Self-adaptive control method for operation parameters of artificial intelligence module

Abstract

The invention discloses an artificial intelligence module operation parameter adaptive control method, and relates to the technical field of parameter control, and the method comprises the steps: predicting future operation index values of an artificial intelligence module under different operation parameter adjustment variables based on a state representation vector and a preliminary control vector, and taking the preliminary control vector as an initial point, the method comprises the following steps of: constructing a multi-objective optimization problem by using a multi-objective optimization algorithm and using a Pareto solution set of the multi-objective optimization problem to perform real-time search and obtain a final control vector, and issuing the final control vector to an execution mechanism of an artificial intelligence module to perform physical adjustment; according to the method, collaborative optimization of energy efficiency and performance is realized by searching the optimal tradeoff control scheme, and the core problem that safety and efficiency are difficult to consider under multi-target conflicts, working condition switching and abnormal working conditions is solved.

Description

Technical Field

[0001] This invention relates to the field of parameter control technology, and in particular to an adaptive control method for the operating parameters of an artificial intelligence module. Background Technology

[0002] With the rapid development of artificial intelligence technology, AI modules are being deployed more and more widely in data centers, edge computing devices, and dedicated acceleration platforms. To meet the stringent requirements of diverse task loads for energy efficiency, latency, and thermal management, dynamic adjustment of operating parameters has become a key link in ensuring module performance and reliability. In recent years, parameter adjustment methods based on feedback control, reinforcement learning, and heuristic strategies have been proposed one after another. Some solutions attempt to introduce runtime monitoring data to achieve a certain degree of adaptability. This not only promotes the evolution of AI modules towards a higher level of autonomy and intelligence, but also lays an important foundation for building an efficient and reliable artificial intelligence operating environment.

[0003] Nevertheless, existing intelligent module operation parameter control methods still have room for improvement. First, they fail to effectively model the spatiotemporal coupling relationship between components, especially neglecting the semantic information of key events contained in logs, which limits the ability to identify operating conditions. Second, the control strategy lacks a conditional mechanism and cannot dynamically adjust the decision logic according to the operating conditions, resulting in weak strategy generalization ability and easy performance fluctuations under new load modes. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides an adaptive control method for the operating parameters of an artificial intelligence module to solve the problems of limited operating condition recognition capability and weak strategy generalization capability.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] This invention provides an adaptive control method for the operating parameters of an artificial intelligence module, comprising:

[0008] Real-time acquisition of multimodal feedback data from artificial intelligence modules to construct spatiotemporal state models;

[0009] Extracting a state representation vector that represents the current operating state of the artificial intelligence module based on a spatiotemporal state model;

[0010] Based on state representation vectors and multimodal feedback data, the current operating condition is identified, and an operating condition feature vector is generated.

[0011] Adaptive control law calculations are performed on the state representation vector and the operating condition feature vector to obtain a preliminary control vector for adjusting the operating parameters of the artificial intelligence module;

[0012] Based on the state representation vector and the preliminary control vector, predict the future operating index values ​​of the artificial intelligence module under different operating parameter adjustments;

[0013] Starting with the initial control vector and optimizing future performance indicators, a multi-objective optimization problem is constructed.

[0014] A multi-objective optimization algorithm is used to perform a real-time search on the Pareto solution set of the multi-objective optimization problem to obtain the final control vector.

[0015] The final control vector is sent to the actuator of the artificial intelligence module for physical adjustment.

[0016] As a preferred embodiment of the adaptive control method for the operating parameters of the artificial intelligence module described in this invention, the multimodal feedback data includes time-series monitoring data, event recording data, and topology data describing the connection relationships between components.

[0017] As a preferred embodiment of the adaptive control method for the operating parameters of the artificial intelligence module described in this invention, the construction of the spatiotemporal state model specifically includes:

[0018] The basic graph structure is constructed by using the hardware components of the artificial intelligence module as physical nodes and the connection relationships between the hardware components represented by topological data as edges.

[0019] By using time-series monitoring data as the time-series attribute sequence of corresponding physical nodes, natural language processing is performed on event record data to extract key events. These key events are then integrated as attribute information into the basic graph structure to form a spatiotemporal state model.

[0020] As a preferred embodiment of the adaptive control method for the operating parameters of the artificial intelligence module described in this invention, the extraction of the state representation vector representing the current operating state of the artificial intelligence module refers to performing correlation analysis on the physical nodes in the spatiotemporal state model and extracting time evolution features by combining time series attributes to obtain the state representation vector.

[0021] As a preferred embodiment of the adaptive control method for the operating parameters of the artificial intelligence module described in this invention, the step of identifying the current operating condition and generating an operating condition feature vector based on the state representation vector and multimodal feedback data specifically includes:

[0022] Statistical analysis is performed on the state representation vector to extract high-order statistical feature vectors. At the same time, trend analysis is performed on the time series monitoring data in the multimodal feedback data to extract time series pattern feature vectors.

[0023] The high-order statistical feature vector is concatenated with the time-series pattern feature vector to form the original working condition description feature vector;

[0024] Principal component analysis is used to perform linear transformation and dimensionality reduction on the original working condition description feature vector, and the first k principal components are selected as the basic components.

[0025] The basic components are matched with the template features in the working condition template library based on similarity. The working condition template with the highest matching degree is selected and the corresponding template code is extracted.

[0026] The template encoding and the basic components are weighted and fused to generate the working condition feature vector.

[0027] As a preferred embodiment of the adaptive control method for the operating parameters of the artificial intelligence module described in this invention, the step of calculating the adaptive control law on the state representation vector and the operating condition feature vector to obtain the preliminary control vector for adjusting the operating parameters of the artificial intelligence module specifically involves:

[0028] Map the state representation vector and the operating condition feature vector to control law modulation parameters;

[0029] By using the modulation parameters of the control law to perform feature modulation on the state representation vector, conditional state features are generated.

[0030] The control strategy is calculated based on the conditional state characteristics to obtain the preliminary control vector.

[0031] As a preferred embodiment of the adaptive control method for the operating parameters of the artificial intelligence module described in this invention, the step of predicting the future operating index values ​​of the artificial intelligence module under different operating parameter adjustment amounts based on the state representation vector and the preliminary control vector specifically includes:

[0032] The state representation vector is concatenated with the preliminary control vector to form a joint feature vector;

[0033] Perform a high-order polynomial expansion on the joint feature vectors to generate a nonlinear feature combination;

[0034] Using a nonlinear feature combination vector as input, a set of predefined index prediction functions are applied to calculate future operating index values.

