Artificial intelligence model control system and control method

Abstract

The present invention relates to the field of artificial intelligence control technology, and discloses an artificial intelligence model control system and control method, comprising: a device data acquisition module for collecting measurement data of edge devices and constructing a virtual object data unit; a data preprocessing module for preprocessing the virtual object data unit and writing it into an instant information area and a simulated memory area respectively; a model inference module for constructing a first feature vector and inputting it into a random forest model to obtain a model prediction result; a credibility calculation module for calculating a credibility weight according to a first preset function; a control execution module for correcting the model prediction result and generating a current control output value; and a model adjustment module for constructing a control behavior change feature and adjusting the artificial intelligence model according to a first control strategy. The present invention improves the real-time performance, adaptability, and abnormal data processing capabilities of the control system.

Description

Technical Field

[0001] The present invention belongs to the field of artificial intelligence control technology, and specifically relates to an artificial intelligence model control system and control method. Background Art

[0002] Currently, traditional control systems in the industrial control and automation fields mostly use classic control strategies, such as PID control. While simple and effective, these control methods often have limitations. For example, PID control cannot effectively handle system nonlinearities and large-scale changes, and its parameter adjustment is cumbersome, making it difficult to achieve automatic optimization under real-time conditions.

[0003] With the rapid development of artificial intelligence technology, more and more control systems are beginning to introduce machine learning and deep learning methods in order to improve control accuracy and system stability. Currently, control systems based on machine learning have gradually been applied to some complex systems, such as autonomous driving, smart grids, and industrial robots. However, these systems still have some problems: 1. Real-time issues: Current control systems based on machine learning often require a long time for training and data processing, making it difficult to adapt to highly dynamic real-time control tasks; 2. Poor model adaptability: Traditional machine learning models have difficulty automatically adjusting parameters according to environmental changes, resulting in a delayed response of the control system when faced with sudden changes; 3. Abnormal data processing issues: Abnormal data or noise in the control system can cause model output errors, affecting the stability and reliability of the control system. Summary of the Invention

[0004] The present invention provides an artificial intelligence model control system and control method to solve the technical problems of poor real-time performance, insufficient model adaptability and improper abnormal data processing in related technologies.

[0005] The present invention provides an artificial intelligence model control system, comprising:

[0006] The device data acquisition module is used to collect measurement data from multiple edge devices and construct corresponding virtual object data units;

[0007] A data preprocessing module is used to preprocess virtual object data units and write them into the real-time information area and the virtual memory area respectively. The real-time information area is used to provide input data for the artificial intelligence model at the current moment, and the virtual memory area is used to store historical status data and execution records.

[0008] A model inference module is used to extract pre-processed virtual object data units from the instant information area, construct a first feature vector, and input the first feature vector into a random forest model for inference to obtain a model prediction result;

[0009] a credibility calculation module, configured to obtain a feedback response value returned by an external device after a previous round of control operations, calculate a prediction error based on the feedback response value and a current model prediction result, and calculate a credibility weight of the artificial intelligence model output according to a first preset function, wherein the external device represents a target control object for executing a control instruction generated by the artificial intelligence model;

[0010] The control execution module is used to modify the model prediction results according to the credibility weight, and perform weighted processing based on the feedback response value of the previous round of control operations to generate the current control output value, and then generate control instructions, which are sent to the external device through the communication module to execute the corresponding control operation;

[0011] The model adjustment module is used to construct the control behavior change characteristics based on the current control output value and the previous round of control output value, and write the prediction error, credibility weight, feedback response value and control behavior change characteristics into the simulated memory area, and then adjust the artificial intelligence model according to the first control strategy.

[0012] Furthermore, the measurement data includes: temperature, current, voltage and frequency, and the virtual object data unit includes: device identification, timestamp, temperature, current, voltage and frequency.

[0013] Furthermore, the pre-processing steps include:

[0014] S201, for each parameter of the measurement data, using a linear regression method based on the measurement values ​​at the current time point and the two previous time points, predicting the trend value of the parameter at the current time point;

[0015] S202, calculating the offset between the measured value and the trend value at the current time point, and using the offset as an input feature of the corresponding parameter;

[0016] S203: Parameters whose offset exceeds the first preset threshold are regarded as abnormal values ​​and replaced with the trend value at the current time point;

[0017] S204: normalize the processed measurement data using a maximum-minimum normalization method.

