Intelligent hardware configuration optimization method based on deep learning

Abstract

The invention discloses an intelligent hardware configuration optimization method based on deep learning. The method comprises the following steps of collecting system operation data and performing preprocessing; generating a high-dimensional input feature matrix by using an input feature analysis module of the improved NCGM model; introducing a microsynaptic plasticity mechanism to form initial local performance characterization; introducing a continuous time topological evolution mechanism, and constructing a dynamic topological structure; forming a global performance representation vector; adopting a multi-domain collaborative adaptive optimization mechanism to obtain a convergence parameter set; and outputting a hardware configuration parameter set, thereby realizing intelligentization and adaptive configuration optimization of the target hardware system, and improving the self-learning ability, performance prediction precision and energy consumption utilization rate of the hardware system.

Description

Technical Field

[0001] This invention relates to the field of intelligent hardware system optimization, and in particular to an intelligent hardware configuration optimization method based on deep learning. Background Technology

[0002] With the rapid development of artificial intelligence and intelligent hardware systems, hardware resource allocation and performance optimization have gradually become significant bottlenecks restricting overall system efficiency. Currently, most hardware configuration optimization methods still rely on fixed configuration strategies or experience-based manual tuning, making it difficult to cope with complex and ever-changing task scenarios and the real-time balancing of multi-dimensional performance goals. For example, in server clusters, edge computing devices, or embedded intelligent systems, different computing tasks vary significantly in terms of load characteristics, energy consumption requirements, and latency constraints. Existing methods generally lack adaptive dynamic configuration capabilities, leading to insufficient hardware utilization, energy waste, and decreased system response performance.

[0003] In existing technologies, some studies attempt to predict and optimize hardware parameters using machine learning models. However, these methods often employ static neural network structures, whose structures are fixed after training and cannot be dynamically adjusted based on changes in input data distribution or hardware operating conditions. Furthermore, these models typically only use gradient descent-based parameter updates during optimization, failing to capture the complex dependencies and nonlinear coupling characteristics between parameters in the hardware system. Due to the lack of structural adaptation and dynamic feedback mechanisms, existing models often suffer from slow convergence, poor generalization performance, and insufficient prediction accuracy. Summary of the Invention

[0004] One objective of this invention is to propose a deep learning-based intelligent hardware configuration optimization method. This invention fully utilizes the self-organizing computational capability and adaptive topological evolution characteristics of the improved NCGM model, and combines differentiable synaptic plasticity mechanism, time constant adaptive method and Laplace domain supervised optimization strategy to achieve adaptive configuration and dynamic optimization of the target hardware system under multi-dimensional performance indicators.

[0005] A method for optimizing the configuration of intelligent hardware based on deep learning according to an embodiment of the present invention includes the following steps: Collect and preprocess system operation data from the target hardware system; The preprocessed system running data is input into the input feature parsing module of the improved NCGM model, and a high-dimensional input feature matrix is ​​generated by using nonlinear function combination mapping and normalization. The high-dimensional input feature matrix is ​​input into the cell state initialization module of the improved NCGM model, and a differentiable synaptic plasticity mechanism is introduced to form an initial local performance representation. Based on the initial local performance characterization, a continuous-time topology evolution mechanism is introduced into the dynamic topology adaptation module of the improved NCGM model to construct a dynamic topology structure. Based on the dynamic topology, a global performance representation vector is formed using an adaptive time constant method. In the supervised learning and parameter optimization module of the improved NCGM model, a multi-domain collaborative adaptive optimization mechanism is adopted to map the global performance representation vector and the target performance index to the Laplace frequency domain to obtain the convergent parameter set. Based on the convergence parameter set, the prediction error is calculated according to the target performance index. The weights of the NCGM model are continuously updated and improved through supervised optimization and topology self-evolution mechanism, and the hardware configuration parameter set is output.

[0006] Optionally, the system operation data includes central processing unit utilization, graphics processing unit utilization, memory usage, bandwidth usage, temperature, power consumption, response time, and task load characteristics. The preprocessing includes noise removal, normalization, and feature standardization.

[0007] Optionally, the generation of the high-dimensional input feature matrix specifically includes: The preprocessed system operation data is used to construct an input feature matrix, which consists of n system operation data features, each corresponding to a monitoring dimension of the target hardware system during operation. The input feature matrix is ​​input into the input feature parsing module of the improved NCGM model, and nonlinear function combination mapping is performed to generate a high-dimensional feature representation vector. The generation of the high-dimensional nonlinear expression is based on the Kolmogorov-Arnold representation structure. Each system running data feature is transformed through a learnable one-dimensional nonlinear function and weighted summation is performed using a combination function. The high-dimensional feature representation vectors are normalized by using the minimum and maximum value standardization method to obtain a normalized feature vector set. The normalized feature vector set is recombined according to a preset dimensional order to generate a high-dimensional input feature matrix.