[0035] The future operating performance indicators include power consumption, latency, and temperature.

[0036] As a preferred embodiment of the adaptive control method for the operating parameters of the artificial intelligence module described in this invention, the construction of the multi-objective optimization problem specifically includes:

[0037] The initial control vector is used as the starting point for the optimization search, and the power consumption, delay, and temperature parameters are defined as objective functions to be minimized.

[0038] Set corresponding constraint boundaries for each objective function;

[0039] The constraints include power consumption upper limit constraints, delay threshold constraints, and temperature safety constraints; integrating the objective function and the constraints forms a multi-objective optimization problem.

[0040] As a preferred embodiment of the adaptive control method for the operating parameters of the artificial intelligence module described in this invention, the method employs a multi-objective optimization algorithm to perform a real-time search within the Pareto solution set of the multi-objective optimization problem to obtain the final control vector. Specifically:

[0041] A decomposition-based multi-objective evolutionary algorithm is used to decompose the multi-objective optimization problem into multiple scalar quantum problems, and a population of candidate solutions is maintained in the neighborhood of each scalar quantum problem.

[0042] After initializing the candidate solution population with the initial control vector, crossover and mutation operations are performed to generate new solutions;

[0043] The power consumption, delay, and temperature values ​​corresponding to each new solution are calculated using an evaluation function. The candidate solution population is then updated based on the dominance relationship and aggregation function. Crossover, mutation, evaluation, and update operations are performed iteratively until the termination condition is met, and the final candidate solution population is obtained.

[0044] After performing non-dominated sorting on the final candidate solution population, an initial control vector that satisfies all constraints is selected as the final control vector at the Pareto front.

[0045] As a preferred embodiment of the adaptive control method for the operating parameters of the artificial intelligence module described in this invention, the step of sending the final control vector to the actuator of the artificial intelligence module for physical adjustment specifically includes:

[0046] The final control vector is parsed into a set of adjustable parameter instructions for the artificial intelligence module;

[0047] The adjustable parameter instruction set is sent sequentially to the actuator of the artificial intelligence module through the hardware communication interface for parameter adjustment.

[0048] The beneficial effects of this invention are as follows: By constructing a spatiotemporal state model and extracting high-order state representations, a deep perception of the operating state of the AI ​​module is achieved. On this basis, conditional calculations are performed on the state representation vector and the operating condition feature vector, enabling the control logic to dynamically adapt to different task requirements. Combined with a multi-objective Pareto optimization algorithm, the optimal trade-off control scheme is searched in real time under the premise of strictly meeting hard constraints such as power consumption, delay, and temperature. This not only improves the accuracy, stability, and adaptability of the control strategy, but also effectively achieves the synergistic optimization of energy efficiency and performance, solving the core problem of difficulty in balancing safety and efficiency under current multi-objective conflicts, operating condition switching, and abnormal operating conditions. Attached Figure Description

[0049] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0050] Figure 1 A flowchart of an adaptive control method for the operating parameters of an artificial intelligence module.

[0051] Figure 2 A flowchart for obtaining the feature vector of the working condition.

[0052] Figure 3 A flowchart for obtaining a spatiotemporal state model.

[0053] Figure 4 To obtain a flowchart for a multi-objective optimization problem. Detailed Implementation

[0054] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0055] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0056] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0057] Reference Figures 1-4 As one embodiment of the present invention, this embodiment provides an adaptive control method for the operating parameters of an artificial intelligence module, comprising the following steps:

[0058] S1. Collect multimodal feedback data from the artificial intelligence module in real time to construct a spatiotemporal state model.

[0059] S1.1 It should be noted that data acquisition probes are deployed on the operating kernel and hardware performance counters of the artificial intelligence module; the probes are configured to continuously collect timing monitoring data; the timing monitoring data includes CPU utilization percentage, memory usage in bytes, hardware accelerator core frequency and power consumption in watts; the event logging agent listens for real-time operation event service and application log file writes; when an event entry is generated, the event logging agent immediately captures and parses the event logging data; the event logging data includes error level, event timestamp and descriptive text; topology data is obtained by reading the device tree description file of the artificial intelligence module; the topology data describes the physical connection relationship and address mapping between the CPU core, memory channels, hardware accelerator and peripheral component interconnect high-speed bus; a shared data buffer is created in memory; the timing monitoring data, event logging data and topology data are aligned with the timestamp of the global synchronization clock and written to the shared data buffer.

[0060] S1.2 It should be noted that, based on the physical connection relationships described in the topology data, each hardware component instance of the artificial intelligence module is defined as a physical node; the hardware components include the central processing unit core, memory controller, hardware accelerator engine, and high-speed bus ports for interconnecting peripheral components; if the topology data indicates that there is a direct physical link or bus connection relationship between two hardware components, then an undirected edge is created between the physical nodes representing these two hardware components, and all hardware components and their corresponding connection relationships are traversed in turn to generate a basic graph structure composed of a set of nodes and a set of edges.

[0061] The system loads a predefined grammar rule base and entity dictionary to process event log data. The grammar rule base includes syntactic patterns for identifying subject-verb-object structures, while the entity dictionary stores a set of known hardware component names and error codes. Event log data is segmented into sentences. For each sentence, the grammar rule base is applied for pattern matching to identify the core verb as the action type, and noun phrases with a subject-verb relationship to the verb are extracted as event subject candidates. The words in the sentences are compared with the entity dictionary, and words belonging to hardware component names are marked as named entities. The noun phrases of the event subject candidates extracted from the grammar rule base are matched against the marked hardware component names. If a noun phrase of an event subject is completely identical to any hardware component name at the string level, it is determined that the noun phrase of the event subject candidate is identical to the hardware component name. The noun phrases of the event subject candidate that are determined to be identical are formally confirmed as the event subjects of the critical events. The severity level field is read from the event log data entries. This field includes standard severity level terms such as "DEBUG," "INFO," "WARNING," "ERROR," and "CRITICAL." These severity level terms are mapped to corresponding severity values: "DEBUG" is mapped to value 1, "INFO" to value 2, "WARNING" to value 3, "ERROR" to value 4, and "CRITICAL" to value 5. The mapped severity values, the identified event subject, and the action type are combined and encapsulated into a critical event data object, completing the construction of the critical event. The event subject identified in the critical event is matched against the physical node. If the event subject string matches the physical node's identifier string, the match is successful, and the critical event is associated with the corresponding physical node. The text summary and timestamp of the critical event are added as dynamic attributes to the attribute information of the corresponding physical node. The basic graph structure, which integrates temporal attribute sequences and dynamic attributes, constitutes the spatiotemporal state model.

