A deep learning optimization system and method for intelligent devices

Abstract

The application belongs to the technical field of data processing, and discloses a deep learning optimization system and method for intelligent equipment, which comprises the following steps: collecting data and performing data cleaning to obtain accurate equipment parameters; identifying the target intelligent equipment resource level based on the accurate equipment parameters; performing energy consumption scheduling on the target intelligent equipment based on the resource level to obtain the scheduled intelligent equipment; performing disturbance simulation on the accurate equipment parameters and outputting disturbance stability indexes; performing suitable parameter screening on the accurate equipment parameters based on the disturbance stability indexes and outputting stable equipment parameters; performing risk assessment on the stable equipment parameters to obtain parameter risk indexes; performing model matching on the scheduled intelligent equipment based on the parameter risk indexes to generate a model adjustment strategy; performing scene simulation based on the model adjustment strategy, loading and optimizing the model adjustment strategy based on the simulation effect, and outputting an optimized model strategy; and the application realizes an efficient deep learning optimization method for intelligent equipment.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and more specifically, to a deep learning optimization system and method for intelligent devices. Background Technology

[0002] With the rapid development of artificial intelligence technology, deep learning-based intelligent models have been widely deployed in various terminal devices and applied to tasks in fields such as image recognition, environmental perception, and intelligent control. In practical applications, accurately adapting deep learning models to terminal devices under different operating conditions and ensuring operational efficiency and security has become an important research direction in related fields, but existing technologies still face challenges in this regard.

[0003] In the actual operation of smart devices, the parameters related to the device's control resources are highly time-varying. For example, parameters such as power consumption, memory usage, and system temperature change with the task type or stage of task execution. At this time, the resource consumption of the deep learning model deployed in the device may no longer match the preset power consumption threshold of the device at a certain task stage. However, existing technologies often lack real-time monitoring and adjustment of energy consumption, resulting in a mismatch between model operation and the actual device resource state. For example, continuously loading a high-precision and high-load model under high-temperature conditions may cause the device to reduce its frequency or the model to stop running, thus causing task interruption. Furthermore, existing model deployment methods generally overlook the instability and high risk that the input data itself may exhibit during actual operation. For example, a certain model may experience performance degradation or crash after receiving certain feature input data. At the same time, existing technologies lack the necessary risk assessment before deploying models for such unstable input data, and cannot identify whether certain input data are more likely to trigger model output errors. For example, in a certain vehicle recognition system, when the car window is heavily fogged up in rainy weather, the intensity of background texture loss in the front-facing image exceeds the system's tolerance threshold, and the built-in model identifies the lane as a null value and triggers an incorrect path decision.

[0004] In view of this, the present invention proposes a deep learning optimization system and method for intelligent devices to solve the above problems. Summary of the Invention

[0005] To overcome the aforementioned deficiencies of the prior art and to achieve the above objectives, the present invention provides the following technical solution: a deep learning optimization method for intelligent devices, comprising:

[0006] S1. Collect the operating status parameters of the target intelligent device, and perform data cleaning on the operating status parameters to obtain accurate device parameters;

[0007] S2. Identify the resource level of the target intelligent device based on accurate device parameters; perform energy consumption scheduling on the target intelligent device based on the resource level to obtain the scheduled intelligent device;

[0008] S3. Perform disturbance simulation on precise equipment parameters and output disturbance stability index; based on the disturbance stability index, perform appropriate parameter selection on precise equipment parameters and output stable equipment parameters;

[0009] S4. Conduct a risk assessment of the stable equipment parameters to obtain parameter risk indicators;

[0010] S5. Based on parameter risk indicators, perform model matching on scheduling intelligent devices and generate model adjustment strategies;

[0011] S6. Simulate the scene based on the model adjustment strategy, load and optimize the model adjustment strategy based on the simulation effect, output the optimized model strategy, and apply the optimized model strategy to the target smart device.

[0012] Furthermore, the method for identifying the resource level of a target intelligent device based on precise device parameters includes:

[0013] Obtain task scenario information, and divide precise equipment parameters into time periods based on the task scenario information to obtain time-segmented equipment parameters; perform statistical feature analysis on the time-segmented equipment parameters of any time period to output state feature items; calculate the resource evolution amount between different time periods based on the state feature items, and construct change trend data based on the resource evolution amount;

[0014] The parameters belonging to each resource dimension in the trend data are weighted by features to obtain the resource status score vector for the corresponding time period; the number of abnormal situations in the corresponding time period is queried, an abnormal fluctuation coefficient is constructed based on the number of abnormal situations, and the abnormal fluctuation coefficient is multiplied with the resource status score vector to output the fluctuation status score vector; a resource level scoring matrix is ​​constructed, and the fluctuation status score vector is matched with the resource level scoring matrix to output the resource level.

