classification. To address the efficiency issue in hyperparameter configuration, surrogate-assisted evolutionary algorithms have emerged. By constructing inexpensive surrogate models to approximate real evaluations, they effectively reduce computational costs and have been applied in some engineering optimization scenarios. However, in the hyperparameter configuration task for robust classification of non-stationary time-series signals, existing technologies still face many key challenges that urgently need to be addressed: First, the surrogate model has insufficient modeling capabilities. There are heterogeneous dependencies between hyperparameters, including local tight coupling and cross-dimensional long-range collaboration. Traditional single-scale surrogate models (such as Gaussian processes and single perceptron networks) cannot accurately capture such complex relationships, resulting in low fitting accuracy and poor prediction stability under small sample conditions, and thus failing to provide reliable support for hyperparameter optimization. Secondly, there is a lack of adaptability to non-stationary environments. Existing hyperparameter optimization strategies are mostly based on the assumption of a "static evaluation environment" and adopt point estimation-based decision-making logic, ignoring the uncertainty of prediction results under non-stationary disturbances. When the data distribution drifts, surrogate models trained on historical static data are prone to producing misleading predictions, causing valuable real evaluation resources to be allocated to high-risk areas, resulting in decreased optimization efficiency. Third, the optimization objective design has limitations. Existing methods mostly focus on static performance indicators such as classification accuracy and computational cost, without incorporating "robustness under non-stationary perturbations" into the explicit optimization objective. This leads to the hyperparameter combinations obtained from the search being prone to performance abrupt changes in distribution drift scenarios, making it difficult to meet the requirements of model stability in practical applications. Fourth, the evaluation cost remains high. Each real training and evaluation of deep networks requires a large amount of computing resources and time. Especially in high-dimensional hyperparameter spaces, traditional optimization methods need to verify the effectiveness of hyperparameters through a large number of repeated real evaluations, resulting in high configuration costs and severely limiting their large-scale application in engineering scenarios. Although some multi-objective hyperparameter optimization methods have emerged in the existing technology, such as configuration schemes based on particle swarm optimization and hierarchical pyramid networks, as well as the Surrogate-Assisted Evolutionary Algorithm (SAEA), these methods still have not solved the three core problems of heterogeneous dependency modeling, dynamic risk perception, and robust optimization in non-stationary environments, making it difficult to achieve efficient and reliable hyperparameter configuration for classification of non-stationary time series signals. Therefore, in response to the technical pain points of hyperparameter configuration relying on expensive evaluation and insufficient robustness under non-stationary perturbations, the development of a hyperparameter adaptive optimization method that can accurately capture the heterogeneous dependence of hyperparameters, dynamically perceive environmental risks, and balance performance and robustness has become an urgent need to promote the application of deep convolutional neural networks in complex engineering scenarios. This is of great significance for reducing the engineering threshold of artificial intelligence technology and helping the intelligent upgrading of industries. Summary of the Invention

[0003] To address the shortcomings of existing technologies, the present invention aims to provide an optimization method for intelligent hyperparameter configuration that is non-manual, low-cost, and highly reliable. This invention is achieved using the following technical solution: A lightweight adaptive optimization method for high-dimensional hyperparameters in CNNs for non-stationary temporal signal classification is proposed. It uses a deep convolutional classification network as the core classification carrier and a meta-learning dual-branch convolutional polynomial surrogate-assisted evolutionary algorithm (MetaDCP-SAEA) as the optimization strategy. The combination of hyperparameters to be optimized is defined as individuals in the population. Through a three-layer collaborative architecture of "perception-evaluation-decision," it achieves efficient multi-objective configuration of hyperparameters under non-stationary perturbations, ultimately outputting the optimal set of non-dominated hyperparameters and their corresponding target values ​​on the Pareto front. The co-evolutionary agent architecture includes a structure-aware dual-branch agent module, a meta-learning-enhanced uncertainty assessment module, and a three-objective adaptive decision-making module. The functions and collaborative logic of each module are as follows: The core architecture's collaborative logic is as follows: First, through a structure-aware dual-branch proxy module, the complex nonlinear mapping relationship between high-dimensional hyperparameters and robust classification performance is efficiently fitted, providing an accurate performance prediction basis for subsequent evaluation and decision-making. Then, through a meta-learning-enhanced uncertainty assessment module, the uncertainty risk in the prediction process is dynamically quantified, enabling perception of data distribution drift in non-stationary environments. Finally, through a three-objective adaptive decision-making module, candidate hyperparameter combinations that balance performance, robustness, and exploratory capabilities are selected. After real evaluation, feedback is used to update the preceding modules, forming a closed-loop optimization and completing the model hyperparameter configuration for robust classification of non-stationary time-series signals. The core functions of each module are as follows: **Structured Awareness Dual-Branch Proxy Module:** Addressing the core pain point in the background technology where "traditional single-scale Proxy models struggle to accurately capture heterogeneous dependencies between hyperparameters, exhibiting low fitting accuracy and poor prediction stability under small sample conditions," this module serves as the core of the perception layer in the "perception-evaluation-decision" architecture of this invention. It employs a heterogeneous architecture with parallel local and long-range branches to collaboratively model the local tight coupling and cross-dimensional long-range dependencies between hyperparameters. Through polynomial nonlinear mapping, it enhances fitting accuracy and prediction stability under small sample conditions, outputting robust classification performance predictions. This provides the basic input for the subsequent meta-learning-enhanced uncertainty assessment module, offering reliable prediction support for risk assessment and hyperparameter decision-making. **Meta-Learning-Enhanced Uncertainty Assessment Module:** Addressing the core pain point in the background technology where "existing hyperparameter optimization strategies are based on static evaluation environment assumptions, neglecting the uncertainty of prediction results under non-stationary perturbations, and easily generating misleading predictions when data distribution drifts," this module serves as the core of the "perception-evaluation-decision" architecture of this invention. The core of the architecture's evaluation layer receives the robust classification performance predictions output by the structural perception dual-branch agent module of the perception layer. It employs a two-stage mechanism of "parameter perturbation integration + meta-regression calibration" to dynamically quantify the uncertainty risk during the prediction process and complete risk calibration. This effectively distinguishes between cognitive uncertainty and data uncertainty, outputting a calibrated prediction risk score. This provides the core basis for risk decision-making in the subsequent three-objective adaptive decision-making module, achieving accurate perception of data distribution drift in non-stationary environments. The three-objective adaptive decision-making module addresses the core pain point in the background technology: "the optimization objective is not incorporated into the robustness under non-stationary perturbations, leading to sudden performance changes in hyperparameter combinations and high evaluation costs in distribution drift scenarios." As the core of the "perception-evaluation-decision" architecture of this invention, it uses the robust classification performance predictions output by the aforementioned structural perception dual-branch agent module and the calibrated risk score output by the meta-learning enhanced uncertainty evaluation module as core inputs. With "robust classification quality - prediction uncertainty - solution diversity" as the core optimization objectives, it utilizes Pareto ranking and adaptive "exploitation-exploration" mechanisms. The strategy enables adaptive risk screening under non-stationary perturbations, balancing convergence and exploration. After the selected candidate hyperparameter combinations are evaluated in practice, they will be fed back to update the model parameters of the preceding perception layer and evaluation layer, completing the closed-loop iteration of the entire optimization architecture. This invention also includes four auxiliary optimization mechanisms designed to work in conjunction with the core architecture, providing full-process support for the core closed loop. The collaborative logic of each mechanism with the core architecture is as follows: (1) Multi-objective modeling and normalization mechanism: As the foundation of the entire optimization scheme, robust classification error rate, computational complexity, and training time are set as the core optimization objectives. Normalization eliminates the difference in dimensions, providing a unified optimization benchmark for the three-layer core architecture and achieving multi-objective balanced optimization; (2) Standardized preprocessing process: As the data input basis of the scheme, through three major operations—time-frequency domain mapping, non-stationary disturbance simulation, and data augmentation—the original non-stationary time series signal is converted into a CNN-adaptive signal. Standardized data for feature extraction and robustness optimization ensures the robustness of the model in non-stationary environments from the source of data; (3) Adaptive reference solution selection mechanism: provides core support for the training of the structure-aware dual-branch surrogate module, selects reference solutions by integrating convergence and diversity dual objectives, generates high-quality supervision signals for surrogate model training, dynamically adapts to distribution drift in non-stationary environments, and improves the fitting accuracy of surrogate models; (4) Cross-network architecture adaptive adaptation mechanism: through standardized hyperparameter encoding, it is compatible with different types of deep convolutional classification networks, ensuring that the core architecture of "perception-evaluation-decision" can be adapted to different CNN architectures without modification, and expands the engineering application scope of the solution. This mechanism can be applied to various deep classification networks based on convolutional operations, without specific network architecture dependencies; Furthermore, the solution of this invention is compatible with various mainstream deep convolutional classification networks, including but not limited to lightweight MobileNet series, ResNet series, EfficientNet series and other architectures; the following uses the MobileNetV2 classification network based on parameter interconnection as a typical embodiment, and the weld defect time-frequency domain image dataset TIDIP as the experimental object. Two targeted datasets were constructed in a stable environment and a non-stationary perturbation environment respectively to verify the technical effect of this solution. This invention addresses four core pain points of existing technologies and achieves the following six beneficial effects through an integrated design of a closed-loop collaborative architecture of "perception-evaluation-decision": To address the shortcomings of existing CNN hyperparameter configuration methods for non-stationary temporal signal classification, such as low efficiency, high cost, and insufficient robustness, this invention proposes a lightweight adaptive optimization method for high-dimensional hyperparameters in CNNs for non-stationary temporal signal classification. The core optimization objectives are robust classification error rate, computational complexity, and training time. Using combinations of high-dimensional hyperparameters as individuals in the population, a MetaDCP-SAEA co-evolutionary proxy mechanism is designed. Through hierarchical collaboration of a structure-aware dual-branch proxy module, a meta-learning-enhanced uncertainty assessment module, and a three-objective adaptive decision module, the complex mapping relationship between hyperparameters and robust classification performance is fully fitted. No manual intervention is required throughout the process, ultimately achieving efficient and adaptive configuration of hyperparameters for deep convolutional neural networks in non-stationary environments with extremely low computational cost. To address the shortcomings of existing surrogate models, such as insufficient modeling capabilities and low fitting accuracy under small sample conditions, this invention proposes a structure-aware bi-branch convolutional multinomial surrogate model. Through heterogeneous parallel design of local and long-range branches, combined with multinomial nonlinear mapping, it explicitly captures the heterogeneous relationships between local tight coupling and cross-dimensional long-range dependence of hyperparameters, iteratively optimizes feature extraction efficiency, and significantly improves the fitting accuracy and prediction stability of the surrogate model under small sample conditions. This provides a reliable prediction basis for hyperparameter optimization and reduces invalid real evaluations. To address the shortcomings of existing technologies, such as lack of adaptability to non-stationary environments and inability to perceive data distribution drift, this invention proposes a meta-learning-enhanced two-stage uncertainty assessment mechanism. It generates a primary uncertainty signal through parameter perturbation integration and combines it with meta-regression calibration through multi-dimensional feature fusion to accurately quantify and predict risks and effectively distinguish between cognitive uncertainty and data uncertainty. This provides a reliable quantitative signal for risk perception decision-making in non-stationary environments and significantly improves the adaptive capability of configuration to data distribution drift in non-stationary environments. To address the limitations of existing technologies in optimizing target design and the pain points of performance abrupt changes after deployment due to the lack of robustness indicators, this invention proposes a three-objective adaptive decision framework of "robust classification quality - prediction uncertainty - solution diversity". This framework explicitly incorporates robustness under non-stationary perturbations into the optimization objective. Through Pareto ranking and a dynamic "exploitation-exploration" strategy, it achieves an adaptive balance between convergence, robustness and exploration, effectively avoiding high-risk areas under non-stationary perturbations and ensuring the robustness and effectiveness of hyperparameter configuration in actual deployment. To address the issues of outdated supervision signals and decreased fitting accuracy in surrogate model training under non-stationary environments, this invention proposes an adaptive reference solution selection mechanism. This mechanism integrates a dual-objective screening strategy of convergence and diversity, dynamically adjusting the composition of reference solutions based on the current frontier scale. This provides surrogate model training with high-quality supervision signals that adapt to changes in non-stationary environments, further improving the accuracy and convergence efficiency of hyperparameter search under non-stationary perturbations. Sixth, the technical superiority of the present invention was verified through multi-dimensional comparative experiments: In the benchmark function test, MetaDCP-SAEA outperformed the comparative algorithms in over 85% of the 28 DTLZ series comparisons, demonstrating its excellent high-dimensional multi-objective optimization capability; In the actual TIDIP weld defect dataset test, the classification model optimized by MetaDCP-SAEA achieved a classification accuracy of 95.83% under stationary conditions and maintained an accuracy of 87.04% under non-stationary perturbation conditions, with an accuracy drop of only 9.17%, far superior to most comparative algorithms. At the same time, the training time was shortened by an average of 12% compared to other surrogate-assisted evolutionary algorithms, fully verifying that the present invention has the advantages of high efficiency, high accuracy and strong robustness in non-stationary time series signal classification tasks, and has outstanding engineering application value; This invention is scientifically and rationally designed, applicable to various tasks such as prediction, identification, classification, diagnosis, and detection, and can save manpower and time costs. This invention can effectively reduce the barriers to the application of artificial intelligence technology, accelerate the engineering application process of automated machine learning, meet the national strategic requirements for promoting the healthy development and independent innovation of the intelligent manufacturing industry, and help industries achieve intelligent transformation and upgrading. It possesses broad market potential and outstanding engineering application value. Attached Figure Description