[0062] Furthermore, the specific construction process of the grammar rule base is as follows: Historical event records from the AI ​​module are collected as analysis corpus. The analysis corpus is reviewed and annotated to identify common sentence structures and core verbs used to describe hardware status, error events, and operational behaviors. Based on the annotation results, fixed syntactic patterns describing specific event types are summarized, such as "<hardware component> + <report / trigger / occurrence> + <error type>". These syntactic patterns are formalized into executable rule entries, each rule entry including a trigger part-of-speech sequence pattern, an action type extraction location, and an event subject extraction location. The rule entries are organized into a structured set and stored to form the grammar rule base.

[0063] The specific process of constructing the entity dictionary is as follows: Official hardware specification documents, device driver manuals, and historical fault reports of the AI ​​module are collected. The official names, aliases, and model codes of all hardware components are extracted from the hardware specification documents and device driver manuals to form a hardware component name set. All recorded error codes and standard status description terms are extracted from historical fault reports to form an error code set and a status identifier set. All entries from the hardware component name set, error code set, and status identifier set are stored in a structured file according to their corresponding categories (hardware component name category, error code category, and status identifier category) to form the entity dictionary.

[0064] S2. Extract the state representation vector representing the current operating state of the artificial intelligence module based on the spatiotemporal state model.

[0065] It should be noted that a historical spatiotemporal state model is constructed (in the same way as obtaining the spatiotemporal state model in step S1.2) and the corresponding state evaluation values ​​are labeled. The state evaluation values ​​are used to characterize the overall operational status of the AI ​​module at a historical moment. The state evaluation values ​​are represented in a single scalar form, with a value range of 0 to 1. The closer the state evaluation value is to 1, the more stable the operation of the AI ​​module is at the corresponding historical moment, the higher the operating efficiency, and the lower the risk. The closer the state evaluation value is to 0, the higher the risk of performance degradation, resource consumption, or thermal instability at the corresponding historical moment. The state evaluation values ​​are calculated based on the actual operating indicators within the time window corresponding to the historical spatiotemporal state model. The actual operating indicators include task processing latency, throughput per unit time, power consumption, and temperature. After normalizing the actual operating indicators, a weighted sum is obtained to obtain the state evaluation value.

[0066] Read the physical node state information from the historical spatiotemporal state model and pair it with the corresponding state evaluation value to form a training set; the physical node state information includes the operating status data of each hardware component at the current moment, such as CPU utilization, hardware accelerator core frequency, memory usage, power consumption value, and component operation log events.

[0067] A spatiotemporal state modeling model is constructed, which has a three-layer structure. Each layer includes a spatial correlation analysis layer, a temporal evolution analysis layer, and a spatiotemporal feature fusion layer. Specifically, the attention weight vector of the spatial correlation analysis layer has a dimension of 128×1, and the linear transformation matrix has a dimension of 64×64. The temporal evolution analysis layer has 64 causal convolution kernels, each with a time length of 3 and a parameter size of 3×64. The gating activation unit weight matrix has a parameter size of 64×64, and the gating activation unit bias parameter has a dimension of 64. The linear transformation matrix of the spatiotemporal feature fusion layer has a dimension of 128×64, and the bias parameter has a dimension of 64.

[0068] The parameter settings for each layer are based on the physical node input feature dimension and the state representation vector target dimension as a unified benchmark, so that the output dimensions of the spatial correlation analysis layer, the temporal evolution analysis layer and the spatiotemporal feature fusion layer are consistent, thereby ensuring the stability of the multi-layer calculation process.

[0069] The training set is input into the spatiotemporal state modeling model and forward propagation is performed to obtain the predicted output. The loss value between the predicted output and the state evaluation value is calculated. The gradient of the loss value with respect to all trainable parameters in the spatiotemporal state modeling model is calculated using the backpropagation algorithm. The parameters of the spatiotemporal state modeling model are updated according to the gradient using the Adam optimizer. The process of forward propagation, loss calculation, backpropagation and parameter update is iteratively executed until the loss value is less than the convergence threshold, and the trained spatiotemporal state modeling model is obtained.

[0070] The convergence threshold is set to characterize whether the loss value of the spatiotemporal state model has entered a stable convergence phase during training. It is calculated by statistically analyzing the loss value sequence for 10 consecutive iterations, calculating the mean and standard deviation of the loss value sequence, and using the sum of the squares of the mean and standard deviation as the convergence threshold. Ten consecutive iterations can suppress random fluctuations caused by single or few iterations, while avoiding prolonged training due to delayed convergence determination caused by an excessively long statistical window.

[0071] The physical node state information in the spatiotemporal state model is input into the trained spatiotemporal state modeling model. In the spatial association analysis layer, all physical nodes and their first-order neighbor nodes in the spatiotemporal state model are traversed. The physical nodes are taken as target nodes. The feature vectors of the target node and its first-order neighbor nodes are concatenated and then dot-producted with the attention weight vector. The dot-product result is input into the LeakyReLU function to obtain the unnormalized association score. The unnormalized association score is normalized using Softmax to obtain the association weight. The feature vectors of the neighbor nodes are weighted and summed according to the association weight to obtain the spatial association feature vector of the physical node.

[0072] The temporal attribute sequence carried by the physical node is read and input into the temporal evolution analysis layer. The temporal evolution analysis layer uses causal convolution kernels to perform convolution calculations on the temporal attribute sequence to generate convolution output feature vectors. These convolution output feature vectors are stacked in chronological order to form an intermediate feature tensor. The intermediate feature tensor is input into a gating activation unit, which performs a linear mapping on the feature dimension and splits it into two components (a gated signal computation component and a candidate feature component). A sigmoid function is applied to the gated signal computation component to generate a gated signal vector. The gated signal vector is then multiplied element-wise with the candidate feature component to obtain the temporal evolution feature vector. The gated signal vector is a weight vector used to retain, suppress, or adjust the elements in the candidate feature component.

[0073] Spatial correlation feature vectors and temporal evolution feature vectors are input into the spatiotemporal feature fusion layer. In this layer, the spatial correlation feature vectors and temporal evolution feature vectors are added element-wise, multiplied by a linear transformation matrix, and a bias parameter is added. The ReLU activation function is applied to the calculation results to obtain the hidden feature vectors of the physical nodes. These hidden feature vectors are then used as input to the next layer, and the spatial correlation analysis, temporal evolution analysis, and spatiotemporal feature fusion process are repeated until all layers are calculated. The hidden feature vectors obtained from all physical nodes in the last layer are stacked to form a two-dimensional feature matrix. The arithmetic mean of each column of the two-dimensional feature matrix is ​​calculated, and the results are arranged sequentially to form a dense vector of fixed dimensions, which is the state representation vector.