[0018] Furthermore, the first eigenvector is composed of offsets of parameters in the preprocessed virtual object data unit, wherein the parameters include temperature, current, voltage, and frequency, wherein the expression of the first eigenvector is: , Indicates the temperature offset, Indicates the current offset, Indicates the voltage offset, Indicates the frequency offset.

[0019] Furthermore, the first eigenvector is input into the random forest model for inference to obtain the model prediction result. The specific steps include:

[0020] S301: Inputting the first feature vector into multiple decision trees in the random forest model respectively, each decision tree traversing the first feature vector based on an internal partitioning rule until traversing to a leaf node, and outputting a sub-prediction value defined by the leaf node, wherein the internal partitioning rule is used to perform feature partitioning on non-leaf nodes, and the feature partitioning includes determining the branching direction of the traversal path according to a feature splitting threshold set on the non-leaf nodes;

[0021] S302, performing result fusion on the sub-prediction values ​​output by multiple decision trees, and using a weighted average method to generate the model prediction result.

[0022] Furthermore, the prediction error is the absolute difference between the current model prediction result and the feedback response value. The first preset function is constructed by the prediction error and the average value of the prediction errors of the past n rounds, where n represents a preset number of historical control operation rounds. The determination process of the first preset function is:

[0023] Obtain the model prediction result of the current round t and the feedback response value returned by the external device, calculate the absolute difference between the two, and obtain the prediction error of the current round;

[0024] Based on the historical control operation rounds, extract the n-round prediction error values ​​from round tn to round t, where n is the preset number of historical rounds, calculate the arithmetic mean of the prediction error values, and obtain the average value of the prediction errors of the past n rounds; add a minimum value between 0 and 1 to the average value of the prediction errors of the past n rounds to obtain an adjusted average value to prevent the divisor from being zero; divide the prediction error of the current round by the adjusted average value to obtain a ratio; raise the ratio to the power of the second weight coefficient to obtain a power processing result; multiply the first weight coefficient by the power processing result to obtain a scaled processing value; add the value 1 to the scaled processing value to obtain a sum value; take the inverse of the sum value to obtain the credibility weight of the current round as the output result of the first preset function.

[0025] Furthermore, correcting the model prediction result according to the credibility weight includes: performing weighted fusion processing according to the deviation between the current model prediction result and the previous round of feedback response value in combination with the current credibility weight to obtain the current corrected control output value.

[0026] Furthermore, the control behavior change characteristics are obtained by the following steps:

[0027] Calculating a first difference between the corrected control output value of the tth round and the corrected control output value of the t-1th round; performing a square operation on the first difference to obtain a first square result;

[0028] Obtaining a standard deviation of the control output value and adding a minimum value between 0 and 1 to the standard deviation to obtain a denominator adjustment value; dividing the first square result by the denominator adjustment value to obtain a first ratio; processing the first ratio to obtain a first sum;

[0029] Performing a natural logarithm operation on the first sum to obtain a first logarithm result; calculating a second difference between the corrected control output value of the t-1th round and the corrected control output value of the t-2th round; and calculating an absolute value difference between the first difference and the second difference;

[0030] Multiplying a preset weight coefficient by the absolute value difference to obtain a product result; calculating the power of the product result of the natural constant e to obtain an exponential operation result;

[0031] The exponential operation result is processed to obtain a second sum value; and the first logarithmic result is multiplied by the second sum value to obtain a final control behavior change characteristic value.

[0032] Furthermore, the first control strategy includes:

[0033] When the duration of the control behavior change characteristic exceeding the second preset threshold exceeds the first preset time period, switching the currently used artificial intelligence model to a preset backup model;

[0034] Based on the paired data stored in the simulated memory area, a local retraining operation is performed on the decision tree in the current artificial intelligence model, wherein the paired data includes: the first eigenvector, the model prediction result and the feedback response value; the local retraining operation includes: adjusting the output value of the leaf node and modifying the feature splitting threshold of the non-leaf node.