[0004] Figure 1 This is a flowchart illustrating the overall process of the present invention. Figure 2 A diagram illustrating the structure of a two-branch proxy model; Figure 3 This indicates that MetaDCP-SAEA and six surrogate-assisted evolutionary algorithms performed IGD on DTLZ1. + Compare the curve convergence plot; Figure 4 This indicates that MetaDCP-SAEA and six surrogate-assisted evolutionary algorithms performed IGD on WFG3. + Compare the curve convergence plot; Figure 5 This indicates that MetaDCP-SAEA and six surrogate-assisted evolutionary algorithms performed IGD on WFG9. + Compare the curve convergence plot; Figure 6 This indicates that MetaDCP-SAEA and six surrogate-assisted evolutionary algorithms perform IGD on DTLZ5. + Compare the curve convergence plot; Figure 7 This represents the target distribution of hyperparameter configurations for the classification network in MetaDCP-SAEA on the non-stationary TIDIP dataset. Detailed Implementation

[0005] Many specific details are set forth in the following description in order to provide a full understanding of the invention, but the invention may also be practiced in other ways different from those described herein; obviously, the embodiments in the specification are only some embodiments of the invention, and not all embodiments. The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings; A lightweight adaptive optimization method for high-dimensional hyperparameters in CNNs for non-stationary time-series signal classification, as shown in Figure 1, uses a closed-loop collaborative optimization architecture of "perception-evaluation-decision" as its core. Hyperparameter combinations are treated as individuals in an evolutionary population. The overall execution flow strictly follows the logic of "preprocessing and target definition - initialization and archive construction - iterative closed-loop optimization - termination and result output." Each step corresponds one-to-one with the flowchart in the attached figure and the three-layer core architecture of the aforementioned invention. The iterative closed-loop optimization step perfectly matches the closed-loop logic of "perception layer model training - evaluation layer risk quantification - decision layer candidate solution screening - real evaluation feedback update." The specific implementation steps are as follows: Process initialization and initial archive construction (corresponding to the starting stage in Figure 1): The process begins with sampling to generate initial sample points: using the Latin hypercube sampling method, N initial population individuals are generated in the hyperparameter search space, each individual corresponding to a set of CNN hyperparameter combinations. A real evaluation is performed on this initial population to calculate the three target values ​​for each individual: robust classification error rate, computational complexity, and training time. The hyperparameter combinations and corresponding target values ​​are stored in the initial archive DB, completing the initialization stage; Reference solution selection and quality label generation (corresponding to the "Selecting Reference Solution" module in Figure 1): After entering the iterative loop, reference solution selection is performed first; Perception-evaluation layer model training (corresponding to the "Robust Enhancement Bi-branch Convolutional Multinomial Network" and "Linear Meta-model" modules in Figure 1): This step is the core modeling stage of the closed loop, corresponding to the training agent process shown by the solid arrows in Figure 1; Decision layer intelligent selection (corresponding to the "Agent-Assisted Evolutionary Optimization" and "Centralized Selection and Diversity Exploration" modules in Figure 1): Based on the predicted values ​​of the perception layer and the risk scores of the evaluation layer, the decision layer performs three-objective adaptive selection, corresponding to... Figure 1The process involves several steps, including "genetic operations," "offspring solution generation," and "filling criteria." The real evaluation and archive update (corresponding to the "real evaluation" and "DB update" modules in Figure 1) are performed: Real evaluation is conducted on the candidate solutions selected by the decision layer, calculating their actual three-objective values. The hyperparameter combination and the real target value are merged into the archived DB, and the number of real evaluations is updated. Then, N individuals are reselected from the updated DB as the population for the next iteration, completing the closed-loop feedback. The iteration termination and result output (corresponding to the termination step in Figure 1) are repeated until the number of real evaluations (FE) reaches a preset upper limit, at which point the iteration terminates. The set of non-dominated solutions is extracted from the final archived DB, which is the optimal hyperparameter configuration set on the Pareto front. This set and its corresponding three-objective values ​​are output, completing the model configuration for robust classification of non-stationary time series signals. After completing the overall closed-loop iterative implementation process of "perception-evaluation-decision," this invention relies on the MetaDCP-SAEA co-evolutionary proxy mechanism to achieve lightweight adaptive optimization of high-dimensional hyperparameters in CNNs for non-stationary temporal signal classification. This mechanism, through hierarchical collaboration and functional complementarity of each core module, accurately fits the mapping relationship between hyperparameters and robust classification performance, achieving low-cost and efficient configuration of hyperparameters in non-stationary environments. The following will, in conjunction with the core design principles and actual implementation details of this invention, elaborate on the specific implementation methods for each of the six beneficial effects of this invention, clarifying the implementation steps, parameter settings, and collaborative logic of each module. First, the lightweight adaptive optimization method for high-dimensional hyperparameters in CNNs for non-stationary temporal signal classification proposed in this invention uses a general deep convolutional classifier network as the core classification carrier. Throughout the process, it minimizes three core objectives: robust classification error rate, computational complexity, and training time. This scheme is adaptable to deep convolutional classifier networks of any structure. For other types of deep convolutional classifier networks such as ResNet, EfficientNet, and ShuffleNet, this scheme does not require modification of the core "perception-evaluation-decision" optimization mechanism. It only needs to integrate the structural parameters and training parameters of the target network into a standardized decision vector through standardized hyperparameter encoding, and dynamically adjust the parameter types and value ranges in the hyperparameter search space to achieve adaptive hyperparameter optimization of the corresponding network, thus completing the robust classification configuration for non-stationary temporal signals. The following section uses the MobileNetV2 classification network based on parameter interconnection as a typical example to describe the specific implementation steps of the scheme in detail. The specific implementation steps are as follows: (1) Determining the optimization objective and quantization modeling: The hyperparameter optimization objective for robust classification of non-stationary sequence signals is modeled as a three-objective minimization problem, and the objective vector is defined as = And normalize each target to eliminate dimensional differences; among which To improve robust classification error rate, To calculate the computational complexity, Training time; (2) Definition and initialization of hyperparameter population: sort out the high-dimensional hyperparameters to be optimized in the target depth convolutional classification network. General hyperparameters include but are not limited to structural parameters and training parameters such as the number of convolution channels, convolution kernel size, dilation rate, learning rate, optimizer type, batch size, training epochs, and dropout rate. For different network architectures, the type and value range of the hyperparameters to be optimized can be adaptively adjusted. Each set of hyperparameter combinations is defined as a population individual in the MetaDCP-SAEA evolutionary algorithm. The initial hyperparameter population is generated by Latin hypercube sampling. The population size is adaptively set according to the hyperparameter dimension to ensure the uniformity and diversity of the initial population and lay the foundation for subsequent evolutionary optimization. (3) The MetaDCP-SAEA co-evolutionary proxy mechanism is constructed by building a core mechanism consisting of a structure-aware dual-branch proxy module, a meta-learning-enhanced uncertainty assessment module, and a three-objective adaptive decision-making module. The hierarchical collaborative logic of the three modules is defined: the structure-aware dual-branch proxy module is used as the perception layer to complete the efficient fitting of the complex nonlinear mapping between hyperparameters and robust classification performance, providing an accurate performance prediction basis for subsequent evaluation and decision-making; the meta-learning-enhanced uncertainty assessment module is used as the evaluation layer to dynamically quantify the uncertainty risk in the prediction process and achieve accurate perception of the data distribution drift in non-stationary environments; the three-objective adaptive decision-making module is used as the decision-making layer to screen candidate hyperparameter combinations that take into account performance, robustness and exploration, forming a hierarchical collaborative operation mode of "perception-evaluation-decision". (4) The implementation of lightweight adaptive optimization and low cost control replace the large amount of real training and evaluation of deep convolutional classification networks with the fitting of the surrogate model. Only the optimal candidate hyperparameter combination selected by the decision layer is evaluated, reducing the number of high-cost real training times. At the same time, each module adopts a lightweight modeling design (such as the linear computational complexity of the dual-branch surrogate model and the lightweight linear regression architecture of the meta-learning evaluation module), reducing the training and computational costs of the module itself. Moreover, the three modules achieve adaptive parameter updates through closed-loop feedback, without the need for manual adjustment of hyperparameter search strategies