[0074] S3. Based on the state representation vector and multimodal feedback data, identify the current operating condition and generate an operating condition feature vector.

[0075] S3.1 It should be noted that after calculating the mean, variance, skewness, and kurtosis of the state representation vector, it is arranged into a higher-order statistical feature vector with a length of 4. The skewness formula is:

[0076] ;

[0077] in, Indicates skewness, The dimension of the state representation vector is represented by . The variance of the state representation vector is represented. This represents the mean of the state representation vector. The first element in the state representation vector represents the state representation vector. Each element value.

[0078] Furthermore, Represents the variance of the state representation vector The power, i.e. the cube of the standard deviation, is used to normalize the third central moment to eliminate the influence of numerical scale and dimensions on the skewness calculation results, making the skewness a dimensionless statistic. This indicates that the first element in the state representation vector is... The difference between each element value and the mean of the state representation vector is calculated and then raised to the power of 3. This is used to enhance the ability to characterize the direction and magnitude of deviation, thereby reflecting the contribution of each element to the asymmetry of the overall distribution.

[0079] S3.2 It should be noted that the kurtosis formula is:

[0080] ;

[0081] in, Indicates kurtosis;

[0082] Furthermore, subtracting a constant of 3 from the kurtosis calculation is to obtain excess kurtosis; by introducing excess kurtosis, the degree of deviation of the state characterization vector distribution from the standard normal distribution can be intuitively reflected, which facilitates a unified comparison of the distribution characteristics under different state conditions. This indicates that the first element in the state representation vector is... After calculating the difference between each element value and the mean of the state representation vector, take... The power of kurtosis, compared to the cube in skewness calculation, focuses more on the magnitude of the deviation rather than the direction of deviation. This aligns with the statistical definition of kurtosis as a measure of the "steepness" of a distribution, thus reflecting the sharpness and extreme value characteristics of the distribution.

[0083] S3.3. It should be noted that the time-series monitoring data is loaded as a matrix of shape [T, M], containing T time steps and M indicators. For each column of the indicator sequence in the time-series monitoring data, i.e., each individual indicator, a first-order linear line is fitted using the least squares method to obtain the slope value of the first-order linear line. The slope value is used as the trend strength value of the indicator sequence, which is expressed by the formula:

[0084] ;

[0085] in, Indicates the slope value. This represents the total number of time steps in the time-series monitoring data. Indicates the first The values ​​of the time independent variables corresponding to each time step Indicates the first The observations at each time step on the index sequence This represents the average of the observations in the index sequence at each time step. It represents the average value of the time independent variable at each time step.

[0086] Simultaneously, for the same indicator sequence, the standard deviation of data points within the most recent 9 time steps is calculated, and this standard deviation is used as the volatility intensity value of the indicator sequence. The trend intensity value and the volatility intensity value are concatenated to form a time series pattern feature vector.

[0087] Furthermore, the time window length corresponding to the most recent 9 time steps can cover the main change range of short-term load fluctuations and thermal inertia response of the artificial intelligence module, while limiting the historical span of standard deviation calculation to avoid dilution of the fluctuation intensity value by the long-term trend.

[0088] S3.4 It should be noted that the higher-order statistical feature vector and the time-series pattern feature vector are concatenated to form the original operating condition description feature vector; all original operating condition description feature vectors are collected and constructed into an original operating condition description feature matrix, where each row represents an original operating condition description feature vector; the arithmetic mean of each column of the original operating condition description feature matrix is ​​calculated to obtain a mean vector (the dimension of the mean vector is the same as the dimension of the original operating condition description feature vector); for each original operating condition description feature vector in the original operating condition description feature matrix, a vector subtraction operation is performed, that is, the difference between each element of the original operating condition description feature vector and the corresponding element in the mean vector is calculated. After the subtraction operation is completed, each vector in the original operating condition description feature matrix becomes a centered original operating condition description feature vector.

[0089] Calculate the covariance matrix of the original working condition description eigenvectors after centering; perform eigenvalue decomposition on the covariance matrix to obtain eigenvalues ​​and corresponding eigenvectors, which are the principal component axes; arrange the eigenvalues ​​in descending order and select the eigenvectors corresponding to the k largest eigenvalues ​​as the projection matrix; multiply the projection matrix with the original working condition description eigenvectors after centering to obtain a k-dimensional vector, which is the basic component.

[0090] The variable 'k' is set to reduce feature redundancy and computational complexity while preserving the main information of the original working condition description features. After arranging the feature values ​​in descending order, the cumulative variance contribution rate is calculated for each value. Feature values ​​with a cumulative variance contribution rate greater than or equal to 0.9 are selected, and their number is counted to obtain the value of 'k'. When the cumulative contribution rate of the feature values ​​reaches 0.9, it covers the "main variation pattern" of the data, effectively suppressing low-contribution noise components while preserving the main variance information and reducing computational complexity, thereby improving the stability and efficiency of feature representation. When the cumulative contribution rate of the feature values ​​is less than 0.9, although it can further reduce feature dimensionality and computational load, it can lead to the loss of the main information of the original working condition description features, increasing the risk of working condition identification errors. A predefined working condition template library is loaded. The working condition template library contains multiple working condition templates, each consisting of a template feature vector and a corresponding template code. The cosine similarity value between the basic components and all template feature vectors is calculated, expressed by the formula:

[0091] ;

[0092] in, Represents the cosine similarity value. Represents the fundamental components. This represents the template feature vector.

[0093] Compare all the calculated cosine similarity values, select the work condition template corresponding to the maximum cosine similarity value, and extract the corresponding template code from the work condition template (the template code is a fixed-dimensional category embedding vector).

[0094] After concatenating the template encoding with the basic components, the vectors are input into a pre-defined linear transformation layer. The linear transformation layer multiplies the concatenated vector with the weight matrix and adds a bias vector to obtain the linear transformation result. This result is then input into the LeakyReLU function for calculation to obtain the working condition feature vector. The sum of the number of principal components (k) and the dimension of the template encoding is multiplied by the dimension of the working condition feature vector to obtain the weight matrix; the bias vector is a zero vector.