[0035] The present invention provides an artificial intelligence model control method, comprising the following steps:

[0036] S401, collecting measurement data of multiple edge devices and constructing corresponding virtual object data units;

[0037] S402, pre-processing the virtual object data unit and writing it into the real-time information area and the virtual memory area respectively, wherein the real-time information area is used to provide input data of the artificial intelligence model at the current moment, and the virtual memory area is used to store historical state data and execution records;

[0038] S403, extracting pre-processed virtual object data units from the instant information area, constructing a first feature vector, and inputting the first feature vector into a random forest model for inference to obtain a model prediction result;

[0039] S404, obtaining a feedback response value returned by the external device after the previous round of control operations, calculating a prediction error based on the feedback response value and a current model prediction result, and calculating a credibility weight of the artificial intelligence model output according to a first preset function, wherein the external device represents a target control object for executing a control instruction generated by the artificial intelligence model;

[0040] S405: Correct the model prediction result according to the credibility weight and perform weighted processing based on the feedback response value of the previous round of control operation to generate the current control output value, and then generate a control instruction, which is sent to the external device through the communication module to execute the corresponding control operation;

[0041] S406, constructing the control behavior change characteristics based on the current control output value and the previous round of control output value, and writing the prediction error, credibility weight, feedback response value and control behavior change characteristics into the simulated memory area, and then adjusting the artificial intelligence model according to the first control strategy.

[0042] The beneficial effects of the present invention are as follows: the present invention sets an instant information area and a simulated memory area to process real-time control data and historical data respectively, optimizes data access speed, reduces I / O conflicts, and supports efficient model reasoning and control strategy adjustment; by combining real-time feedback and historical data, the model output is dynamically adjusted using a credibility calculation module to ensure that the model can be adaptively adjusted under different control environments, thereby improving the stability and reliability of the system; the present invention adopts dynamic correction and model adaptive adjustment strategies from data acquisition, preprocessing, model reasoning to control execution, thereby ensuring control accuracy and system stability, and can effectively cope with various complex control tasks. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 It is a module schematic diagram of an artificial intelligence model control system of the present invention. DETAILED DESCRIPTION

[0044] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed solely to enable those skilled in the art to better understand and implement the subject matter described herein, and that the functions and arrangements of the elements discussed may be varied without departing from the scope of this specification. Various examples may omit, substitute, or add various processes or components as needed. In addition, features described with respect to some examples may also be combined in other examples.

[0045] like Figure 1 As shown, an artificial intelligence model control system includes:

[0046] The device data acquisition module 101 is used to collect measurement data of multiple edge devices and construct corresponding virtual object data units;

[0047] A data preprocessing module 102 is used to preprocess virtual object data units and write them into a real-time information area and a simulated memory area, respectively. The real-time information area is used to provide input data for the current artificial intelligence model, and the simulated memory area is used to store historical status data and execution records.

[0048] The model inference module 103 is used to extract the pre-processed virtual object data unit from the instant information area, construct a first feature vector, and input the first feature vector into the random forest model for inference to obtain a model prediction result;

[0049] a credibility calculation module 104 for obtaining a feedback response value returned by an external device after the previous round of control operations, calculating a prediction error based on the feedback response value and a current model prediction result, and calculating a credibility weight of the artificial intelligence model output according to a first preset function, wherein the external device represents a target control object for executing a control instruction generated by the artificial intelligence model;

[0050] The control execution module 105 is used to modify the model prediction results according to the credibility weight, and perform weighted processing based on the feedback response value of the previous round of control operation to generate the current control output value, and then generate a control instruction, which is sent to the external device through the communication module to execute the corresponding control operation;

[0051] The model adjustment module 106 is used to construct the control behavior change characteristics based on the current control output value and the previous round of control output value, and write the prediction error, credibility weight, feedback response value and control behavior change characteristics into the simulated memory area, and then adjust the artificial intelligence model according to the first control strategy.

[0052] In one embodiment of the present invention, the measurement data includes: temperature, current, voltage and frequency. The virtual object data unit refers to the encapsulation of the measurement data into a standardized and reusable data structure with a single edge device as the granularity. The virtual object data unit includes: device identification, timestamp, temperature, current, voltage and frequency.