[0008] Optionally, the formation of the initial local performance characterization specifically includes: The high-dimensional input feature matrix is ​​input into the cell state initialization module of the improved NCGM model. According to the mapping relationship of the high-dimensional input feature matrix, a mapping matrix is ​​obtained to form an initial state parameter set. The initial state parameter set is formed by linear combination of the mapping matrix and combined with nonlinear function modulation. Finally, the initial state parameter set of each computation cell is output. The initial state parameter set includes state vector, synaptic weight parameters and plasticity trace parameters. Based on the initial state parameter set, a corresponding input feature vector is assigned to each computational cell, and the input feature vector is linearly transformed using a mapping matrix to generate the original initial state value. Based on the original initial state values, a set of connection parameters between computational cells is defined. The set of connection parameters includes synaptic connection weight parameters and a plasticity memory factor. The synaptic connection weight refers to the signal transmission strength between computational cells, and the plasticity memory factor refers to the influence of historical states on the current synaptic strength. Using a differentiable synaptic plasticity mechanism, the set of connection parameters is dynamically updated according to the activity correlation between computational cells. The update process is that the synaptic connection between each pair of computational cells decays historical state information through a forgetting factor, and the weight is updated by accumulating based on the current activity strength of the computational cell. Based on the updated set of connection parameters, neighborhood information is integrated into the initial state vector of each computational cell to generate the initial local performance representation of the computational cell.

[0009] Optionally, the construction of the dynamic topology specifically includes: The initial local performance characterization is input into the dynamic topology adaptive module of the improved NCGM model to establish the computational cell state matrix. Based on the computational cell state matrix, the connection strength is calculated using an exponential decay function to generate the initial connection matrix. The initial connection matrix is ​​dynamically topologically evolved in the continuous time dimension. The state vector of the computation cell state matrix is ​​dynamically updated by the state evolution weight matrix and nonlinear activation function. The connection strength in the initial connection matrix is ​​adaptively adjusted according to the feedback results of the performance objective function and the regularization constraint. During the dynamic topology evolution process, the initial connection matrix is ​​normalized, and the proportion of connection weights of each computational cell is calculated to obtain the distribution result of the normalized connection weights. The normalization process refers to calculating the sum of the connection weights of each computational cell and scaling it proportionally. Based on the distribution of normalized connection weights, the neighborhood set of each computational cell at the current time is obtained. When the connection strength is greater than a preset threshold, the corresponding computational cell is included in the neighborhood set. During the continuous time evolution process, the neighborhood sets at all times are recorded and averaged to generate a neighborhood evolution feature set. Based on the neighborhood evolution feature set, the stability of the dynamic topology is determined. When the rate of change of the neighborhood set is less than a set threshold, the final normalized connection matrix and the corresponding neighborhood set are output, forming a dynamic topology.

[0010] Optionally, the generation of the global performance representation vector specifically includes: Based on the connection weight distribution and state activation intensity of each computational cell in the dynamic topology, the average connection influence factor of each computational cell is calculated, and a time constant vector is generated. The time constant vector refers to the rate of state change of the computational cell in the continuous time dimension. Each computational cell corresponds to a time constant parameter, and all time constant parameters constitute the time constant vector. An adaptive time constant method is used to dynamically adjust the time constant vector to generate an adjusted time constant vector. The adjusted time constant vector is generated by calculating the rate of change of the time constant by the current state intensity of each computation cell and the time decay coefficient, so that the time constant increases when the state intensity of the computation cell is large and gradually decreases when the state tends to be stable, thus obtaining the adjusted time constant vector. Based on the adjusted time constant vector, the state of each computation cell is updated to obtain the updated computation cell. The updated computation cell is obtained by weighted summation of the current state of the computation cell and the states of the neighboring computation cells according to the normalized connection matrix, and then linear combination of the self-state mapping matrix, the neighboring state mapping matrix and the bias vector. The updated computational cells are expanded into matrix form, and the state matrices of all computational cells are updated synchronously. Under the constraint mechanism of the time constant vector, the state matrix of the computational cells at the global time endpoint is obtained by integrating the continuous time changes of the state of each computational cell. The column vectors of the computational cell state matrix at the global time endpoint are aggregated to obtain an aggregated vector. The aggregated vector is then linearly mapped using the readout matrix and readout bias vector to generate a global performance characterization vector.