[0095] Furthermore, the specific construction process of the operating condition template library is as follows: A large number of historical operating condition description feature vectors are collected (using the same method as step S3.4), forming a historical operating condition description feature vector set; K-Means clustering is performed on the historical operating condition description feature vector set, dividing it into K clusters, each representing a typical operating condition pattern; the centroid vector of all operating condition description feature vectors within each cluster is calculated, and this centroid vector is defined as the template feature vector of the operating condition template; a unique identifier is assigned to each cluster, and this identifier is encoded as a vector, serving as the template encoding for the operating condition template. The K operating condition templates, composed of template feature vectors and template encodings, are stored to obtain the operating condition template library. The template feature vectors represent the typical distribution characteristics of the corresponding operating condition template in the dimensionality-reduced feature space, and the template encoding uniquely identifies the corresponding operating condition template.

[0096] S4. Perform adaptive control law calculation on the state representation vector and operating condition feature vector to obtain the preliminary control vector for adjusting the operating parameters of the artificial intelligence module.

[0097] S4.1 It should be noted that a control strategy calculation model is constructed. The control strategy calculation model is a multi-layer parameterized calculation structure, including a conditional coding layer and a control strategy calculation layer. The linear transformation matrix parameter dimension of the first conditional coding layer is 64×32, and the bias parameter dimension is 64. The linear transformation matrix parameter dimension of the second conditional coding layer is 128×64, and the bias parameter dimension is 128 (the second conditional coding layer outputs 128-dimensional control law modulation parameters, where the first 64 dimensions are used as a scaling factor vector, and the last 64 dimensions are used as a bias factor vector). The first control... The linear transformation matrix parameter dimension of the strategy calculation layer is 128×64, and the bias parameter dimension is 128; the linear transformation matrix parameter dimension of the second control strategy calculation layer is 64×128, and the bias parameter dimension is 64; the linear transformation matrix parameter dimension of the third control strategy calculation layer is 4×64, and the bias parameter dimension is 4; historical state representation vectors and historical operating condition feature vectors are collected; within the time neighborhood corresponding to the historical state representation vectors and historical operating condition feature vectors, multiple adjustable parameters actually executed by the artificial intelligence module and recorded in the logs are collected as supervision labels. The historical state representation vector, historical operating condition feature vector, and supervision labels are used as the training set for the control policy calculation model. The training set is input into the control policy calculation model for forward prediction calculation, and the predicted control vector is output. The loss value between the predicted control vector and the supervision label is calculated. The gradient of the loss value with respect to all trainable parameters in the control policy calculation model is calculated using the backpropagation algorithm. The Adam optimizer is used to update the trainable parameters in the control policy calculation model according to the calculated gradient. The process of forward prediction, loss calculation, backpropagation, and parameter update is iteratively executed until the loss value is less than the stability threshold, and the trained control policy calculation model is obtained.

[0098] The stability threshold is set to determine whether the loss value of the control strategy computational model has entered a stable convergence phase during training. It is calculated by taking the median and interquartile range of the loss values ​​over 15 consecutive iterations, and using the sum of the median and interquartile range as the stability threshold. 15 consecutive iterations cover a longer range of short-term fluctuations in the loss value and improve the robustness of the median and interquartile range statistics, thereby reducing the impact of abnormal fluctuations in a single iteration on the stability threshold, while keeping the time window required for determining the stability threshold within an acceptable range.

[0099] S4.2 It should be noted that the state representation vector and the operating condition feature vector are input into the trained control strategy calculation model in parallel. The conditional encoding layer of the control strategy calculation model multiplies the operating condition feature vector with the weight matrix and superimposes a bias vector to obtain the linearly transformed operating condition feature vector. The linearly transformed operating condition feature vector is input into the ReLU function for calculation to obtain the control law modulation parameters (scaling factor vector and bias factor vector). Each element of the state representation vector is multiplied with the corresponding element of the scaling factor vector, and the product result is added to the corresponding element of the bias factor vector to generate the conditional state features. The conditional state features are input into the control strategy calculation model. The control strategy calculation layer consists of three sequentially connected linear transformation layers. Each linear transformation layer performs a linear transformation on the input features and applies the ReLU activation function. Specifically, the first linear transformation layer maps the conditional state features to 128-dimensional hidden features, the second linear transformation layer maps the 128-dimensional hidden features to 64-dimensional hidden features, and the third linear transformation layer maps the 64-dimensional hidden features to output features. The Tanh activation function is applied to the output features to obtain the preliminary control vector. Each dimension of the initial control vector corresponds one-to-one with each adjustable parameter of the artificial intelligence module; by consulting the hardware register manual, the upper bound and lower bound (i.e., the maximum and minimum values) of each adjustable parameter can be determined.

[0100] S5. Based on the state representation vector and the preliminary control vector, predict the future operating index values ​​of the artificial intelligence module under different operating parameter adjustments.

[0101] The state representation vector and the initial control vector are concatenated end-to-end to form a joint feature vector; each element in the joint feature vector is considered a first-order basis. The product term of two distinct elements in the joint feature vector is calculated as a second-order interactive feature. Simultaneously, the square term of each element in the joint feature vector is calculated as a second-order polynomial feature. All first-order bases, second-order interactive features, and second-order polynomial features are concatenated to form a nonlinear feature combination vector.

[0102] Using a nonlinear feature combination vector as input, a set of predefined index prediction functions are applied to calculate future operating index values. Specifically, the power consumption index prediction function multiplies the voltage square term, voltage-frequency product term, and constant term in the nonlinear feature combination vector with their corresponding physical coefficients and then sums the results to output the power consumption index value. The delay index prediction function quotients the load characterization term and processing capacity characterization term in the nonlinear feature combination vector and outputs the delay index value. The temperature recursive function substitutes the temperature index value from the previous time step, the current power consumption index value, and the heat dissipation characterization term in the nonlinear feature combination vector into a predefined thermodynamic difference equation to calculate the temperature index value for the next time step. By sequentially applying the power consumption index prediction function, the delay index prediction function, and iteratively calling the temperature recursive function, the power consumption index value, delay index value, and temperature index value for future time steps are generated, i.e., the future operating index values.

[0103] Furthermore, the process of obtaining the temperature index value of the previous moment is as follows: the temperature reading at the beginning of the current control cycle is collected by the temperature sensor of the artificial intelligence module, and the temperature reading is used as the temperature index value of the previous moment.