and model parameters, further reducing manual costs and trial and error costs. (5) Output and configuration of the optimal hyperparameter set: Through multiple rounds of iterative optimization using the MetaDCP-SAEA co-evolutionary proxy mechanism, the candidate hyperparameter combinations in each round are sorted by three objectives. Finally, the optimal set of non-dominated hyperparameters and their corresponding objective values ​​on the Pareto front are extracted. This set is then directly applied to the hyperparameter configuration of the deep convolutional neural network to complete the construction of the non-stationary time-series signal classification model and achieve low-cost and efficient configuration of hyperparameters of deep convolutional neural networks in non-stationary environments. Second, the core objective of the structure-aware bi-branch convolutional multinomial surrogate model proposed in this invention is to explicitly model the local tight coupling and cross-dimensional long-range dependencies between hyperparameters through a heterogeneous parallel architecture, and to enhance the fitting ability under small sample conditions by combining multinomial nonlinear mapping, thereby providing accurate predictive support for subsequent risk assessment and decision-making. The specific implementation steps are detailed in Figure 2 below: (1) As shown in Figure 2, the surrogate model adopts a heterogeneous parallel architecture of “input layer - dual-branch feature extraction layer - multinomial mapping layer - fusion output layer”. The input is a high-dimensional hyperparameter decision vector (the dimension is consistent with the total number of hyperparameters to be optimized), and the output is a robust classification performance prediction value. Among them, the dual-branch feature extraction layer includes local branches and long-range branches. The two operate independently in parallel and share the input layer and fusion output layer. Differentiated receptive field design captures different dependencies between hyperparameters to ensure the comprehensiveness and specificity of feature extraction. (2) Local branches focus on capturing the local tight coupling relationship between hyperparameters (such as the relationship between the number of channels and the kernel size within the same convolutional layer). The specific implementation is as follows: Kernel configuration: A 1×1 convolutional kernel is used. This size design can act independently on each hyperparameter decision variable, only realizing channel expansion without introducing cross-dimensional information interaction, and accurately focusing on local feature extraction; Feature extraction process: The input hyperparameter decision vector is reshaped into a 1× D A one-dimensional tensor of ×1 ( D (where the hyperparameter dimension is 1), after inputting a local branch, a linear transformation is performed through a 1×1 convolution kernel, and the number of output channels is set to 1. This yields a local feature map C1; dimensionality adjustment: the local feature map C1 is flattened into a one-dimensional feature vector F1 with dimension 1. This prepares for subsequent nonlinear mapping; (3) Implementation Design of Long-Range Branches Long-range branches are used to capture cross-dimensional long-range collaborative relationships between hyperparameters (such as the correlation between the initial learning rate and the dropout rate of deep networks). The specific implementation is as follows: Convolutional kernel and padding configuration: A 9×1 convolutional kernel is used, combined with a "same" padding strategy to ensure that the output dimension is consistent with the input dimension. Information from adjacent hyperparameter dimensions is aggregated through a sliding window to achieve long-range dependency capture; Feature extraction process: The same input tensor is input into the long-range branch, and a convolution operation is performed through a 9×1 convolutional kernel. The number of output channels is consistent with that of the local branch, resulting in a long-range feature map C2; Dimension adjustment: The long-range feature map C2 is flattened into a one-dimensional feature vector F2, with the same dimension. This ensures that the feature vectors of the two branches have the same dimension; (4) Implementation of polynomial nonlinear mapping: The eigenvectors F1 and F2 of the two branches are respectively connected to independent 3rd order polynomial networks. The nonlinear mapping capability is enhanced by power expansion. The specific implementation is as follows: Polynomial mapping calculation: For the eigenvector F k (k=1 corresponds to the local branch, k=2 corresponds to the long-range branch) Perform power expansion from 0 to 3, and generate feature vectors and learnable coefficient vectors for each order. Multiplying (i=0,1,2,3) and summing the results yields the branch output Z. k The calculation formula is: in, This is a coefficient vector that can be updated through training after random initialization. Indicates to Each element is taken independently i Power; Output characteristic: Z k As a scalar value, it directly reflects the predictive contribution of the corresponding branch to the performance of the hyperparameter combination; (5) Feature fusion and output implementation adopts an equal-weighted averaging strategy to fuse the outputs of the two branches, avoiding the risk of small-sample overfitting caused by the introduction of additional trainable parameters. Specifically, the local branch output Z1 and the long-range branch output Z2 are added with a weight of 1:1 to obtain the final robust classification performance prediction value. ,Right now: ; (6) Model training configuration and optimization; training data source: using hyperparameter combinations and corresponding quality labels that have been evaluated in the archive (generated by the adaptive reference solution selection mechanism); the training set and validation set are divided in an 8:2 ratio; objective function and optimizer: using mean squared error (MSE) as the objective function to calculate the predicted value. With genuine quality label y The deviation is expressed by the formula: ; in, θFor all learnable parameters of the model, The training sample size is set to a certain value. The Adam optimizer is used for parameter updates, with an initial learning rate of 0.001, which decays to 0.9 times the previous learning rate every 50 rounds. An early stopping mechanism is introduced, using a validation set comprising 20% ​​of the training data to construct a loss monitoring window. If the loss on the validation set decreases by less than 10% for 10 consecutive rounds, the loss is stopped. −4 If the current optimal parameters are not found, then training will be terminated and the current optimal parameters will be saved. Third, the meta-learning-enhanced two-stage uncertainty assessment mechanism of the core architecture evaluation layer of this invention addresses the problems of existing methods in the background art being unable to perceive the distribution drift of non-stationary environments and ignoring prediction uncertainty, resulting in low optimization efficiency. The core design objective is to accurately quantify prediction uncertainty and distinguish its sources (cognitive uncertainty and data uncertainty) through a two-stage collaborative operation of "parameter perturbation integration + meta-regression calibration". This mechanism uses the output of the aforementioned structure-aware dual-branch surrogate model as the sole input basis. The calibrated risk score of its output will directly serve as one of the core optimization objectives of the three-objective adaptive decision-making module, providing a reliable signal for risk perception decision-making in non-stationary environments and improving the adaptability of configuration to data distribution drift. The specific implementation steps are detailed below: The mechanism requires pre-implementation preparation and input definition. Its input consists of a structure-aware dual-branch surrogate model (already trained and converged) and a set of candidate hyperparameter solutions to be evaluated. The output is a calibrated prediction risk score (range [0,1], where a higher score indicates a greater prediction risk). Before implementation, the dual-branch surrogate model must be trained and converged (refer to the training configuration in point two) to ensure the model possesses basic performance prediction capabilities. Simultaneously, the feature dimensions of the candidate hyperparameter solutions to be evaluated must be determined to lay the foundation for subsequent feature extraction and processing. Phase 1: Parameter Perturbation Integration to Generate Primary Uncertainty Signals. This phase generates multiple perturbation model variants by applying controllable parameter perturbations to the converged bi-branch surrogate model. The primary uncertainty is quantified using the statistical volatility of the predicted output. Specifically, the perturbation model is generated by obtaining the optimal parameter set after the bi-branch surrogate model has converged. Five levels of linearly increasing Gaussian noise are applied to the parameter set to generate five perturbation models. The parameter update formula is as follows: ; The noise intensity increases linearly: , The total number of perturbation models is set so that the first model retains its original parameters, and the perturbation of subsequent models gradually increases to a maximum value of 0.01; multi-model prediction: each candidate solution of the hyperparameters to be evaluated is... x Input five perturbation models respectively to obtain five predicted values; Primary uncertainty calculation: calculate the mean of the five predicted values. with standard deviation ,in As a primary uncertainty signal, it directly reflects the degree of fluctuation in the model's prediction of the candidate solution. The greater the fluctuation, the worse the prediction stability and the higher the potential risk. Phase Two: Meta-Regression Calibration for Refined Uncertainty Modeling. This phase constructs a meta-regression model by fusing multi-dimensional complementary features to calibrate the primary uncertainty signal, accurately distinguish the sources of uncertainty, and output the final risk score. The specific implementation is as follows: Feature Extraction and Construction: For each candidate solution of the hyperparameter to be evaluated... x Five categories of features with a total of 15 dimensions were extracted to construct a comprehensive feature vector. z Basic features (3 dimensions): hyperparameter decision variable vector, ensemble mean of the first stage. Integrated standard deviation Statistical features (5 dimensions): mean, variance, skewness, kurtosis, and quartiles of the