[0011] Optionally, the output of the convergence parameter set specifically includes: The global performance representation vector and the target performance index are input into the supervised learning and parameter optimization module of the improved NCGM model to establish a mapping model between the global performance representation vector and the target performance index. Based on the mapping model, the global performance representation vector and the target performance index are mapped to the Laplace frequency domain to obtain the corresponding frequency domain representation. By calculating the difference between the global performance representation vector and the target performance index in the Laplace frequency domain, a frequency domain error function is formed. A supervised learning loss function is constructed by jointly using the frequency domain error function with the difference between the global performance representation vector and the target performance index. The supervised learning loss function is a weighted sum of the frequency domain error constraint term and the time domain error constraint term through a joint time and frequency balance coefficient. Using the supervised learning loss function as the optimization objective, a multi-domain collaborative adaptive optimization mechanism is adopted to minimize the parameter set of the mapping model. The gradient of the supervised learning loss function relative to the parameter set of the mapping model is calculated using the frequency domain backpropagation algorithm, and then mapped back to the time domain to perform parameter updates, resulting in the updated parameter set. The updated parameter set is re-input into the improved NCGM model, and the Laplacian frequency domain mapping, supervised learning loss function calculation and parameter update process are repeated until the convergence condition is met, and the converged parameter set is output.

[0012] Optionally, the output of the hardware configuration parameter set specifically includes: The NCGM model is improved based on the configuration of the convergence parameter set. Real-time collected system operation data is input into the improved NCGM model to generate a predictive performance output vector. The difference between the predicted performance output vector and the target performance index is calculated to obtain the prediction error vector, which refers to the degree of deviation between the output of the improved NCGM model and the target performance index. Based on the supervised learning loss function, the parameter set of the improved NCGM model is optimized in a supervised manner according to the prediction error vector to obtain the optimized and updated parameter set. The optimized and updated parameter set is obtained by taking the sum of squares of the prediction error vector as the prediction error term in the loss function, and introducing the squared difference between the current parameter set and the convergent parameter set as the parameter constraint term. Based on the optimized and updated parameter set, a topology self-evolution mechanism is enabled. The connection strength between computational cells is dynamically adjusted according to the updated weight distribution and node activation intensity to obtain the updated normalized connection matrix. The updated normalized connection matrix is ​​input into the improved NCGM model for performance evaluation. When the prediction error vector is less than a preset threshold, a set of hardware configuration parameters is output. The set of hardware configuration parameters includes computing resource allocation parameters, storage and cache configuration parameters, communication and bus control parameters, energy consumption control parameters, and topology mapping and structure scheduling parameters.

[0013] The beneficial effects of this invention are: This invention proposes an adaptive configuration optimization method for intelligent hardware systems by introducing an improved NCGM model. This method utilizes the feature representation capabilities and self-organizing evolution mechanism of deep learning models to achieve a high-dimensional mapping relationship between hardware configuration parameters and performance indicators. It can autonomously learn and continuously optimize system performance under complex operating environments, effectively overcoming the shortcomings of existing technologies such as reliance on human experience in hardware configuration, low optimization efficiency, and insufficient dynamic response.

[0014] This invention introduces differentiable synaptic plasticity and dynamic topology adaptation mechanisms, enabling the improved NCGM model to dynamically adjust its internal structure and connection strength based on the characteristic distribution of system operating data and performance feedback. This achieves adaptive collaboration between computational cells and high-precision generation of global performance representations. Simultaneously, by combining a time constant adaptation method, the model exhibits better time responsiveness and stability during state updates, thereby improving the continuity and convergence of the optimization process.

[0015] Furthermore, this invention employs a supervised optimization strategy based on the Laplace domain, achieving efficient parameter learning and weight updates under joint constraints in the frequency and time domains. This mechanism enhances the model's generalization ability and stability in a multi-dimensional feature space, resulting in higher accuracy and robustness in the hardware configuration optimization process. In summary, this invention realizes the transformation of intelligent hardware systems from static configuration to dynamic self-evolutionary optimization, possessing high accuracy, high adaptability, and strong generalization ability, significantly improving the performance utilization and operational reliability of the hardware system. Attached Figure Description