[0104] The process of obtaining the physical coefficients for the voltage square term, voltage-frequency product term, and constant term is as follows: Based on the timing power consumption model file of the standard cell library provided by the AI ​​module designer or manufacturer, and the parasitic parameter file for a specific process corner, power consumption analysis is performed using electronic design automation tools. Specifically, the dynamic power consumption coefficient is obtained by summing the product of the output pin load capacitance and the switching activity of all standard cells; its physical essence is the total effective switching capacitance. The interactive power consumption coefficient is obtained by integrating the internal short-circuit current pulses of all standard cells during the input signal transition in the time domain and calculating the average power consumption, characterizing the additional power consumption generated by the transistor being simultaneously turned on during the switching transient. The static power consumption coefficient is obtained by multiplying the sum of the subthreshold leakage current and gate leakage current of all standard cells at the current operating voltage and junction temperature with the operating voltage. The calculated dynamic power consumption coefficient, interactive power consumption coefficient, and static power consumption coefficient are used as the physical coefficients for the voltage square term, voltage-frequency product term, and constant term, respectively.

[0105] S6. Using the initial control vector as the starting point and the future operating index value as the optimization objective, construct a multi-objective optimization problem.

[0106] It should be noted that the initial control vector is used as the initial solution for the multi-objective optimization search. The power consumption metric is defined as the first objective function, aiming to minimize total power consumption; the delay metric is defined as the second objective function, aiming to minimize average delay; and the temperature metric is defined as the third objective function, aiming to minimize peak temperature.

[0107] A power consumption upper limit constraint is set for the power consumption objective function, requiring that the power consumption index value corresponding to any candidate control vector generated during the optimization process must not exceed a preset power consumption wall threshold. A delay threshold constraint is set for the delay objective function, requiring that the delay index value corresponding to any candidate control vector generated during the optimization process must not exceed a preset service level agreement (SLA) threshold. A temperature safety constraint is set for the temperature objective function, requiring that the temperature index value corresponding to any candidate control vector generated during the optimization process must not exceed a preset hardware thermal protection threshold. The power consumption objective function, delay objective function, temperature objective function, power consumption upper limit constraint, delay threshold constraint, and temperature safety constraint are integrated and formally constructed into a multi-objective optimization problem containing three minimization objectives and three inequality constraints.

[0108] Furthermore, the power consumption wall threshold is set to prevent hardware damage, instability, and triggering of protective shutdown due to excessive power consumption; based on the maximum allowable power consumption specified in the AI ​​module technical specifications, 80%-90% of the maximum allowable power consumption is taken as the power consumption wall threshold.

[0109] Service Level Agreement (SLA) thresholds are set to ensure the performance and business continuity of upper-layer applications. They are based on the service performance targets required by the AI ​​module (clearly defined in the SLA signed by the user or business party, for example, stipulating that "95% of inference tasks will have a latency of no more than 100 milliseconds"). Therefore, the user-defined latency target value in the SLA can be used as the SLA threshold.

[0110] Hardware thermal protection thresholds are set to prevent transistor leakage current surges, timing errors, performance drops, and even permanent physical damage caused by overheating. Based on the maximum junction temperature parameter specified in the technical specifications provided by the AI ​​module chip manufacturer, a safety margin (e.g., 5-10°C) is set as the hardware thermal protection threshold. This safety margin can offset some measurement uncertainties, ensuring that even under the most unfavorable combination of errors, the actual junction temperature will not exceed the limit.

[0111] S7. After performing non-dominated sorting on the final candidate solution population, select an initial control vector that satisfies all constraints at the Pareto front as the final control vector.

[0112] S7.1 It should be noted that the DasandDennis method is used to generate a set of weight vectors. Specifically, a partitioning parameter (positive integer) is set. For the power consumption objective function, delay objective function, and temperature objective function, all combinations of three non-negative integers not exceeding the partitioning parameter are enumerated (the sum of these three non-negative integers must equal the partitioning parameter). Each such non-negative integer combination is divided by the partitioning parameter to obtain three decimals, the sum of which is one, forming a weight vector. All non-negative integer combinations satisfying the condition are traversed to generate corresponding weight vectors; these weight vectors are uniformly distributed in the tradeoff space formed by the three objective functions. Each weight vector aggregates the three objective functions in the multi-objective optimization problem into a scalar quantum problem; the initial control vector is used as the initial center vector of the candidate solution population, and random perturbations are applied within the range of parameter values ​​of the initial control vector to generate perturbed control vectors; the initial control vector and the perturbed control vectors constitute the candidate solution population; a candidate solution population is maintained for each scalar quantum problem, and each individual in the candidate solution population is a complete control vector.

[0113] S7.2. It should be noted that, from the candidate solution population of the scalar quantum problem, two random integers uniformly distributed within the population index range are generated using a pseudo-random number generator. Based on these two random integers, two different individuals are randomly selected without replacement from the candidate solution population as tournament candidates. The aggregation function value of the first tournament candidate on the current scalar quantum problem is calculated, which is derived from the weight vector of the scalar quantum problem and the objective function values ​​(power consumption objective function, delay objective function, and temperature objective function). The aggregation function value of the second tournament candidate on the current scalar quantum problem is calculated. The aggregation function values ​​of the two tournament candidates are compared, and the individual with the smaller aggregation function value is selected as the winner. The winner is determined as the parent solution (first parent solution) selected by the binary tournament selection method. An adjacent scalar quantum problem is randomly selected from the neighborhood list of the scalar quantum problem, and another parent solution (parent solution in vector form) is also selected from the candidate solution population of the adjacent scalar quantum problem using the binary tournament selection method.

[0114] For each dimension in the first and second parent solution vectors, a random number uniformly distributed in the interval [0,1] is generated using a pseudo-random number generator. If the random number of a dimension in the parent solution vector is greater than the crossover probability, the dimension value of the first parent solution is selected as the value of the intermediate child solution in the corresponding dimension; if the random number of a dimension in the parent solution vector is less than or equal to the crossover probability, a simulated binary crossover operation is performed on the dimension; the mixing coefficient is calculated according to the formula for simulated binary crossover, which is:

[0115] ;

[0116] in, Represents the mixing coefficient. This indicates a crossover random number (which needs to be regenerated using a pseudo-random number generator). This represents the preset cross-distribution index.

[0117] The value of the intermediate sub-solution in the corresponding dimension is calculated based on the mixing coefficient, using the following formula:

[0118] ;

[0119] in, Indicates the intermediate sub-solution at the th Dimension value, This indicates that the first parent solution is in the [missing information]. Dimension value of a dimension This indicates that the second parent solution is in the... The dimension value of a dimension.

[0120] After iterating through all dimensions of the two parent solutions in sequence, a complete intermediate child solution vector is generated.