[0053] In one embodiment of the present invention, the edge device refers to a field data acquisition device used to collect measurement data such as temperature, current, voltage and frequency, including: temperature sensor, current sensor, voltage monitoring unit, frequency acquisition component; the external device refers to the target control object used to execute the control instructions generated by the artificial intelligence model, including electric actuators, drive modules or other response devices, which can return status information or response values ​​after executing the instructions.

[0054] The device data acquisition module establishes a connection with each sensor through the standardized acquisition interface RS485 serial port, and polls the data of all edge devices according to the preset adoption cycle; further, the data acquisition module performs field parsing and structured packaging on the measurement data collected by each edge device, using the device identifier as the index primary key, and writes the collected temperature, current, voltage and frequency into a structured data format according to the preset fields, and associates it with the current timestamp to construct a virtual object data unit. This unit serves as the standard interface structure for the input of subsequent artificial intelligence models, supporting automatic feature vector generation and historical data tracing functions.

[0055] In one embodiment of the present invention, the pre-processing step includes:

[0056] S201: For each parameter of the measured data, a linear regression method is used to predict the trend value of the parameter at the current time point based on the measured values ​​at the current time point and the two previous time points. Specifically, the regression coefficient is calculated using the least squares method based on the measured values ​​at the three time points, and a regression equation is constructed. The measured values ​​at the two previous time points are then substituted into the regression equation to obtain the trend value at the current time point. This trend value is used to reflect the expected development direction of the parameter and helps determine whether there is a deviation in the current state.

[0057] S202, calculating the offset between the measured value and the trend value at the current time point, and using the offset as an input feature of the corresponding parameter;

[0058] S203: Parameters whose offset exceeds the first preset threshold are considered as outliers and replaced with trend values ​​at the current time point to maintain the continuity and stability of the feature data and prevent outliers from affecting the model prediction results;

[0059] S204, use the maximum and minimum normalization method to normalize the processed measurement data, and standardize all parameters to the range of 0 to 1. This method can effectively eliminate the numerical differences between different parameters due to different units and scales, and improve the numerical stability, calculation accuracy and convergence efficiency of the artificial intelligence model during training and inference.

[0060] In one embodiment of the present invention, the preprocessed data is written into two types of storage areas respectively, including: an instant information area and a simulated memory area, wherein the instant information area is implemented as a circular cache resident in the memory, indexed by device ID + timestamp, and only stores data of the most recent two rounds of control operation cycles for high-speed reading in the model inference stage; the simulated memory area is implemented as a local MySQL database, which writes complete virtual object data units and control feedback records on a minute-by-minute basis, and stores data of at least 50 control operation cycles; through this dual-area structure, the present invention realizes the decoupling of real-time control and historical learning, significantly reduces I / O conflicts without increasing the memory peak, and improves the convergence stability of the artificial intelligence model.

[0061] In one embodiment of the present invention, the first eigenvector is composed of offsets of parameters in the preprocessed virtual object data unit, wherein the parameters include temperature, current, voltage, and frequency. The expression of the first eigenvector is: , Indicates the temperature offset, Indicates the current offset, Indicates the voltage offset, Represents the frequency offset; the first eigenvector can accurately reflect the degree of deviation of the current state of the device compared to the predicted trend, improving the model's ability to identify device behavior fluctuations and control anomalies.

[0062] In one embodiment of the present invention, the first eigenvector is input into a random forest model for inference to obtain a model prediction result. The specific steps include:

[0063] S301: Inputting the first feature vector into multiple decision trees in the random forest model respectively, each decision tree traversing the first feature vector based on an internal partitioning rule until traversing to a leaf node, and outputting a sub-prediction value defined by the leaf node, wherein the internal partitioning rule is used to perform feature partitioning on non-leaf nodes, and the feature partitioning includes determining the branching direction of the traversal path according to a feature splitting threshold set on the non-leaf nodes;

[0064] Specifically, the internal division rule refers to the threshold judgment logic executed by the non-leaf nodes of each decision tree in the random forest model on the specific dimensional parameters in the first eigenvector, which is used to determine the path direction of the input data in the decision tree structure. Specifically, each non-leaf node sets a splitting condition, such as a frequency parameter ≤ 60Hz. When the first eigenvector meets the judgment condition, the system sends it to the left sub-branch of the node; otherwise, it enters the right sub-branch. This path traversal process continues in the tree structure until it reaches a leaf node, and the leaf node outputs a corresponding prediction value as the sub-prediction result of the decision tree. The above-mentioned division rules and leaf node output values ​​are automatically generated by minimizing the variance during the model training phase to form a decision logic structure for efficient classification or regression. The structure of each tree remains independent, supporting subsequent sub-structure replacement, leaf node update or split condition adjustment operations;

[0065] S302, performing result fusion on the sub-prediction values ​​output by multiple decision trees, and using a weighted average method to generate the model prediction result.