hyperparameter combination; Interaction features (1 dimension): the product of the ensemble mean and ensemble standard deviation. Density features (3D): 3-nearest neighbor sample density, 5-nearest neighbor sample density, 7-nearest neighbor sample density; Distance features (3D): Euclidean distance to the centroid of the training sample, Euclidean distance to the most similar training sample; Feature preprocessing: processing the 15-dimensional feature vector. z Z-score standardization was performed to eliminate dimensional differences between different features. The meta-regression model was constructed using a single-layer linear regression architecture, with the output layer incorporating a sigmoid activation function to limit the output risk score to the [0,1] interval. Model parameters included linear weights. (Initialized to random values ​​following an N(0,0.01) distribution) and bias term (Initialized to 0), the model output formula is: Supervision signal generation: Supervision signals are constructed based on hyperparameter samples that have undergone real evaluation. For the first... i Calculate quality deviation from historical samples. With integrated fluctuations After normalizing the two signals, they are linearly fused to generate a monitoring signal: ; Model training optimization: Using mean squared error (MSE) as the base loss function, contrastive learning loss and L2 regularization are introduced. The total loss function is: Among them, the contrastive learning loss is Regularization term For L2 weight decay term, regularization coefficient Used to avoid overfitting of the meta-regression model; by minimizing the total loss through the Adam optimizer, the parameters of the meta-regression model are updated and iterated, thereby improving the accuracy of risk identification; The two-stage mechanism of overall coordination and application of uncertainty assessment forms a collaborative logic of "initial perception - fine calibration": parameter perturbation integration captures and predicts fluctuations, providing a basis for preliminary risk screening in non-stationary environments; meta-regression calibration integrates multi-dimensional features to achieve accurate quantification and source differentiation of uncertainty. In practical applications, the output risk score of this mechanism directly serves as one of the core inputs of the three-objective adaptive decision-making framework. When data distribution drift leads to an increase in the risk score, the decision layer automatically adjusts the search strategy to avoid high-risk hyperparameter combinations and ensure the robustness of hyperparameter configuration. Fourth, the core architecture of this invention employs a three-objective adaptive decision-making framework for "robust classification quality, prediction uncertainty, and solution diversity" in its decision layer. Addressing the issues in the background technology where robustness is not incorporated into the optimization objectives, high-dimensional space search is prone to getting trapped in local optima, and real evaluation costs are too high, the core design aims to achieve an adaptive balance of convergence, robustness, and exploratory power in non-stationary perturbation environments by explicitly modeling three complementary objectives and combining Pareto ranking with a dynamic "exploitation-exploration" strategy. This framework uses the robust classification quality prediction value output by the structure-aware dual-branch surrogate module and the risk score output by the meta-learning enhanced uncertainty assessment module as core inputs. After real evaluation, the selected candidate hyperparameter combinations are fed back to update the training data and model parameters of the two preceding modules, completing the closed loop of the entire optimization process. This effectively avoids high-risk areas, reduces invalid real evaluations, and ensures the reliability and efficiency of hyperparameter configuration. Its specific implementation revolves around the collaborative logic of objective modeling, ranking and selection, and strategy adjustment. The quantitative modeling of the three objectives forms the foundation of the decision-making framework. Each candidate solution for the hyperparameters to be evaluated is mapped to a three-dimensional objective vector. The objectives are both independent and complementary in terms of constraints. Among them, the "robust classification quality" objective is directly derived from the prediction output of the structure-aware bi-branch surrogate model, with the reciprocal of the robust classification error rate as the core quantitative indicator. The higher the value, the better the classification performance of the candidate solution in non-stationary environments. The "prediction uncertainty" objective is taken from the calibrated risk score output by the meta-learning enhanced bi-stage uncertainty assessment mechanism. The lower the score, the more reliable the prediction result and the stronger the robustness in the face of data distribution drift. The "solution diversity" objective is quantified by calculating the distribution characteristics of candidate solutions in the hyperparameter search space. It comprehensively considers factors such as the Euclidean distance between the candidate solution and the selected high-quality solution, and the crowding degree of the region. The greater the distance and the lower the crowding degree, the higher the diversity index, thereby avoiding the search process from getting trapped in local optima and ensuring a comprehensive exploration of the hyperparameter space. Pareto sorting is a core step in multi-objective optimization screening. In practice, it first collects a set of all candidate solutions with hyperparameters to be decided, calculates the three-dimensional objective vector for each candidate solution, and then uses a non-dominated sorting algorithm to hierarchically divide the candidate solutions. This algorithm divides the candidate solutions into multiple Pareto front levels by judging the dominance relationship between them. The first front level contains the optimal non-dominated solutions that are not dominated by any other candidate solutions, and the dominance relationship of subsequent levels decreases sequentially. Simultaneously, the proportion of the first front level in the entire candidate solution set is calculated. ( F1 is the first frontier set of individuals. C (This is the set of candidate solutions), and this proportion directly reflects the convergence status of the current search process, providing a key basis for adjusting the subsequent "exploitation-exploration" strategy; The adaptive adjustment of dynamic "exploitation-exploration" strategies is key to coping with non-stationary environments, and its core is based on historical data. Median μ of the sequence t and absolute median difference σ t Construct a Sigmoid decision function to dynamically adjust the execution probabilities of the two strategies. The decision function formula is as follows: ; (in (For smoothing terms). When A higher value indicates that the current search is approaching a high-quality region, and the algorithm uses probability... The "utilization" strategy is implemented by calculating the crowding distance of each candidate solution in the first frontier level to filter individuals in sparse regions, focusing on in-depth mining of high-quality solutions; when A lower value indicates that the current search area has high uncertainty or no high-quality solution has been found. The algorithm then uses probability... The "exploration" strategy is implemented, and the selection quota is allocated in descending order of the frontier level, with the first frontier accounting for 50% of the quota, and the remaining levels allocated in order of rank, to ensure that the search scope covers more potential high-quality areas and avoid search stagnation caused by data distribution drift. To further adapt to the dynamic changes of non-stationary environments, the decision framework also incorporates an adaptive adjustment mechanism for the number of selection positions. When the size of the first frontier level exceeds 30% of the candidate solution set, the number of selection positions is increased by 20% to retain more high-quality solutions while maintaining diversity. If no new Pareto front is generated for three consecutive iterations, indicating that the search may be trapped in a local optimum or facing severe distribution drift, an additional 15% of selection positions are added, and the weight of the "exploration" strategy is strengthened to guide the algorithm to expand into the uncovered hyperparameter space. Through this series of dynamic adjustments, the decision framework can flexibly balance the needs of "focusing on high-quality solutions" and "exploring new regions" under non-stationary perturbations, effectively avoiding high-risk regions and ensuring that the hyperparameter configuration achieves a dynamic balance between convergence, robustness, and exploration, ultimately outputting an optimal non-dominated hyperparameter combination that takes into account multiple objectives. Fifth, the adaptive reference solution selection mechanism proposed in this invention is based on the dual-objective screening logic of convergence and diversity. It dynamically adjusts the composition of reference solutions according to the frontier scale of the current archived solution set, providing high-quality supervision signals that are adapted to non-stationary environments for the structure-aware dual-branch surrogate model, thereby improving the accuracy and efficiency of hyperparameter search. Its specific implementation revolves around the closed-loop process of "frontier partitioning - dynamic screening - label generation". The prerequisite for implementing the mechanism is input definition and initialization. This requires clearly defining the complete set of hyperparameter solutions and their corresponding target values ​​(robust classification error rate, computational complexity, training time) in the current archive database, while also pre-setting the number of reference solutions. Q = 7. First, perform a non-dominated sort on the archived solution set, and then divide the frontier levels according to the dominance relationship of the objective values ​​of each solution. ,in F Level 1 represents the first frontier level, which contains the optimal non-dominated solution that is not dominated by any other solution. The solutions in subsequent levels are successively dominated by solutions in better levels. This division directly reflects the convergence state of the current search and provides a basis for selecting reference solutions. The dynamic selection of reference solutions is the core of the mechanism, and it needs to be based on the scale of the first frontier level. F 