[0121] Furthermore, the crossover probability is set to balance the inheritance of parent solutions and the generation of new solutions. It is calculated as the reciprocal of the total dimension of the initial control vector.

[0122] The cross-distribution index is set to adjust the degree of concentration of the child solution relative to the parent solution in the simulated binary cross-distribution; the total dimension of the initial control vector is used as the cross-distribution index.

[0123] S7.3 It should be noted that for each dimension of the intermediate sub-solution vector, a random number is generated using a pseudo-random number generator; the random number is compared with the preset mutation probability. If the random number is greater than the mutation probability, the value of the intermediate sub-solution in that dimension is directly assigned to the new solution. After traversing all dimensions, a complete new solution vector is generated. If the random number is less than or equal to the mutation probability, the perturbation value is calculated according to the formula for polynomial mutation, which is:

[0124] ;

[0125] in, Indicates the disturbance value. This indicates a perturbed random number (which needs to be regenerated using a pseudo-random number generator). This represents the variation distribution index.

[0126] The dimension value of the new solution in the corresponding dimension is calculated based on the perturbation value, using the following formula:

[0127] ;

[0128] in, The new solution is in the first... Dimension value of a dimension Indicates the first Upper bound of the dimension parameter, Indicates the first The lower bound of the dimension parameter.

[0129] After iterating through all dimensions of the intermediate child solutions, a complete new solution is generated.

[0130] The mutation probability is set to control the balance between search diversity and solution stability, and to prevent the candidate solution population from converging too early or being subject to invalid random perturbations. The mean and standard deviation of the upper and lower bounds of the parameters of the initial control vector are calculated, and the mean and standard deviation are summed and normalized to obtain the mutation probability.

[0131] The variation distribution index is set to adjust the degree of concentration of the probability distribution of the disturbance amplitude in the polynomial variation; similarly, the total dimension of the initial control vector is used as the variation distribution index.

[0132] S7.4 It should be noted that for each new solution generated through crossover and mutation operations, the aforementioned prediction process is re-executed based on the state representation vector and the new solution to calculate the future operating index values ​​corresponding to the new solution, namely, power consumption index value, delay index value, and temperature index value. These three calculated values ​​together constitute the objective function value vector of the new solution. The new solution and the objective function value vector are treated as a complete entity. All individuals in the candidate solution population of the current scalar quantum problem and its adjacent scalar quantum problems are compared with the new solution individual. Based on the Pareto dominance relationship, if the new solution individual dominates any existing individual in the candidate solution population, then the dominated existing individual is replaced by the new solution individual; if the new solution individual is dominated by any existing individual in the candidate solution population, then the new solution individual is discarded; if the new solution individual and the existing individuals in the candidate solution population do not dominate each other, then the values ​​of the new solution individual and the existing individuals under the aggregation function of the current scalar quantum problem are further calculated. If the aggregation function value of the new solution individual is better than that of a certain existing individual, then the corresponding existing individual is replaced. This update process ensures that the candidate solution population of each scalar quantum problem is continuously optimized in terms of dominance relationship and aggregation function value. Repeatedly performing crossover, mutation, evaluation, and update operations constitutes a complete iterative cycle. The iterative cycle continues until the candidate solution population no longer improves, at which point the iteration terminates; the candidate solution populations of all scalar quantum problems in the final generation are then combined into the final candidate solution population.

[0133] A fast non-dominated sort is performed on all individuals in the final candidate solution population. Pareto dominance is calculated based on the individuals' power consumption, delay, and temperature parameters, and all individuals are divided into multiple Pareto levels. Individuals not dominated by any other individual are assigned to the first Pareto front, those dominated only by the first Pareto front are assigned to the second Pareto front, and so on. Within the first Pareto front, if an individual's power consumption does not exceed the upper limit constraint, delay does not exceed the delay threshold constraint, and temperature does not exceed the temperature safety constraint, then the corresponding individual is marked as a feasible solution. From all individuals marked as feasible solutions, one individual is arbitrarily selected as the preferred solution. If no feasible solution exists in the first Pareto front, the next Pareto front is checked sequentially until at least one feasible solution satisfying all constraints is found and selected. For the finally selected individual, the corresponding control vector is extracted as the final control vector.

[0134] S8. Send the final control vector to the actuator of the artificial intelligence module for physical adjustment.

[0135] The final control vector is compared with a preset parameter mapping table, which defines the correspondence between the index of each dimension in the control vector and the address and bit field of the specific adjustable hardware register of the AI ​​module. Based on the parameter mapping table, the value of each dimension in the final control vector is converted into the format and value required by the target hardware register. Following the instruction format specified by the AI ​​module's underlying driver interface, the converted register address, opcode (write), and data value are encapsulated into a single machine instruction. All dimensions of the final control vector are traversed to generate multiple corresponding machine instructions. These machine instructions are then combined sequentially to form the set of adjustable parameter instructions for the AI ​​module.

[0136] The set of adjustable parameter instructions for the AI ​​module is transmitted in data packets via a high-speed bus interface interconnecting peripheral components. These data packets are sequentially sent to the actuator of the AI ​​module. The actuator receives and parses the data packets through the bus interface, extracting the set of adjustable parameter instructions for the AI ​​module. The actuator interprets each instruction in the set of adjustable parameter instructions for the AI ​​module, identifying the hardware register address, opcode, and data value within the instruction. Based on the instruction content, the actuator generates corresponding control signals to drive one or more parameter regulators to adjust the operating state of the AI ​​module.

[0137] The specific construction process of the parameter mapping table is as follows: The hardware register manual of the AI ​​module is consulted to extract the physical addresses, bit field definitions, and valid value ranges of all programmable registers related to dynamic adjustment. Simultaneously, the device driver source code of the AI ​​module is analyzed to identify the application programming interface functions and parameter structures used by the driver to configure the hardware registers. Each dimension of the final control vector is associated with an adjustable hardware parameter, and a mapping record is established for each association. The mapping record includes the index of the corresponding dimension in the final control vector, the physical address of the target hardware register, the operation bit field, the numerical conversion formula, and optional verification rules. The collection of all mapping records is stored in a file to form the parameter mapping table.

[0138] This embodiment also provides a computer device applicable to the adaptive control method for operating parameters of an artificial intelligence module, comprising: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the adaptive control method for operating parameters of an artificial intelligence module as proposed in the above embodiment.