[0066] In one embodiment of the present invention, the feedback response value refers to the result data returned by the external device after receiving and executing the control instruction, including but not limited to: whether the control instruction is successfully executed, such as whether the switch is successfully switched, whether the device starts or stops as expected, etc.; after executing the control instruction, the status value of the external device, such as the temperature, current, voltage and other parameters returned by the sensor, as verification data for the effect of the control action; other auxiliary information generated after the device is executed, such as delay time, fault status, etc.

[0067] In one embodiment of the present invention, the prediction error is the absolute difference between the current model prediction result and the feedback response value. The first preset function is constructed by the prediction error and the average value of the prediction errors of the past n rounds, where n represents a preset number of historical control operation rounds. The first preset function is determined as follows:

[0068] Obtain the model prediction result of the current round t and the feedback response value returned by the external device, calculate the absolute difference between the two, and obtain the prediction error of the current round;

[0069] Based on the historical control operation rounds, extract the n-round prediction error values ​​from round tn to round t, where n is the preset number of historical rounds, calculate the arithmetic mean of the prediction error values, and obtain the average value of the prediction errors of the past n rounds; add a minimum value between 0 and 1 to the average value of the prediction errors of the past n rounds to obtain an adjusted average value to prevent the divisor from being zero; divide the prediction error of the current round by the adjusted average value to obtain a ratio; raise the ratio to the power of the second weight coefficient to obtain a power processing result; multiply the first weight coefficient by the power processing result to obtain a scaled processing value; add the value 1 to the scaled processing value to obtain a sum value; take the inverse of the sum value to obtain the credibility weight of the current round as the output result of the first preset function.

[0070] When specifically applied, the calculation formula of the first preset function can be implemented by the following formula, for example:

[0071] ;

[0072] in, Represents the credibility weight, which is used to dynamically evaluate the credibility of the prediction results of the current artificial intelligence model. Represents the model prediction results, Indicates the feedback response value, represents the prediction error, Represents the minimum value between 0 and 1. represents the average value of the prediction error from the past tnth round to the tth round, t represents the index of the control operation round, and Represent the first weight coefficient and the second weight coefficient respectively.

[0073] When the current prediction error is significantly greater than the average of the historical prediction errors, the calculated credibility weight will drop rapidly, reflecting the unreliability of the current model output; conversely, when the prediction error remains within the normal fluctuation range, the credibility weight remains at a high level, maintaining the model's dominant role in control instruction generation.

[0074] In one embodiment of the present invention, correcting the model prediction result according to the credibility weight includes: performing weighted fusion processing based on the deviation between the current model prediction result and the previous round of feedback response value in combination with the current credibility weight to obtain the current corrected control output value; wherein the calculation formula for correcting the model prediction result according to the credibility weight is:

[0075] ;

[0076] in, Represents the current control output value, that is, the corrected model prediction result, which is used to generate control instructions. Represents the feedback response value of the t-1th round.

[0077] The control instructions generated by the control execution module are sent to the external device via the communication module. The communication module is a wired communication interface, such as RS485 or CAN bus, which is used to reliably transmit the control instructions to the target control object.

[0078] In one embodiment of the present invention, the control behavior change characteristics are obtained by the following steps:

[0079] Calculating a first difference between the corrected control output value of the tth round and the corrected control output value of the t-1th round; performing a square operation on the first difference to obtain a first square result;

[0080] Obtaining a standard deviation of the control output value and adding a minimum value between 0 and 1 to the standard deviation to obtain a denominator adjustment value; dividing the first square result by the denominator adjustment value to obtain a first ratio; processing the first ratio to obtain a first sum;

[0081] Performing a natural logarithm operation on the first sum to obtain a first logarithm result; calculating a second difference between the corrected control output value of the t-1th round and the corrected control output value of the t-2th round; and calculating an absolute value difference between the first difference and the second difference;

[0082] Multiplying a preset weight coefficient by the absolute value difference to obtain a product result; calculating the power of the product result of the natural constant e to obtain an exponential operation result;

[0083] The exponential operation result is processed to obtain a second sum value; and the first logarithmic result is multiplied by the second sum value to obtain a final control behavior change characteristic value.