[0016] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0017] Figure 1 This is an overall flowchart of a deep learning-based intelligent hardware configuration optimization method proposed in this invention. Figure 2 This is a schematic diagram of the module structure of an improved NCGM model for a deep learning-based intelligent hardware configuration optimization method proposed in this invention. Detailed Implementation

[0018] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0019] refer to Figure 1-2 A method for optimizing the configuration of intelligent hardware based on deep learning includes the following steps: Collect and preprocess system operation data from the target hardware system; The preprocessed system running data is input into the input feature parsing module of the improved NCGM model, and a high-dimensional input feature matrix is ​​generated by using nonlinear function combination mapping and normalization. The high-dimensional input feature matrix is ​​input into the cell state initialization module of the improved NCGM model, and a differentiable synaptic plasticity mechanism is introduced to form an initial local performance representation. Based on the initial local performance characterization, a continuous-time topology evolution mechanism is introduced into the dynamic topology adaptation module of the improved NCGM model to construct a dynamic topology structure. Based on the dynamic topology, a global performance representation vector is formed using an adaptive time constant method. In the supervised learning and parameter optimization module of the improved NCGM model, a multi-domain collaborative adaptive optimization mechanism is adopted to map the global performance representation vector and the target performance index to the Laplace frequency domain to obtain the convergent parameter set. Based on the convergence parameter set, the prediction error is calculated according to the target performance index. The weights of the NCGM model are continuously updated and improved through supervised optimization and topology self-evolution mechanism, and the hardware configuration parameter set is output.

[0020] In this embodiment, the system operation data includes central processing unit utilization, graphics processing unit utilization, memory usage, bandwidth usage, temperature, power consumption, response time, and task load characteristics. The preprocessing includes noise removal, normalization, and feature standardization.

[0021] In this embodiment, the generation of the high-dimensional input feature matrix specifically includes: The preprocessed system operation data is used to construct an input feature matrix, which consists of n system operation data features, each corresponding to a monitoring dimension of the target hardware system during operation. The input feature matrix is ​​input into the input feature parsing module of the improved NCGM model, and nonlinear function combination mapping is performed to generate a high-dimensional feature representation vector. The generation of the high-dimensional nonlinear expression is based on the Kolmogorov-Arnold representation structure. Each system running data feature is transformed through a learnable one-dimensional nonlinear function and weighted summation is performed using a combination function. The high-dimensional feature representation vectors are normalized by using the minimum and maximum value standardization method to obtain a normalized feature vector set. The normalized feature vector set is recombined according to a preset dimensional order to generate a high-dimensional input feature matrix.

[0022] In this embodiment, the formation of the initial local performance characterization specifically includes: The high-dimensional input feature matrix is ​​input into the cell state initialization module of the improved NCGM model. According to the mapping relationship of the high-dimensional input feature matrix, a mapping matrix is ​​obtained to form an initial state parameter set. The initial state parameter set is formed by linear combination of the mapping matrix and combined with nonlinear function modulation. Finally, the initial state parameter set of each computation cell is output. The initial state parameter set includes state vector, synaptic weight parameters and plasticity trace parameters. Based on the initial state parameter set, a corresponding input feature vector is assigned to each computational cell, and the input feature vector is linearly transformed using a mapping matrix to generate the original initial state value. Based on the original initial state values, a set of connection parameters between computational cells is defined. The set of connection parameters includes synaptic connection weight parameters and a plasticity memory factor. The synaptic connection weight refers to the signal transmission strength between computational cells, and the plasticity memory factor refers to the influence of historical states on the current synaptic strength. Using a differentiable synaptic plasticity mechanism, the set of connection parameters is dynamically updated according to the activity correlation between computational cells. The update process is that the synaptic connection between each pair of computational cells decays historical state information through a forgetting factor, and the weight is updated by accumulating based on the current activity strength of the computational cell. Based on the updated set of connection parameters, neighborhood information is integrated into the initial state vector of each computational cell to generate the initial local performance representation of the computational cell.