[0139] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0140] This embodiment also provides a storage medium storing a computer program. When executed by a processor, the program implements the adaptive control method for the operating parameters of an artificial intelligence module as proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0141] In summary, this invention achieves deep perception of the operating state of an AI module by constructing a spatiotemporal state model and extracting high-order state representations. Based on this, conditional calculations are performed on the state representation vector and operating condition feature vector, enabling the control logic to dynamically adapt to different task requirements. Combined with a multi-objective Pareto optimization algorithm, the optimal trade-off control scheme is searched in real time under strict constraints such as power consumption, latency, and temperature. This not only improves the accuracy, stability, and adaptability of the control strategy but also effectively achieves synergistic optimization of energy efficiency and performance, solving the core problem of balancing safety and efficiency under current multi-objective conflicts, operating condition switching, and abnormal operating conditions.

[0142] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. An adaptive control method for operating parameters of an artificial intelligence module, characterized in that: include, Real-time acquisition of multimodal feedback data from artificial intelligence modules to construct spatiotemporal state models; Extracting a state representation vector that represents the current operating state of the artificial intelligence module based on a spatiotemporal state model; Based on state representation vectors and multimodal feedback data, the current operating condition is identified, and an operating condition feature vector is generated. Adaptive control law calculations are performed on the state representation vector and the operating condition feature vector to obtain a preliminary control vector for adjusting the operating parameters of the artificial intelligence module; Based on the state representation vector and the preliminary control vector, predict the future operating index values ​​of the artificial intelligence module under different operating parameter adjustments; Starting with the initial control vector and optimizing future performance indicators, a multi-objective optimization problem is constructed. A multi-objective optimization algorithm is used to perform a real-time search on the Pareto solution set of the multi-objective optimization problem to obtain the final control vector. The final control vector is sent to the actuator of the artificial intelligence module for physical adjustment.

2. The adaptive control method for operating parameters of an artificial intelligence module as described in claim 1, characterized in that: The multimodal feedback data includes time-series monitoring data, event log data, and topology data describing the connection relationships between components.

3. The adaptive control method for operating parameters of an artificial intelligence module as described in claim 2, characterized in that: The construction of the spatiotemporal state model specifically involves: The basic graph structure is constructed by using the hardware components of the artificial intelligence module as physical nodes and the connection relationships between the hardware components represented by topological data as edges. By using time-series monitoring data as the time-series attribute sequence of corresponding physical nodes, natural language processing is performed on event record data to extract key events. These key events are then integrated as attribute information into the basic graph structure to form a spatiotemporal state model.

4. The adaptive control method for operating parameters of an artificial intelligence module as described in claim 3, characterized in that: The extraction of the state representation vector representing the current operating state of the artificial intelligence module refers to performing correlation analysis on the physical nodes in the spatiotemporal state model and extracting time evolution features by combining time series attributes to obtain the state representation vector.

5. The adaptive control method for operating parameters of an artificial intelligence module as described in claim 4, characterized in that: The process of identifying the current operating condition and generating an operating condition feature vector based on the state representation vector and multimodal feedback data is as follows: Statistical analysis is performed on the state representation vector to extract high-order statistical feature vectors. At the same time, trend analysis is performed on the time series monitoring data in the multimodal feedback data to extract time series pattern feature vectors. The high-order statistical feature vector is concatenated with the time-series pattern feature vector to form the original working condition description feature vector; Principal component analysis is used to perform linear transformation and dimensionality reduction on the original working condition description feature vector, and the first k principal components are selected as the basic components. The basic components are matched with the template features in the working condition template library based on similarity. The working condition template with the highest matching degree is selected and the corresponding template code is extracted. The template encoding and the basic components are weighted and fused to generate the working condition feature vector.

6. The adaptive control method for operating parameters of an artificial intelligence module as described in claim 5, characterized in that: The adaptive control law calculation of the state representation vector and the operating condition feature vector yields a preliminary control vector for adjusting the operating parameters of the artificial intelligence module, specifically as follows: Map the state representation vector and the operating condition feature vector to control law modulation parameters; By using the modulation parameters of the control law to perform feature modulation on the state representation vector, conditional state features are generated. The control strategy is calculated based on the conditional state characteristics to obtain the preliminary control vector.

7. The adaptive control method for operating parameters of an artificial intelligence module as described in claim 6, characterized in that: The prediction of future operating index values ​​of the artificial intelligence module under different operating parameter adjustments, based on the state representation vector and the preliminary control vector, is as follows: The state representation vector is concatenated with the preliminary control vector to form a joint feature vector; Perform a high-order polynomial expansion on the joint feature vectors to generate a nonlinear feature combination; Using a nonlinear feature combination vector as input, a set of predefined index prediction functions are applied to calculate future operating index values. The future operating performance indicators include power consumption, latency, and temperature.

8. The adaptive control method for operating parameters of an artificial intelligence module as described in claim 7, characterized in that: The construction of the multi-objective optimization problem is specifically as follows: The initial control vector is used as the starting point for the optimization search, and the power consumption, delay, and temperature parameters are defined as objective functions to be minimized. Set corresponding constraint boundaries for each objective function; The constraints include power consumption upper limit constraints, delay threshold constraints, and temperature safety constraints; integrating the objective function and the constraints forms a multi-objective optimization problem.

9. The adaptive control method for operating parameters of an artificial intelligence module as described in claim 8, characterized in that: The process employs a multi-objective optimization algorithm to perform a real-time search within the Pareto solution set of the multi-objective optimization problem to obtain the final control vector. Specifically: The multi-objective optimization problem is decomposed into multiple scalar quantum problems using a decomposition-based multi-objective evolutionary algorithm, and a population of candidate solutions is maintained in the neighborhood of each scalar quantum problem. After initializing the candidate solution population with the initial control vector, crossover and mutation operations are performed to generate new solutions; The power consumption, delay, and temperature values ​​corresponding to each new solution are calculated using an evaluation function. The candidate solution population is then updated based on the dominance relationship and aggregation function. Crossover, mutation, evaluation, and update operations are performed iteratively until the termination condition is met, and the final candidate solution population is obtained. After performing non-dominated sorting on the final candidate solution population, an initial control vector that satisfies all constraints is selected as the final control vector at the Pareto front.

10. The adaptive control method for operating parameters of an artificial intelligence module as described in claim 9, characterized in that: The process of sending the final control vector to the actuator of the artificial intelligence module for physical adjustment specifically involves: The final control vector is parsed into a set of adjustable parameter instructions for the artificial intelligence module; The adjustable parameter instruction set is sent sequentially to the actuator of the artificial intelligence module through the hardware communication interface for parameter adjustment.

Images