[0084] When applied specifically, the above control behavior change characteristics can be realized by the following formula, for example:

[0085] The control behavior change characteristics are calculated through nonlinear composite function, and the calculation formula is:

[0086] ;

[0087] in, Indicates the change characteristics of control behavior, which is used to measure the fluctuation intensity and mutation degree of control output value in the time dimension. represents the control output value of the t-1th round, represents the control output value of the t-2th round, Indicates the standard deviation of the control output value, Represents the exponential adjustment coefficient, which is used to adjust the sensitivity of the output value of the current artificial intelligence model to error changes;

[0088] The control behavior change characteristic takes into account the change amplitude between the current control output value and the previous round of control output value, as well as the difference in the change rate between two consecutive rounds of control values. By performing a nonlinear combination of these two parts, this embodiment can give differentiated responses when facing both slow-changing and sudden-changing behaviors. The higher the control behavior change characteristic, the more intense the control behavior and the more unstable the system state, which is more likely to trigger model switching or local adjustment operations at the control strategy layer.

[0089] In one embodiment of the present invention, the first control strategy includes:

[0090] When the duration of the control behavior change characteristic exceeding the second preset threshold exceeds the first preset time period, the currently used artificial intelligence model is switched to a preset backup model; the backup model is a stable model version obtained through historical training, and its structure is consistent with the main model, allowing seamless switching;

[0091] Based on the paired data stored in the simulated memory area, a local retraining operation is performed on the decision tree in the current artificial intelligence model, and the paired data includes: the first eigenvector, the model prediction result and the feedback response value.

[0092] In one embodiment of the present invention, the local retraining operation includes: adjusting the output value of the leaf node and modifying the feature splitting threshold of the non-leaf node;

[0093] Specifically, adjusting the output value of a leaf node means re-evaluating the target output value corresponding to a leaf node when the historical prediction error or accumulated deviation is large, without changing the decision tree structure. This embodiment collects the first eigenvector and feedback response value that hit the leaf node based on the paired data recorded in the virtual memory area, calculates the feedback mean, and updates the output result of the leaf node, thereby improving the prediction accuracy of the branch.

[0094] Modifying the feature splitting thresholds for non-leaf nodes refers to readjusting the feature partitioning conditions for certain non-leaf nodes while maintaining the tree's node structure. This operation is suitable for split nodes that frequently experience misjudgment. This embodiment re-evaluates the optimality of the current splitting conditions under the current data distribution based on the first eigenvector and its feedback response value in the historical paired data. By minimizing local errors, the original feature splitting threshold is replaced, thereby optimizing the node's judgment logic and improving the accuracy of path selection.

[0095] In one embodiment of the present invention, an artificial intelligence model control method is also provided, comprising the following steps:

[0096] S401, collecting measurement data of multiple edge devices and constructing corresponding virtual object data units;

[0097] S402, pre-processing the virtual object data unit and writing it into the real-time information area and the virtual memory area respectively, wherein the real-time information area is used to provide input data of the artificial intelligence model at the current moment, and the virtual memory area is used to store historical state data and execution records;

[0098] S403, extracting pre-processed virtual object data units from the instant information area, constructing a first feature vector, and inputting the first feature vector into a random forest model for inference to obtain a model prediction result;

[0099] S404, obtaining a feedback response value returned by the external device after the previous round of control operations, calculating a prediction error based on the feedback response value and a current model prediction result, and calculating a credibility weight of the artificial intelligence model output according to a first preset function, wherein the external device represents a target control object for executing a control instruction generated by the artificial intelligence model;

[0100] S405: Correct the model prediction result according to the credibility weight and perform weighted processing based on the feedback response value of the previous round of control operation to generate the current control output value, and then generate a control instruction, which is sent to the external device through the communication module to execute the corresponding control operation;

[0101] S406, constructing the control behavior change characteristics based on the current control output value and the previous round of control output value, and writing the prediction error, credibility weight, feedback response value and control behavior change characteristics into the simulated memory area, and then adjusting the artificial intelligence model according to the first control strategy.