[0023] In this embodiment, the construction of the dynamic topology specifically includes: The initial local performance characterization is input into the dynamic topology adaptive module of the improved NCGM model to establish the computational cell state matrix. Based on the computational cell state matrix, the connection strength is calculated using an exponential decay function to generate the initial connection matrix. The initial connection matrix is ​​dynamically topologically evolved in the continuous time dimension. The state vector of the computation cell state matrix is ​​dynamically updated by the state evolution weight matrix and nonlinear activation function. The connection strength in the initial connection matrix is ​​adaptively adjusted according to the feedback results of the performance objective function and the regularization constraint. During the dynamic topology evolution process, the initial connection matrix is ​​normalized, and the proportion of connection weights of each computational cell is calculated to obtain the distribution result of the normalized connection weights. The normalization process refers to calculating the sum of the connection weights of each computational cell and scaling it proportionally. Based on the distribution of normalized connection weights, the neighborhood set of each computational cell at the current time is obtained. When the connection strength is greater than a preset threshold, the corresponding computational cell is included in the neighborhood set. During the continuous time evolution process, the neighborhood sets at all times are recorded and averaged to generate a neighborhood evolution feature set. Based on the neighborhood evolution feature set, the stability of the dynamic topology is determined. When the rate of change of the neighborhood set is less than a set threshold, the final normalized connection matrix and the corresponding neighborhood set are output, forming a dynamic topology.

[0024] In this embodiment, the generation of the global performance representation vector specifically includes: Based on the connection weight distribution and state activation intensity of each computational cell in the dynamic topology, the average connection influence factor of each computational cell is calculated, and a time constant vector is generated. The time constant vector refers to the rate of state change of the computational cell in the continuous time dimension. Each computational cell corresponds to a time constant parameter, and all time constant parameters constitute the time constant vector. An adaptive time constant method is used to dynamically adjust the time constant vector to generate an adjusted time constant vector. The adjusted time constant vector is generated by calculating the rate of change of the time constant by the current state intensity of each computation cell and the time decay coefficient, so that the time constant increases when the state intensity of the computation cell is large and gradually decreases when the state tends to be stable, thus obtaining the adjusted time constant vector. Based on the adjusted time constant vector, the state of each computation cell is updated to obtain the updated computation cell. The updated computation cell is obtained by weighted summation of the current state of the computation cell and the states of the neighboring computation cells according to the normalized connection matrix, and then linear combination of the self-state mapping matrix, the neighboring state mapping matrix and the bias vector. The updated computational cells are expanded into matrix form, and the state matrices of all computational cells are updated synchronously. Under the constraint mechanism of the time constant vector, the state matrix of the computational cells at the global time endpoint is obtained by integrating the continuous time changes of the state of each computational cell. The column vectors of the computational cell state matrix at the global time endpoint are aggregated to obtain an aggregated vector. The aggregated vector is then linearly mapped using the readout matrix and readout bias vector to generate a global performance characterization vector.

[0025] In this embodiment, the output of the convergence parameter set specifically includes: The global performance representation vector and the target performance index are input into the supervised learning and parameter optimization module of the improved NCGM model to establish a mapping model between the global performance representation vector and the target performance index. Based on the mapping model, the global performance representation vector and the target performance index are mapped to the Laplace frequency domain to obtain the corresponding frequency domain representation. By calculating the difference between the global performance representation vector and the target performance index in the Laplace frequency domain, a frequency domain error function is formed. A supervised learning loss function is constructed by jointly using the frequency domain error function with the difference between the global performance representation vector and the target performance index. The supervised learning loss function is a weighted sum of the frequency domain error constraint term and the time domain error constraint term through a joint time and frequency balance coefficient. Using the supervised learning loss function as the optimization objective, a multi-domain collaborative adaptive optimization mechanism is adopted to minimize the parameter set of the mapping model. The gradient of the supervised learning loss function relative to the parameter set of the mapping model is calculated using the frequency domain backpropagation algorithm, and then mapped back to the time domain to perform parameter updates, resulting in the updated parameter set. The updated parameter set is re-input into the improved NCGM model, and the Laplacian frequency domain mapping, supervised learning loss function calculation and parameter update process are repeated until the convergence condition is met, and the converged parameter set is output.

[0026] In this embodiment, the output of the hardware configuration parameter set specifically includes: The NCGM model is improved based on the configuration of the convergence parameter set. Real-time collected system operation data is input into the improved NCGM model to generate a predictive performance output vector. The difference between the predicted performance output vector and the target performance index is calculated to obtain the prediction error vector, which refers to the degree of deviation between the output of the improved NCGM model and the target performance index. Based on the supervised learning loss function, the parameter set of the improved NCGM model is optimized in a supervised manner according to the prediction error vector to obtain the optimized and updated parameter set. The optimized and updated parameter set is obtained by taking the sum of squares of the prediction error vector as the prediction error term in the loss function, and introducing the squared difference between the current parameter set and the convergent parameter set as the parameter constraint term. Based on the optimized and updated parameter set, a topology self-evolution mechanism is enabled. The connection strength between computational cells is dynamically adjusted according to the updated weight distribution and node activation intensity to obtain the updated normalized connection matrix. The updated normalized connection matrix is ​​input into the improved NCGM model for performance evaluation. When the prediction error vector is less than a preset threshold, a set of hardware configuration parameters is output. The set of hardware configuration parameters includes computing resource allocation parameters, storage and cache configuration parameters, communication and bus control parameters, energy consumption control parameters, and topology mapping and structure scheduling parameters.