1 | With preset quantity Q The relationship between the reference solutions and the adaptive adjustment of the selection strategy ensures that the reference solutions possess both strong convergence and sufficient diversity. F 1|= Q When, select directly F All individuals in step 1 are used as reference solutions. At this point, the optimal solution of the first frontier can completely cover the Pareto optimal region, and no additional supplementation is needed; when | F 1|< Q At that time, first F All individuals of 1 are included in the candidate set of the reference solution, and the remaining vacancies are filled from the subsequent frontier levels ( F 2,F 3, ...) Replacements are made according to quality score, quality score formula: ; in k i For the target vector of the solution, this score integrates convergence (the reciprocal of the target vector's magnitude) and diversity (the minimum Euclidean distance to other solutions), ensuring that the supplementary solution is both close to the optimal frontier and fills the distribution gap; when | F 1|> Q At that time, a two-stage screening process of "prioritizing high-quality products + supplementing with diversity" was adopted. In the first stage, products with a quality score of ⌈0.8× were selected in descending order. Q ⌉ Individuals (focusing on convergence), the second stage selects the remaining individuals from F1 using the maximum-minimum distance strategy. ,in S For the selected set, C for F 1. Remaining set) to supplement the quantity Q To avoid the reference solution being overly concentrated in local areas; Once the reference solutions are determined, quality labels needed for training the surrogate model must be generated to provide precise guidance for supervised learning. Specifically, the minimum distance from each hyperparameter solution in the archive to all reference solutions is calculated. (in For reference j The target vector, based on which the quality score is defined. The closer the distance, the better the quality of the solution and the higher the score. Then, all solutions are sorted according to their quality scores, and the top 80% of the solutions are labeled as positive samples and the bottom 20% as negative samples, generating a binary quality label. This label highlights the supervision weight of high-quality solutions and avoids model overfitting through a reasonable ratio of positive and negative samples, ensuring that the surrogate model can accurately learn the mapping relationship between hyperparameters and robust classification performance. This mechanism dynamically adjusts the composition of the reference solution, naturally adapting to data distribution drift in non-stationary environments: when environmental changes cause the supervision signal corresponding to the original reference solution to become outdated, a new round of screening will re-divide the frontier level and supplement the reference set with high-quality solutions that adapt to the new distribution, avoiding the surrogate model training being misled by outdated signals; at the same time, by integrating the screening logic of convergence and diversity, the reference solution can both anchor the optimal direction of hyperparameter search and cover potential high-quality regions, effectively improving the accuracy and efficiency of hyperparameter search under non-stationary perturbations, and providing key support for the efficient operation of the entire MetaDCP-SAEA co-evolutionary surrogate mechanism; Sixth, this invention systematically verifies the effectiveness and superiority of the MetaDCP-SAEA algorithm in the hyperparameter configuration task of non-stationary time series signals through multi-dimensional progressive experimental comparison. The experimental design corresponds one-to-one with the aforementioned core technical solutions, following the progressive logic of "basic optimization capability verification - core module effectiveness verification - actual engineering scenario performance verification". Among them, the benchmark function verification is used to test the algorithm's general high-dimensional multi-objective optimization capability, the variant effectiveness verification is used to verify the necessity of the design of each core module of this invention, and the actual dataset verification is used to test the engineering application effect of the algorithm in the industrial non-stationary time series signal classification scenario. The experimental steps cover the selection of comparison algorithms, dataset construction, evaluation index setting, experimental process execution and result analysis. The core experimental design and results are described in detail below. Before the experiment, the hardware and software environment configuration and experimental parameters must be completed to ensure the reproducibility of the experiment. The hardware environment uses an Intel(R) Core(TM) i5-124900F CPU and an NVIDIA GeForce RTX 3060 Ti graphics card. The software environment is built on MATLAB R2024a, relying on the PlatEMO platform to achieve unified operation of all comparison algorithms. Regarding parameter settings, the population size is set to 50, and the maximum number of true evaluations is FE. max Based on hyperparameter dimensions D Set to (11) D -1)+120, Number of reference solutions Q =7, the upper limit of the number of iterations within the proxy model, g max =100, evolutionary algorithm parameters such as crossover probability and mutation probability are configured according to mainstream standards (crossover probability...). proC =1, Cross-distribution index disC =15, Probability of Mutation proM =1 / D Variation distribution index disM =5), all algorithms are run independently 30 times to eliminate the influence of random errors; The comparison objects were selected into two categories. The first category consisted of six representative surrogate-assisted evolutionary algorithms, including DirHV-EI, MO-L2SMEA, EDN-ARMOEA, REMO, ADSAPSO, and DCSCSAEA, covering various surrogate modeling paradigms and different model management mechanisms such as Gaussian processes, Dropout neural networks, and radial basis functions, to ensure the comprehensiveness of the comparison. The second category consisted of four variants of MetaDCP-SAEA: (1) MetaDCP-SAEA-Q relied solely on quality scores; (2) MetaDCP-SAEA-R used completely random selection; (3) MetaDCP-SAEA-RW used roulette wheel selection based on the quality probability distribution; and (4) MetaDCP-SAEA-D considered only the diversity of solution set distributions. Evaluation metrics included the benchmark function test metric (Inverse Generation Distance IGD). + The comparison between IGD (increased volume HV) and actual task metrics (classification accuracy, accuracy reduction, training time) is as follows. + The value is used to measure the convergence and uniformity of the solution set (the smaller the value, the better the performance). HV is used to evaluate the coverage quality of the solution set (the larger the value, the better). The accuracy drop is used to quantify the performance robustness in non-stationary environments (the smaller the drop, the stronger the robustness). The benchmark function validation experiments selected the DTLZ series, classic test functions covering various challenging scenarios such as high-dimensionality, multi-modal, nonlinear, and complex frontier structures, to verify the algorithm's general optimization capabilities. The experimental procedure was as follows: MetaDCP-SAEA and six comparative algorithms were run on each test function, and the IGD performance of each run was recorded. + The metrics, including the mean and standard deviation of 30 runs, are shown in Table 1. Convergence curves (Figures 3-6) were also plotted to visually compare the convergence speed and stability of each algorithm. The results show that MetaDCP-SAEA outperformed the compared algorithms in over 85% of the 28 DTLZ series comparisons, exhibiting a steeper and more stable convergence curve with no obvious plateau, thus validating its fundamental capabilities in high-dimensional space modeling and robust optimization. Table 1 shows the IGD+ values ​​obtained by EDN-ARMOEA, ADSAPSO, REMO, MO-L2SMEA, DirHV-EI, DCSCSAEA and MetaDCP-SAEA on the three-objective DTLZ test suite with decision variables of 20 / 40 / 60 / 100. The best results are highlighted. Problem D EDN-ARMOEA ADSAPSO REMO MO-L2SMEA DirHV-EI DCSCSAEA MetaDCP-SAEA DTLZ1 20 3.3964e+2(3.25e+1) - 2.8100e+2(5.62e+1) - 1.9444e+2(4.24e+1) - 3.2265e+2(3.83e+1) - 2.3439e+2(5.43e+1) - 2.1917e+2(5.23e+1) - 1.5046e+2(2.54e+1) 40 8.6402e+2(4.62e+1) - 5.5760e+2(8.32e+1) - 4.3639e+2(9.35e+1) - 8.3575e+2(5.21e+1) - 8.3848e+2(1.02e+2) - 4.7391e+2(5.95e+1) - 3.2475e+2(3.54e+1) 60 1.4210e+3(5.77e+1) - 8.1914e+2(1.44e+2) - 6.4860e+2(9.97e+1) - 1.3844e+3(6.27e+1) - 1.2281e+3(2.92e+2) - 7.1689e+2(9.89e+1) - 5.5657e+2(6.83e+1) 100 2.5686e+3(6.74e+1) - 1.3691e+3(2.05e+2) - 1.0659e+3(1.26e+2) - 2.3141e+3(4.99e+2) - 2.6390e+3(4.56e+1) - 1.0660e+3(1.02e+2) - 9.4418e+2(1.30e+2) DTLZ2 20 8.2642e-1(6.97e-2) - 4.6168e-1(1.27e-1) - 3.6281e-1(8.51e-2) = 5.4106e-1(5.38e-2) - 4.7324e-1(4.69e-2) - 4.4427e-1(8.52e-2) - 3.6309e-1(9.74e-2) 40 2.0700e+0(7.30e-2) - 9.3230e-1(1.89e-1) - 6.6512e-1(1.37e-1) = 7.7349e-1(1.03e-1) - 7.1027e-1(9.81e-2) = 8.3046e-1(1.43e-1) - 6.9046e-1(1.05e-1) 60 3.3565e+0(1.03e-1) - 1.6685e+0(5.56e-1) - 1.1160e+0(1.64e-1) - 8.8865e-1(1.15e-1) + 1.4924e+0(3.32e-1) - 1.2386e+0(2.59e-1) - 9.9767e-1(1.47e-1) 100 6.0711e+0(2.43e-1) - 3.1642e+0(5.90e-1) - 2.0492e+0(3.88e-1) - 1.0611e+0(1.45e-1) + 6.1349e+0(2.07e-1) - 2.3401e+0(5.13e-1) - 1.8068e+0(2.38e-1) DTLZ3 20 1.0612e+3(1.16e+2) - 6.9149e+2(1.73e+2) - 5.6266e+2(1.04e+2) - 1.0221e+3(9.83e+1) - 5.9944e+2(9.13e+1) - 6.2722e+2(9.94e+1) - 4.5842e+2(6.20e+1) 40 2.7549e+3(1.60e+2) - 1.5289e+3(2.57e+2) - 1.2777e+3(2.00e+2) - 2.5173e+3(3.33e+2) - 1.8397e+3(3.57e+2) - 1.3758e+3(2.00e+2) - 1.0632e+3(1.21e+2) 60 4.5336e+3(2.42e+2) - 2.1768e+3(4.04e+2) - 2.0591e+3(2.36e+2) - 4.2508e+3(2.09e+2) - 2.5207e+3(1.11e+3) = 2.1224e+3(2.41e+2) - 1.7661e+3(1.64e+2) 100 8.2280e+3(2.31e+2) - 4.1291e+3(8.36e+2) - 3.3676e+3(3.18e+2) - 7.5313e+3(1.19e+3) - 8.3335e+3(1.98e+2) - 3.5810e+3(3.66e+2) - 3.0465e+3(3.23e+2) DTLZ4 20 7.2487e-1(1.43e-1) - 6.9784e-1(1.45e-1) - 3.9758e-1(1.00e-1) = 6.2037e-1(7.94e-2) - 6.0384e-1(8.68e-2) - 4.3801e-1(9.76e-2) = 4.0058e-1(9.44e-2) 40 2.1171e+0(1.76e-1) - 1.4744e+0(4.33e-1) - 7.1599e-1(1.46e-1) = 9.2372e-1(1.07e-1) - 7.7627e-1(9.52e-2) - 8.8285e-1(2.20e-1) - 7.1504e-1(1.28e-1) 60 3.5861e+0(1.46e-1) - 2.0984e+0(5.91e-1) - 1.1479e+0(1.90e-1) = 9.7540e-1(1.54e-1) + 8.7238e-1(8.09e-2) + 1.2529e+0(3.08e-1) = 1.1167e+0(1.80e-1) 100 6.3422e+0(2.00e-1) - 3.2924e+0(9.35e-1) - 1.9937e+0(2.73e-1) = 1.1529e+0(1.55e-1) + 6.4606e+0(1.78e-1) - 2.2577e+0(4.49e-1) = 2.0924e+0(3.27e-1) DTLZ5 20 7.1426e-1(7.60e-2) - 3.9351e-1(1.27e-1) - 3.7662e-1(8.61e-2) - 