[0102] It should be noted that the intervals and thresholds are set for ease of comparison. The threshold size depends on the amount of sample data and the cardinality set by those skilled in the art for each set of sample data, as long as it does not affect the proportional relationship between the parameter and the quantized value. Furthermore, the above formulas are all dimensionless numerical calculations. These formulas are derived from software simulations of the most recent real-world conditions using large amounts of data. The preset parameters in these formulas are set by those skilled in the art based on actual conditions.

[0103] The above describes the embodiments of the present invention, but the present invention is not limited to the above specific implementation methods. The above specific implementation methods are merely illustrative and not restrictive. Ordinary technicians in this field can also make many forms based on the inspiration of this embodiment, all of which are protected by this embodiment.

Claims

1. An artificial intelligence model control system, characterized in that: include: The device data acquisition module is used to collect measurement data from multiple edge devices and construct corresponding virtual object data units; A data preprocessing module is used to preprocess virtual object data units and write them into the real-time information area and the virtual memory area respectively. The real-time information area is used to provide input data for the artificial intelligence model at the current moment, and the virtual memory area is used to store historical status data and execution records. A model inference module is used to extract pre-processed virtual object data units from the instant information area, construct a first feature vector, and input the first feature vector into a random forest model for inference to obtain a model prediction result; a credibility calculation module, configured to obtain a feedback response value returned by an external device after a previous round of control operations, calculate a prediction error based on the feedback response value and a current model prediction result, and calculate a credibility weight of the artificial intelligence model output according to a first preset function, wherein the external device represents a target control object for executing a control instruction generated by the artificial intelligence model; The prediction error is the absolute difference between the current model prediction result and the feedback response value, and the first preset function is constructed by the prediction error and the average value of the prediction errors of the past n rounds, where n represents the preset number of historical control operation rounds; The control execution module is used to modify the model prediction results according to the credibility weight, and perform weighted processing based on the feedback response value of the previous round of control operations to generate the current control output value, and then generate control instructions, which are sent to the external device through the communication module to execute the corresponding control operation; A model adjustment module is used to construct a control behavior change feature based on the current control output value and the previous round of control output value, and write the prediction error, credibility weight, feedback response value and control behavior change feature into the quasi-physical memory area, and then adjust the artificial intelligence model according to the first control strategy; The control behavior change characteristics are obtained by the following steps: Calculating a first difference between the corrected control output value of the tth round and the corrected control output value of the t-1th round; performing a square operation on the first difference to obtain a first square result; Obtaining a standard deviation of the control output value and adding a minimum value between 0 and 1 to the standard deviation to obtain a denominator adjustment value; dividing the first square result by the denominator adjustment value to obtain a first ratio; processing the first ratio to obtain a first sum; Performing a natural logarithm operation on the first sum to obtain a first logarithm result; calculating a second difference between the corrected control output value of the t-1th round and the corrected control output value of the t-2th round; and calculating an absolute value difference between the first difference and the second difference; Multiplying a preset weight coefficient by the absolute value difference to obtain a product result; calculating the power of the product result of the natural constant e to obtain an exponential operation result; The exponential operation result is processed to obtain a second sum value; the first logarithmic result is multiplied by the second sum value to obtain a final control behavior change characteristic value.

2. An artificial intelligence model control system according to claim 1, characterized in that: The measurement data includes: temperature, current, voltage and frequency, and the virtual object data unit includes: device identification, timestamp, temperature, current, voltage and frequency.

3. An artificial intelligence model control system according to claim 1, characterized in that: The preprocessing steps include: S201, for each parameter of the measurement data, using a linear regression method based on the measurement values ​​at the current time point and the two previous time points, predicting the trend value of the parameter at the current time point; S202, calculating the offset between the measured value and the trend value at the current time point, and using the offset as an input feature of the corresponding parameter; S203: Parameters whose offset exceeds the first preset threshold are regarded as abnormal values ​​and replaced with the trend value at the current time point; S204: normalize the processed measurement data using a maximum-minimum normalization method.