[0027] Example 1: In an industrial edge computing environment, multiple nodes need to run tasks such as video recognition, data aggregation, and intelligent decision-making simultaneously. Due to the dynamic nature of the task load, traditional hardware resource configuration methods often rely on manual experience or static parameter tables, leading to problems such as slow system response, increased energy consumption, and uneven resource utilization when the load fluctuates. For example, when the number of tasks suddenly increases or the complexity of AI models rises, traditional systems cannot reallocate CPU, GPU, and memory resources in a timely manner, resulting in increased inference latency and performance bottlenecks.

[0028] To address the aforementioned issues, this embodiment constructs a hardware configuration optimization system based on an improved NCGM model. This system collects real-time node runtime data, including CPU utilization, GPU utilization, memory utilization, average task latency, and system power consumption, and inputs this data into the improved NCGM model. The model first performs nonlinear mapping and normalization on the runtime data during the input feature parsing stage, generating a high-dimensional feature matrix. Subsequently, it utilizes the differentiable synaptic plasticity mechanism to establish the connection states between computational cells and continuously optimizes the connection weights based on performance feedback through a dynamic topology adaptive module. Combined with a time constant adaptive mechanism, the model can dynamically adjust the state update rate according to performance changes at different time scales, giving the hardware configuration optimization process self-learning and self-stabilizing characteristics.

[0029] In the experimental tests, four typical edge nodes were selected as experimental samples. Each node was equipped with an Intel Xeon Silver processor, an NVIDIA RTX A5000 GPU, and 64GB of memory. The system was compared and tested under two operating strategies: the traditional static configuration strategy (fixed CPU / GPU resource ratio);

[0030] This invention proposes an improved NCGM optimization strategy. The test task involves running object detection (YOLOv8) and temporal prediction (LSTM) models in parallel for 48 hours. Experimental data is automatically collected and recorded through the system monitoring module.

[0031] Table 1 Performance comparison between the improved NCGM method and the traditional method

[0032] The experimental results in Table 1 demonstrate that the improved NCGM optimization method of this invention outperforms the traditional configuration strategy in multiple metrics. The average response latency is reduced by approximately 29.2%, task throughput is increased by approximately 31%, and unit power efficiency is improved by over 40%, indicating that this invention can improve overall system performance while maintaining stable energy consumption. Simultaneously, the model prediction error is significantly reduced, indicating that the model possesses stronger adaptability and generalization ability under complex load environments.

[0033] Further analysis reveals that the improved NCGM model achieves coordinated updates of structure and parameters during training through Laplace domain-supervised optimization and topological self-evolution mechanisms, enabling the connection patterns between computational cells to dynamically evolve based on task feedback. When the system operating environment changes, the model can quickly adjust its internal weights to maintain stable performance. Compared to traditional static strategies, this invention enables intelligent allocation and continuous optimization of hardware resources, effectively avoiding resource imbalances such as CPU overload and GPU idleness, thereby significantly improving the overall system operating efficiency and energy utilization.

[0034] In summary, this embodiment verifies the feasibility and superiority of the method of the present invention in practical edge computing scenarios. This method possesses the characteristics of high precision, adaptability, and low energy consumption, providing a scalable and reliable technical path for the dynamic configuration and optimization of intelligent hardware systems, and has promising engineering application prospects.

Claims

1. A method for optimizing the configuration of intelligent hardware based on deep learning, characterized in that, Includes the following steps: Collect and preprocess system operation data from the target hardware system; The preprocessed system running data is input into the input feature parsing module of the improved NCGM model, and a high-dimensional input feature matrix is ​​generated by using nonlinear function combination mapping and normalization. The high-dimensional input feature matrix is ​​input into the cell state initialization module of the improved NCGM model, and a differentiable synaptic plasticity mechanism is introduced to form an initial local performance representation. Based on the initial local performance characterization, a continuous-time topology evolution mechanism is introduced into the dynamic topology adaptation module of the improved NCGM model to construct a dynamic topology structure. Based on the dynamic topology, a global performance representation vector is formed using an adaptive time constant method. In the supervised learning and parameter optimization module of the improved NCGM model, a multi-domain collaborative adaptive optimization mechanism is adopted to map the global performance representation vector and the target performance index to the Laplace frequency domain to obtain the convergent parameter set. Based on the convergence parameter set, the prediction error is calculated according to the target performance index. The weights of the NCGM model are continuously updated and improved through supervised optimization and topology self-evolution mechanism, and the hardware configuration parameter set is output.