4.9096e-1(7.32e-2) - 4.0470e-1(7.15e-2) - 4.2032e-1(7.22e-2) - 2.7712e-1(6.99e-2) 40 1.9988e+0(1.13e-1) - 9.9403e-1(2.47e-1) - 6.6980e-1(1.11e-1) - 6.8818e-1(1.13e-1) - 7.5362e-1(1.66e-1) - 8.7720e-1(1.81e-1) - 5.9109e-1(7.80e-2) 60 3.3316e+0(1.14e-1) - 1.7197e+0(3.86e-1) - 1.0734e+0(1.63e-1) = 8.3643e-1(1.13e-1) + 1.3630e+0(2.50e-1) - 1.2191e+0(2.24e-1) - 1.0120e+0(1.49e-1) 100 6.0611e+0(1.76e-1) - 3.1779e+0(7.00e-1) - 1.9878e+0(2.89e-1) - 9.7421e-1(1.36e-1) + 6.1222e+0(1.98e-1) - 2.2903e+0(4.35e-1) - 1.7636e+0(2.50e-1) DTLZ6 20 1.4468e+1(3.95e-1) - 1.0188e+1(1.43e+0) = 1.2095e+1(1.18e+0) - 1.5108e+1(4.45e-1) - 1.1611e+0(6.64e-1) + 1.2256e+1(7.13e-1) - 1.0615e+1(9.84e-1) 40 3.2028e+1(4.45e-1) - 2.1993e+1(2.75e+0) + 2.7660e+1(1.09e+0) - 3.2303e+1(7.56e-1) - 1.0518e+1(3.20e+0) + 2.7283e+1(1.26e+0) - 2.4441e+1(1.13e+0) 60 4.9796e+1(6.24e-1) - 3.0606e+1(4.25e+0) + 4.2916e+1(1.26e+0) - 4.9529e+1(1.07e+0) - 2.2433e+1(9.23e+0) + 4.2193e+1(1.50e+0) - 3.9015e+1(1.65e+0) 100 8.5436e+1(7.13e-1) - 4.8971e+1(7.32e+0) + 7.5733e+1(2.83e+0) - 8.4170e+1(2.16e+0) - 8.6341e+1(2.41e-1) - 7.3306e+1(1.97e+0) - 6.8431e+1(2.11e+0) DTLZ7 20 2.8587e+0(6.09e-1) - 5.8726e+0(1.09e+0) - 1.5108e+0(5.11e-1) + 8.0095e+0(7.93e-1) - 1.8250e-1(6.41e-2) + 3.3173e+0(7.56e-1) - 2.2628e+0(8.23e-1) 40 4.0270e+0(6.50e-1) - 6.6498e+0(1.08e+0) - 1.9924e+0(6.47e-1) + 9.0414e+0(6.28e-1) - 7.8041e-1(3.21e-1) + 2.8743e+0(8.48e-1) - 2.3394e+0(6.00e-1) 60 4.8052e+0(6.27e-1) - 7.1714e+0(1.04e+0) - 2.7333e+0(5.56e-1) + 9.6736e+0(4.65e-1) - 2.8664e+0(1.88e+0) + 3.5133e+0(9.62e-1) = 3.2418e+0(6.97e-1) 100 6.1393e+0(6.02e-1) - 7.3034e+0(1.15e+0) - 3.4360e+0(4.77e-1) = 9.9075e+0(4.16e-1) - 9.9561e+0(2.16e-1) - 3.7497e+0(4.86e-1) - 3.4000e+0(6.94e-1) + / - / = 0 / 28 / 0 3 / 24 / 1 3 / 17 / 8 6 / 22 / 0 7 / 19 / 2 0 / 24 / 4 In experiments with MetaDCP-SAEA and its four variants, MetaDCP-SAEA achieved the best IGD⁺ and HV values ​​on most test problems, as shown in Tables 2 and 3. In contrast, strategies relying solely on quality or diversity performed poorly in terms of convergence and distribution, respectively, while random and probabilistic selection methods exhibited the worst overall stability. These results indicate that a lack of structure-guided reference solution selection makes it difficult to support robust training of surrogate models. Table 2 shows the IGD+ values ​​of MetaDCP-SAEA and its four variants on the three-objective DTLZ test suite with 3 objectives and 40 decision variables. The best results are highlighted. Problem MetaDCP-SAEA-Q MetaDCP-SAEA-R MetaDCP-SAEA-RW MetaDCP-SAEA-D MetaDCP-SAEA DTLZ1 2.9300e+2 (3.47e+1) + 7.9362e+2 (5.77e+1) - 7.2830e+2 (5.37e+1) - 7.3516e+2 (8.86e+1) - 3.2475e+2 (3.54e+1) DTLZ2 7.4603e-1 (1.25e-1) - 1.2664e+0 (2.05e-1) - 1.1176e+0 (2.01e-1) - 1.5499e+0 (2.05e-1) - 6.9046e-1 (1.05e-1) DTLZ3 1.0395e+3 (1.94e+2) = 2.1071e+3 (1.97e+2) - 1.9920e+3 (2.07e+2) - 1.9751e+3 (2.59e+2) - 1.0632e+3 (1.21e+2) DTLZ4 8.2028e-1 (1.57e-1) - 8.9050e-1 (2.17e-1) - 8.7539e-1 (1.51e-1) - 1.2197e+0 (2.57e-1) - 7.1504e-1 (1.28e-1) DTLZ5 6.0461e-1 (1.32e-1) = 1.2513e+0 (1.84e-1) - 1.1572e+0 (1.64e-1) - 1.5406e+0 (2.23e-1) - 5.9109e-1 (7.80e-2) DTLZ6 2.3961e+1 (1.15e+0) = 3.1173e+1 (5.11e-1) - 3.1088e+1 (6.57e-1) - 3.1593e+1 (8.04e-1) - 2.4441e+1 (1.13e+0) DTLZ7 3.5669e+0 (1.33e+0) - 3.1675e+0 (6.69e-1) - 3.0725e+0 (7.31e-1) - 2.7941e+0 (7.14e-1) - 2.3394e+0 (6.00e-1) + / - / = 1 / 3 / 3 0 / 7 / 0 0 / 7 / 0 0 / 7 / 0 Table 3 shows the HV values ​​of MetaDCP-SAEA and its four variants on a three-objective WFG test suite with 3 objectives and 40 decision variables. The best results are highlighted. Problem MetaDCP-SAEA-Q MetaDCP-SAEA-R MetaDCP-SAEA-RW MetaDCP-SAEA-D MetaDCP-SAEA WFG1 3.6169e-2 (6.19e-2) - 1.5915e-2 (2.32e-2) - 1.9128e-2 (2.52e-2) - 2.2169e-1 (4.06e-2) = 1.9944e-1 (5.91e-2) WFG2 5.9193e-1 (2.88e-2) - 6.0335e-1 (1.16e-2) - 6.0408e-1 (1.24e-2) = 6.3612e-1 (1.28e-2) + 6.1627e-1 (2.64e-2) WFG3 1.4322e-1 (1.41e-2) = 1.2063e-1 (4.24e-3) - 1.2042e-1 (3.60e-3) - 1.2101e-1 (7.60e-3) - 1.5056e-1 (2.18e-2) WFG4 3.4791e-1 (1.32e-2) - 3.2407e-1 (4.54e-3) - 3.2375e-1 (5.65e-3) - 3.4713e-1 (7.83e-3) - 3.8003e-1 (1.25e-2) WFG5 3.0356e-1 (2.12e-2) - 2.3692e-1 (5.51e-3) - 2.3513e-1 (5.10e-3) - 2.6523e-1 (1.38e-2) - 3.1563e-1 (1.57e-2) WFG6 2.1872e-1 (2.13e-2) - 1.7379e-1 (4.27e-3) - 1.7455e-1 (4.08e-3) - 1.9790e-1 (1.20e-2) - 2.2965e-1 (1.59e-2) WFG7 2.8532e-1 (1.53e-2) - 2.5677e-1 (5.08e-3) - 2.5647e-1 (5.71e-3) - 2.7283e-1 (8.96e-3) - 2.9653e-1 (2.43e-2) WFG8 2.6816e-1 (1.75e-2) - 2.3295e-1 (9.53e-3) - 2.3438e-1 (5.73e-3) - 2.5781e-1 (1.23e-2) - 2.7954e-1 (1.74e-2) WFG9 2.8609e-1 (5.70e-2) = 1.8145e-1 (1.03e-2) - 1.8449e-1 (1.28e-2) - 1.8955e-1 (2.04e-2) - 2.8012e-1 (5.37e-2) + / - / = 1 / 3 / 3 0 / 7 / 0 0 / 7 / 0 0 / 7 / 0 The real-world dataset validation experiment uses the TIDIP time-frequency domain image dataset for weld defects as the research object. Two sets of experimental datasets are constructed: a stationary dataset and a non-stationary dataset. The stationary dataset is the original TIDIP dataset, containing time-frequency domain image samples of five types of weld defects, including cracks, porosity, and slag inclusions. The non-stationary dataset simulates data distribution drift in real-world scenarios (such as sensor drift and environmental noise fluctuations) by adding Gaussian noise to the original samples. The core task of the experiment is to verify the generalization optimization effect of this scheme. The lightweight MobileNetV2 classification network, widely used in industrial scenarios, is selected as the validation vehicle. Its 14 key hyperparameters (Table 4) are optimized using a three-objective approach, and the classification performance and efficiency of each algorithm after optimization are compared. The experimental procedure is as follows: the two datasets are divided into training and testing sets in an 8:2 ratio. MetaDCP-SAEA is used to optimize the network hyperparameters. The classification accuracy and total training time of the optimized model on the testing set are recorded, and the accuracy reduction of the non-stationary state relative to the stationary state is calculated. Simultaneously, the class sensitivity (multi-class recognition balance) and single evaluation time of each algorithm are statistically analyzed. Figure 7The Pareto front distribution obtained under non-stationary conditions is shown. Three typical configurations in regions with accuracy above 0.8 are selected: point A (optimal accuracy, 95.83%), point B (lowest computational complexity), and point C (shortest training time). The hyperparameter results of the algorithm optimization for the two datasets are shown in Table 5. It is worth noting that in this embodiment, compared with the stationary state, the algorithm adaptively adjusts key parameters for the MobileNetV2 network under non-stationary conditions: the learning rate is increased from 0.00195 to 0.0043, the initial number of channels is increased from 16 to 24, and the number of training epochs is increased from 5 to 9. This adjustment mechanism originates from the risk-aware architecture of MetaDCP-SAEA: when data distribution drift causes an increase in prediction uncertainty, the evaluation layer captures prediction fluctuations through parameter perturbation integration, combines the meta-regression model to accurately quantify risk signals, and then drives the decision layer to shift to a more robust parameter space that is suitable for the target network. This adaptive adjustment mechanism can be applied to various deep convolutional network types. Specifically, the increased learning rate enhances the model's convergence adaptability in noisy data, the increased number of channels resists signal attenuation by improving feature extraction redundancy, and the increased number of training epochs ensures that the model fully learns effective patterns, ultimately enabling the model to maintain an accuracy of 87.04% under non-stationary noise, with the decrease controlled within 9.17%. Table 6 shows the performance comparison of the experimental results with six advanced comparison algorithms: MetaDCP-SAEA achieves a classification accuracy of 95.83% in stationary conditions and 87.04% in non-stationary conditions, with an accuracy drop of only 9.17%, lower than most comparison algorithms (such as EDN-ARMOEA with a drop of 9.85% and REMO with a drop of 10.46%). The training time is reduced by an average of 12% compared to the six comparison algorithms, and the class sensitivity reaches 86.98%. While maintaining high classification accuracy and robustness, it shows better computational efficiency, fully verifying the effectiveness and efficiency of the proposed algorithm in dealing with the hyperparameter configuration of non-stationary temporal signal networks. Table 4. Optimization parameters and optimization range of typical deep convolutional integral networks (taking MobileNetV2 as an example); Parameters to be optimized Optimization range LearningRate [1e-4,1e-2] OptimizerType ["Adam""RMSprop"] BatchSize [4, 16] NumEpochs [5, 10] InitChannels ["16”"24”] Stage1ExpansionRatio ["2”"4”] Stage1OutputChannels ["16”"24”] Stage2ExpansionRatio ["3”"4”] Stage2OutputChannels ["32”"40”"48”] Stage3ExpansionRatio ["4”"6”] Stage3OutputChannels ["64”"80”"96”] Stage4ExpansionRatio ["4”"6”] Stage4OutputChannels ["96”"112”"128”] DropoutRate [0.1, 0.6] Table 5 shows the 14-dimensional decision variables searched by the MobileNetV2 network in typical embodiments when the accuracy is optimal in both stationary and non-stationary states. Parameters to be optimized Stable state Non-stationary state LearningRate 0.00195 0.0043 OptimizerType RMSprop Adam BatchSize 8 8 NumEpochs 5 9 InitChannels 16 24 Stage1ExpansionRatio 4 2 Stage1OutputChannels 16 24 Stage2ExpansionRatio 3 3 Stage2OutputChannels 32 32 Stage3ExpansionRatio 4 4 Stage3OutputChannels 64 64 Stage4ExpansionRatio 4 4 Stage4OutputChannels 96 112 DropoutRate 0.427 0.423 Table 6 compares six state-of-the-art algorithms with the MetaDCP-SAEA proposed in this paper, with the best results highlighted. EDN-ARMOEA ADSAPSO REMO MO-L2SMEA DirHVEI DCSCSAEA MetaDCP-SAEA Steady-state accuracy / % 93.98 94.44 93.06 94.44 94.91 95.37 95.83 Non-stationary accuracy / % 84.72 85.19 83.33 87.04 85.65 85.65 87.04 Accuracy decrease under non-stationary conditions / % 9.85 9.79 10.46 7.84 9.76 10.19 9.17 Class sensitivity under non-stationary conditions / % 84.56 85.16 82.5 86.76 85.32 84.68 86.98 Total optimization time under non-stationary conditions / h 28.17 29.8 24.47 29.32 27.23 26.39 24.27 The above description is merely a specific embodiment of the present invention, enabling those skilled in the art to understand or implement the present invention. Although detailed descriptions have been provided with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments, and they should all be covered within the protection scope of the claims.