4. An artificial intelligence model control system according to claim 2, characterized in that: The first feature vector is composed of offsets of parameters in the pre-processed virtual object data unit, and the parameters include: temperature, current, voltage and frequency.

5. The artificial intelligence model control system according to claim 1, characterized in that: The first eigenvector is input into the random forest model for inference to obtain the model prediction result. The specific steps include: S301: Inputting the first feature vector into multiple decision trees in the random forest model respectively, each decision tree traversing the first feature vector based on an internal partitioning rule until traversing to a leaf node, and outputting a sub-prediction value defined by the leaf node, wherein the internal partitioning rule is used to perform feature partitioning on non-leaf nodes, and the feature partitioning includes determining the branching direction of the traversal path according to a feature splitting threshold set on the non-leaf nodes; S302, performing result fusion on the sub-prediction values ​​output by multiple decision trees, and using a weighted average method to generate the model prediction result.

6. An artificial intelligence model control system according to claim 1, characterized in that: The process of determining the first preset function is: Obtain the model prediction result of the current round t and the feedback response value returned by the external device, calculate the absolute difference between the two, and obtain the prediction error of the current round; Based on historical control operation rounds, extracting prediction error values ​​for n rounds from round tn to round t, where n is a preset number of historical rounds, calculating the arithmetic mean of the prediction error values ​​to obtain an average value of the prediction errors for the past n rounds; adding a minimum value between 0 and 1 to the average value of the prediction errors for the past n rounds to obtain an adjusted average value to prevent division by zero; dividing the prediction error for the current round by the adjusted average value to obtain a ratio; raising the ratio to a power of a second weight coefficient to obtain a power processing result; and multiplying the power processing result by the first weight coefficient to obtain a scaled value; The value 1 is added to the scaling value to obtain a sum value; the reciprocal of the sum value is taken to obtain the credibility weight of the current round as the output result of the first preset function.

7. An artificial intelligence model control system according to claim 6, characterized in that: Correcting the model prediction result according to the credibility weight includes: performing weighted fusion processing based on the deviation between the current model prediction result and the previous round of feedback response value, combined with the current credibility weight, to obtain the current corrected control output value.

8. The artificial intelligence model control system according to claim 1, characterized in that: The first control strategy includes: When the duration of the control behavior change characteristic exceeding the second preset threshold exceeds the first preset time period, switching the currently used artificial intelligence model to a preset backup model; Based on the paired data stored in the simulated memory area, a local retraining operation is performed on the decision tree in the current artificial intelligence model, wherein the paired data includes: the first eigenvector, the model prediction result and the feedback response value; the local retraining operation includes: adjusting the output value of the leaf node and modifying the feature splitting threshold of the non-leaf node.

9. An artificial intelligence model control method, characterized in that: An artificial intelligence model control system according to any one of claims 1 to 8 is adopted, comprising the following steps: S401, collecting measurement data of multiple edge devices and constructing corresponding virtual object data units; S402, pre-processing the virtual object data unit and writing it into the real-time information area and the virtual memory area respectively, wherein the real-time information area is used to provide input data of the artificial intelligence model at the current moment, and the virtual memory area is used to store historical state data and execution records; S403, extracting pre-processed virtual object data units from the instant information area, constructing a first feature vector, and inputting the first feature vector into a random forest model for inference to obtain a model prediction result; S404, obtaining a feedback response value returned by the external device after the previous round of control operations, calculating a prediction error based on the feedback response value and a current model prediction result, and calculating a credibility weight of the artificial intelligence model output according to a first preset function, wherein the external device represents a target control object for executing a control instruction generated by the artificial intelligence model; S405: Correct the model prediction result according to the credibility weight and perform weighted processing based on the feedback response value of the previous round of control operation to generate the current control output value, and then generate a control instruction, which is sent to the external device through the communication module to execute the corresponding control operation; S406, constructing the control behavior change characteristics based on the current control output value and the previous round of control output value, and writing the prediction error, credibility weight, feedback response value and control behavior change characteristics into the simulated memory area, and then adjusting the artificial intelligence model according to the first control strategy.

Images