2. The method for optimizing intelligent hardware configuration based on deep learning according to claim 1, characterized in that, The system operation data includes central processing unit utilization, graphics processing unit utilization, memory usage, bandwidth usage, temperature, power consumption, response time, and task load characteristics. The preprocessing includes noise removal, normalization, and feature standardization.

3. The method for compliance constraint of intelligent agent behavior based on multi-source data traceability chain according to claim 1, characterized in that, The generation of the high-dimensional input feature matrix specifically includes: The preprocessed system operation data is used to construct an input feature matrix, which consists of n system operation data features, each corresponding to a monitoring dimension of the target hardware system during operation. The input feature matrix is ​​input into the input feature parsing module of the improved NCGM model, and nonlinear function combination mapping is performed to generate a high-dimensional feature representation vector. The generation of the high-dimensional nonlinear expression is based on the Kolmogorov-Arnold representation structure. Each system running data feature is transformed through a learnable one-dimensional nonlinear function and weighted summation is performed using a combination function. The high-dimensional feature representation vectors are normalized by using the minimum and maximum value standardization method to obtain a normalized feature vector set. The normalized feature vector set is recombined according to a preset dimensional order to generate a high-dimensional input feature matrix.

4. The method for optimizing intelligent hardware configuration based on deep learning according to claim 1, characterized in that, The formation of the initial local performance characterization specifically includes: The high-dimensional input feature matrix is ​​input into the cell state initialization module of the improved NCGM model. According to the mapping relationship of the high-dimensional input feature matrix, a mapping matrix is ​​obtained to form an initial state parameter set. The initial state parameter set is formed by linear combination of the mapping matrix and combined with nonlinear function modulation. Finally, the initial state parameter set of each computation cell is output. The initial state parameter set includes state vector, synaptic weight parameters and plasticity trace parameters. Based on the initial state parameter set, a corresponding input feature vector is assigned to each computational cell, and the input feature vector is linearly transformed using a mapping matrix to generate the original initial state value. Based on the original initial state values, a set of connection parameters between computational cells is defined. The set of connection parameters includes synaptic connection weight parameters and a plasticity memory factor. The synaptic connection weight refers to the signal transmission strength between computational cells, and the plasticity memory factor refers to the influence of historical states on the current synaptic strength. Using a differentiable synaptic plasticity mechanism, the set of connection parameters is dynamically updated according to the activity correlation between computational cells. The update process is that the synaptic connection between each pair of computational cells decays historical state information through a forgetting factor, and the weight is updated by accumulating based on the current activity strength of the computational cell. Based on the updated set of connection parameters, neighborhood information is integrated into the initial state vector of each computational cell to generate the initial local performance representation of the computational cell.

5. The method for optimizing intelligent hardware configuration based on deep learning according to claim 1, characterized in that, The construction of the dynamic topology specifically includes: The initial local performance characterization is input into the dynamic topology adaptive module of the improved NCGM model to establish the computational cell state matrix. Based on the computational cell state matrix, the connection strength is calculated using an exponential decay function to generate the initial connection matrix. The initial connection matrix is ​​dynamically topologically evolved in the continuous time dimension. The state vector of the computation cell state matrix is ​​dynamically updated by the state evolution weight matrix and nonlinear activation function. The connection strength in the initial connection matrix is ​​adaptively adjusted according to the feedback results of the performance objective function and the regularization constraint. During the dynamic topology evolution process, the initial connection matrix is ​​normalized, and the proportion of connection weights of each computational cell is calculated to obtain the distribution result of the normalized connection weights. The normalization process refers to calculating the sum of the connection weights of each computational cell and scaling it proportionally. Based on the distribution of normalized connection weights, the neighborhood set of each computational cell at the current time is obtained. When the connection strength is greater than a preset threshold, the corresponding computational cell is included in the neighborhood set. During the continuous time evolution process, the neighborhood sets at all times are recorded and averaged to generate a neighborhood evolution feature set. Based on the neighborhood evolution feature set, the stability of the dynamic topology is determined. When the rate of change of the neighborhood set is less than a set threshold, the final normalized connection matrix and the corresponding neighborhood set are output, forming a dynamic topology.