Claims

1. A lightweight adaptive optimization method for high-dimensional hyperparameters of CNNs for classification of non-stationary time-series signals, characterized by: Using a deep CNN adapted for classifying non-stationary temporal signals as the core classification carrier, the non-stationary temporal signals include industrial weld defect echo signals, medical heart sound / ECG signals, and industrial equipment vibration monitoring signals. Addressing the pain points of CNN classification performance degradation under non-stationary perturbations, high-dimensional coupling of hyperparameters, and high costs of real-world training and evaluation, a closed-loop co-evolutionary proxy architecture of "perception-evaluation-decision" is constructed. The structural hyperparameters and training hyperparameters of the CNN are used as the evolutionary population individuals. Through the MetaDCP-SAEA optimization strategy, multi-objective efficient configuration of CNN hyperparameters under non-stationary perturbations is achieved. The co-evolutionary proxy architecture includes a sequentially connected, closed-loop feedback structural perception dual-branch proxy module, a meta-learning enhanced uncertainty evaluation module, and a three-objective adaptive decision-making module. These three modules collaboratively execute the following steps: Perception Phase: The structure-aware dual-branch proxy module fits the complex nonlinear mapping relationship between the high-dimensional hyperparameter combination of CNN and the robust classification performance of time-series signals under non-stationary perturbations. The structure-aware dual-branch proxy module adopts a heterogeneous architecture with parallel local and long-range branches to capture the local tight coupling features and cross-dimensional long-range collaborative features between CNN hyperparameters, respectively. After multinomial nonlinear mapping, it outputs the robust classification performance prediction value, providing an accurate performance prediction basis for subsequent evaluation and decision-making. Evaluation phase: The uncertainty assessment module is enhanced through meta-learning. Based on the robust classification performance prediction value, a two-stage mechanism of "parameter perturbation integration + meta-regression calibration" is adopted to dynamically quantify the uncertainty risk in the CNN hyperparameter performance prediction process and output the calibrated risk score. This enables the perception of the distribution drift of time series signal data in non-stationary environments and provides a risk decision basis for the robustness screening of CNN hyperparameters. Decision and closed-loop update phase: Through the three-objective adaptive decision module, with "robust classification quality - prediction uncertainty - solution diversity" as the core optimization objective, and combining the robust classification performance prediction value and the calibrated risk score, candidate hyperparameter combinations that take into account CNN classification performance, robustness in non-stationary environments and search space exploration are selected. For the selected candidate hyperparameter combinations, perform real training and evaluation of the CNN model in the non-stationary time-series signal classification task, obtain the three real target values ​​of robust classification error rate, computational complexity, and training time, and feed the evaluation results back to update the model parameters of the two preceding modules; finally, output the optimal set of non-dominated hyperparameters and their corresponding target values ​​on the Pareto front, and complete the adaptive configuration of hyperparameters of the CNN model for the robust classification task of non-stationary time-series signals.

2. The method according to claim 1, characterized in that, The optimization objective for robust classification of non-stationary time-series signals is pre-modeled as a three-objective minimization problem, with the objective vector defined as: , Dimensional differences are eliminated through target normalization; where Robust classification error rate This represents the average classification accuracy under non-stationary perturbations. To calculate the computational complexity, the floating-point operations of the convolutional layer and the fully connected layer are summed. Training time is defined as the total time from model initialization to convergence.