6. The method for optimizing intelligent hardware configuration based on deep learning according to claim 1, characterized in that, The generation of the global performance representation vector specifically includes: Based on the connection weight distribution and state activation intensity of each computational cell in the dynamic topology, the average connection influence factor of each computational cell is calculated, and a time constant vector is generated. The time constant vector refers to the rate of state change of the computational cell in the continuous time dimension. Each computational cell corresponds to a time constant parameter, and all time constant parameters constitute the time constant vector. An adaptive time constant method is used to dynamically adjust the time constant vector to generate an adjusted time constant vector. The adjusted time constant vector is generated by calculating the rate of change of the time constant by the current state intensity of each computation cell and the time decay coefficient, so that the time constant increases when the state intensity of the computation cell is large and gradually decreases when the state tends to be stable, thus obtaining the adjusted time constant vector. Based on the adjusted time constant vector, the state of each computation cell is updated to obtain the updated computation cell. The updated computation cell is obtained by weighted summation of the current state of the computation cell and the states of the neighboring computation cells according to the normalized connection matrix, and then linear combination of the self-state mapping matrix, the neighboring state mapping matrix and the bias vector. The updated computational cells are expanded into matrix form, and the state matrices of all computational cells are updated synchronously. Under the constraint mechanism of the time constant vector, the state matrix of the computational cells at the global time endpoint is obtained by integrating the continuous time changes of the state of each computational cell. The column vectors of the computational cell state matrix at the global time endpoint are aggregated to obtain an aggregated vector. The aggregated vector is then linearly mapped using the readout matrix and readout bias vector to generate a global performance characterization vector.

7. The method for optimizing intelligent hardware configuration based on deep learning according to claim 1, characterized in that, The output of the convergence parameter set specifically includes: The global performance representation vector and the target performance index are input into the supervised learning and parameter optimization module of the improved NCGM model to establish a mapping model between the global performance representation vector and the target performance index. Based on the mapping model, the global performance representation vector and the target performance index are mapped to the Laplace frequency domain to obtain the corresponding frequency domain representation. By calculating the difference between the global performance representation vector and the target performance index in the Laplace frequency domain, a frequency domain error function is formed. A supervised learning loss function is constructed by jointly using the frequency domain error function with the difference between the global performance representation vector and the target performance index. The supervised learning loss function is a weighted sum of the frequency domain error constraint term and the time domain error constraint term through a joint time and frequency balance coefficient. Using the supervised learning loss function as the optimization objective, a multi-domain collaborative adaptive optimization mechanism is adopted to minimize the parameter set of the mapping model. The gradient of the supervised learning loss function relative to the parameter set of the mapping model is calculated using the frequency domain backpropagation algorithm, and then mapped back to the time domain to perform parameter updates, resulting in the updated parameter set. The updated parameter set is re-input into the improved NCGM model, and the Laplacian frequency domain mapping, supervised learning loss function calculation and parameter update process are repeated until the convergence condition is met, and the converged parameter set is output.

8. The method for optimizing intelligent hardware configuration based on deep learning according to claim 1, characterized in that, The output of the hardware configuration parameter set specifically includes: The NCGM model is improved based on the configuration of the convergence parameter set. Real-time collected system operation data is input into the improved NCGM model to generate a predictive performance output vector. The difference between the predicted performance output vector and the target performance index is calculated to obtain the prediction error vector, which refers to the degree of deviation between the output of the improved NCGM model and the target performance index. Based on the supervised learning loss function, the parameter set of the improved NCGM model is optimized in a supervised manner according to the prediction error vector to obtain the optimized and updated parameter set. The optimized and updated parameter set is obtained by taking the sum of squares of the prediction error vector as the prediction error term in the loss function, and introducing the squared difference between the current parameter set and the convergent parameter set as the parameter constraint term. Based on the optimized and updated parameter set, a topology self-evolution mechanism is enabled. The connection strength between computational cells is dynamically adjusted according to the updated weight distribution and node activation intensity to obtain the updated normalized connection matrix. The updated normalized connection matrix is ​​input into the improved NCGM model for performance evaluation. When the prediction error vector is less than a preset threshold, a set of hardware configuration parameters is output. The set of hardware configuration parameters includes computing resource allocation parameters, storage and cache configuration parameters, communication and bus control parameters, energy consumption control parameters, and topology mapping and structure scheduling parameters.

Images