3. The method according to claim 1, characterized in that, It also includes a standardized preprocessing process adapted to non-stationary time series signals, providing standardized training data for hyperparameter optimization of deep convolutional integral network, specifically including: (1) Time-frequency domain mapping: using short-time Fourier transform to convert one-dimensional time series signals into two-dimensional time-frequency domain images adapted to feature extraction of deep convolutional integral network; (2) Non-stationary disturbance simulation: adding Gaussian noise of intensity 0-0.1 to the time-frequency domain image to simulate data distribution drift in actual scenarios and construct a non-stationary disturbance dataset; (3) Data augmentation: performing horizontal / vertical flipping operations on the preprocessed time-frequency domain samples to expand the diversity of training data and improve the robustness of the model after hyperparameter optimization; the preprocessing process is adapted to non-stationary time series data such as industrial weld defect echo signals, medical heart sounds / electrocardiogram signals, and industrial equipment vibration monitoring signals.

4. The method according to claim 1, characterized in that: The structure-aware dual-branch proxy module is a convolutional multinomial proxy model, employing a heterogeneous architecture with parallel local and long-range branches. The local branches act independently on each decision variable using 1×1 convolutional kernels, extracting locally tightly coupled hyperparameter features and only expanding channels without introducing cross-dimensional information interaction. The long-range branches use 9×1 convolutional kernels with "same" padding, aggregating adjacent dimensional information through a sliding window to capture cross-dimensional long-range collaborative features, maintaining consistency between the output and input dimensions. Both branches are followed by a 3rd-order multinomial network to achieve nonlinear mapping, performing power expansion on the output features of each branch and mapping them to scalar outputs. The output of the k-th branch is calculated as follows: , Let i be the learnable i-th order coefficient vector. The feature vectors extracted from the two branches are used to obtain the robust classification performance prediction value by averaging the outputs of the two branches with equal weights.

5. The method according to claim 4, characterized in that: The structure-aware dual-branch surrogate model is trained using a supervised learning paradigm, combining adaptive optimization strategies and regularization mechanisms to improve generalization ability. Training data is derived from archived hyperparameter combinations and corresponding quality labels that have undergone real-world evaluation. The objective function uses mean squared error. , These are the model's predicted values. Authentic quality label The number of training samples is [number]; the Adam optimizer is used for parameter updates, with an initial learning rate set to [value]. The learning rate is 0.001, and it decays to 0.9 times the previous learning rate every 50 rounds. An early stopping mechanism is introduced, with a loss monitoring window set for the validation set (20% of the training data). If the loss on the validation set decreases by less than 0.001 for 10 consecutive rounds, the early stopping mechanism will be implemented. If the training stops and the current optimal parameters are saved, then training will be terminated.

6. The method according to claim 1, characterized in that, The meta-learning-enhanced uncertainty assessment module employs a two-stage mechanism of "parameter perturbation integration + meta-regression calibration" to quantify prediction uncertainty and distinguish its sources, providing a risk decision basis for the three-objective adaptive decision-making module. In the parameter perturbation integration stage, a five-level linearly increasing Gaussian noise perturbation model is applied to the parameters of the converged structure-aware dual-branch surrogate model during training, and the prediction standard deviation is calculated. , These are the predicted values ​​for each disturbance model. To predict the mean, the primary uncertainty is quantified by the predicted standard deviation; In the meta-regression calibration stage, a meta-regression model is constructed by integrating multi-dimensional features to calibrate the primary uncertainty and output a calibrated predicted risk score in the range of [0,1].

7. The method according to claim 6, characterized in that, The input features of the meta-regression model are standardized by Z-score and fixed to 15 dimensions, covering five complementary features: basic features (hyperparameter decision variable vector, parameter perturbation ensemble mean and standard deviation), statistical features (mean, variance, skewness, kurtosis and quartiles of the hyperparameter combination), interaction features (product term of ensemble mean and standard deviation), density features (density of 3 / 5 / 7 nearest neighbor samples), and distance features (Euclidean distance to the centroid of the training sample and the most similar sample). The meta-regression model adopts a single-layer linear regression architecture, and the output layer is connected to the sigmoid activation function to limit the predicted risk score to the [0,1] interval. The linear weight parameters are initialized to random values ​​following the N (0,0.01) distribution, and the bias term is initialized to 0. The monitoring signal is generated by normalized linear fusion of quality deviation and integrated fluctuation: Let the quality deviation of the i-th sample be... , The predicted values ​​are from the structure-aware dual-branch surrogate model. For accurate assessment, fluctuations are integrated. Let the standard deviation of the predictions be the integrated parameter perturbation, then the supervision signal is: , The training process introduces contrastive learning loss. ,in For marginal parameters, , These are the predicted scores for high- and low-risk samples, respectively, to improve the accuracy of the discrimination.

8. The method according to claim 1, characterized in that, The three-objective adaptive decision-making module takes "robust classification quality, prediction uncertainty, and solution diversity" as its core optimization objectives and achieves adaptive screening in non-stationary environments through a multi-stage decision-making mechanism. First, it performs non-dominated sorting on the candidate hyperparameter combination set, divides it into multiple Pareto fronts according to dominance relationships, and calculates the proportion of the first front. ,in For the first frontier individual set, C For candidate set; based on history median of sequence and absolute median Constructing the Sigmoid decision function: , For smoothing terms, the probability of the "exploitation-exploration" strategy is adaptively adjusted; when the "exploitation" strategy (probability) is adopted... When, based on crowd distance: , Individuals in sparse regions of the first frontier are selected; when an "exploration" strategy (probability-based) is adopted. When selecting frontiers, the selection quota is allocated in descending order of frontier level, with the first frontier accounting for 50% of the quota, and the remaining levels allocated in order of rank to ensure the coverage of the search scope; at the same time, the number of selections is dynamically adjusted according to the size of the frontier. When the size of the frontier is greater than 30% of the candidate set, 20% of the selection slots are added, and when no new frontiers are generated for 3 consecutive rounds, an additional 15% of the selection slots are added to balance convergence and exploration.

9. The method according to claim 1, characterized in that, It also includes an adaptive reference solution selection mechanism, which generates a surrogate model training benchmark by integrating a dual-objective selection strategy that combines convergence and diversity, providing high-quality supervision signals for the structure-aware dual-branch surrogate module; the input is the current archived solution set and the preset number of reference solutions. First, the frontier levels are divided using non-dominated sorting. According to the scale of the first frontier Adaptive adjustment of selection logic: like Select directly All individuals in the solution are considered as reference solutions; like From the subsequent frontier, quality scoring ( (For individual target vectors) replenished to the quantity Q ; like First select in descending order of quality score. Each individual, then using the maximum-minimum distance strategy ( S For the selected set, (For the candidate set) supplement the remaining individuals to the number Q ; Calculate the minimum distance from an individual in the population to the reference solution. Define quality score The top 80% of the scores were labeled as positive samples and the bottom 20% as negative samples, thus generating training labels for the surrogate model.

10. The method according to claim 1, the cross-network architecture adaptive adaptation mechanism of the present invention is implemented through a three-level process of "dynamic adaptation of search space - hyperparameter normalization encoding - core architecture unmodified adaptation", as follows: The first level is dynamic adaptation of search space: for different types of target depth convolutional neural network (including but not limited to mainstream convolutional neural network architectures such as MobileNet series, ResNet series, EfficientNet series, ShuffleNet series, etc.), the dimension and value range of the hyperparameter search space are dynamically matched. The hyperparameters to be optimized are uniformly divided into two categories: structural parameters and training parameters. The structural parameters include parameters directly related to the network structure, such as the number of convolution channels, convolution kernel size, dilation rate, bottleneck layer dilation ratio, and the number of output channels of each network layer. The training parameters include parameters related to the model training process, such as learning rate, optimizer type, batch size, training epochs, and dropout rate. For lightweight networks (such as the MobileNet series), the upper limit of the number of convolution channels and the dilation rate is reduced to adapt to the lightweight characteristics; for deep residual networks (such as the ResNet series), the range of training epochs and the learning rate is expanded to adapt to the convergence characteristics of deep networks, so as to achieve personalized adaptation of the search space for different CNN architectures. The second level is hyperparameter standardization encoding: a one-dimensional vector unified encoding method of "structural parameters first, training parameters last" is adopted. First, all the hyperparameters to be optimized in the target CNN are uniformly encoded into continuous or discrete standardized decision variables according to their numerical type. Discrete parameters (such as optimizer type, channel number level) are converted into standardized variables through one-hot encoding or integer encoding, and continuous parameters (such as learning rate, dropout rate) are mapped to the [0,1] interval through min-max normalization. Then, according to the preset order of "hierarchical structural parameters from first to last, global training parameters last", all standardized decision variables are concatenated into a one-dimensional standardized decision vector, and the dimension of this vector is completely matched with the input dimension of the structure-aware dual-branch surrogate module, ensuring that the hyperparameter combination of heterogeneous networks can be adapted to the local branch and long-range branch feature extraction logic of the dual-branch surrogate model. The third level is the core architecture adaptation without modification: the encoded standardized decision vector can be directly input into the "perception-evaluation-decision" closed-loop collaborative architecture of this invention. No modifications are needed to the network structure, computational logic, or iterative optimization mechanism of the structure perception dual-branch proxy module, the meta-learning enhanced uncertainty evaluation module, and the three-objective adaptive decision module. Through the hierarchical collaboration of the three modules, the hyperparameter adaptive optimization of the corresponding CNN architecture in the robust classification scenario of non-stationary temporal signals can be completed, realizing the universal adaptation of the core mechanism of this invention to convolutional network architectures of different depths.

Images