[0015] Furthermore, the method of energy consumption scheduling for target intelligent devices based on resource levels includes:

[0016] Based on resource level, resource control points are marked for target intelligent devices, and resource control point labels are output. Based on the resource control point labels, the key control structures of the target intelligent devices are identified, and the regulation fields of the key control structures are deployed. The structure control function is constructed, and the specific parameter values ​​of the regulation fields are calculated using the structure control function. Based on resource level, a preset deep learning model library is screened, and the screened models are matched with the regulated target intelligent devices for energy consumption to obtain the scheduled intelligent devices.

[0017] Furthermore, the method for deploying the control fields of the key control structure includes:

[0018] For each critical control structure, define corresponding control fields and limit the energy consumption boundary of the critical control structure based on the control fields; deploy the control fields to the relevant configuration interfaces of the critical control structure.

[0019] The methods for matching the energy consumption of the screened model with the adjusted target smart device include:

[0020] The energy consumption of the target smart devices after regulation is assessed over time periods, and a stable operating power consumption range is set based on the assessment results. The load standard index of the screened model is obtained, and the stable operating power consumption range is compared with the load standard index. If the load standard index is within the stable operating power consumption range of the time period, the screened model is deployed to the target smart devices during this time period. If the energy consumption of the target smart devices changes over time, the model is hot-switched, and a model whose load standard index conforms to the corresponding stable operating power consumption range of this time period is loaded.

[0021] Furthermore, the method for simulating disturbances to precise equipment parameters includes:

[0022] Modeling of intelligent device operation scenarios based on accurate device parameters; constructing device disturbance templates and injecting dynamic disturbance events into the intelligent device operation scenario modeling based on the device disturbance templates;

[0023] Collect response change information and final simulation feedback data during the disturbance simulation process, and construct the disturbance response path for each precise device parameter from the time series dimension based on the response change information and final simulation feedback data; extract local disturbance segments in the disturbance response path, perform response feature analysis on the local disturbance segments, and output local disturbance response feature vectors;

[0024] The local stability quantity of each local disturbance response feature vector is calculated using a stability function; all local disturbance response feature vectors are concatenated according to the time sequence to output the overall disturbance feature vector; the macroscopic stability quantity of the overall disturbance feature vector is calculated; the local stability quantities are weighted and summed and added to the macroscopic stability quantity to output the disturbance stability parameters; and all disturbance stability parameters are integrated to construct the disturbance stability index.

[0025] Furthermore, the method for selecting appropriate parameters for precise equipment parameters based on disturbance stability indices includes:

[0026] The abnormal response fluctuation values ​​of precise equipment parameters in any round of disturbance simulation are calculated based on the disturbance stability index; parameters with abnormal response fluctuation values ​​less than the preset fluctuation threshold are taken as stable equipment parameters.

[0027] Furthermore, the methods for conducting risk assessments on stable equipment parameters include:

[0028] Abnormal response trajectories are identified based on the abnormal response fluctuation values ​​of each round of disturbance simulation using stable equipment parameters; abnormal response trajectories are analyzed to obtain abnormal fluctuation characteristics; and an abnormal mapping relationship between abnormal fluctuation characteristics and stable equipment parameters is established.

[0029] Determine the corresponding model for stable equipment parameters and construct a performance sensitivity map of the model under stable equipment parameter conditions; extract anomaly correlation subgraphs from the performance sensitivity map based on anomaly mapping relationships; quantify the offset degree of each anomaly correlation subgraph to obtain an offset mapping vector; construct a risk propagation model, use the offset mapping vector as the basis for the operation of the risk propagation model, calculate the risk score of stable equipment parameters using the risk propagation model, and construct parameter risk indicators based on the risk score.

[0030] Furthermore, the method for performing model matching on the scheduling intelligent device includes:

[0031] Identify affected structures in the current deployment model based on parameter risk indicators; obtain functional association information of the affected structures, and adjust the parameters of the affected structures based on the functional association information to obtain a parameter correction model; calculate the predicted power consumption of the parameter correction model.

[0032] By combining the resource status parameters of the scheduling intelligent device for any given time period, it is determined whether the predicted operating power consumption value conforms to the stable operating power consumption range of the scheduling intelligent device within that time period. If it does, the parameter correction model is matched with the scheduling intelligent device for that time period; otherwise, a secondary adjustment is made to the parameter correction model. Power consumption analysis is performed on the parameter correction model to identify high-energy-consuming structures. Hierarchical reconstruction is performed on the high-energy-consuming structures to generate a power-saving parameter model and match it with the corresponding scheduling intelligent device. The model matching process is structured and the model adjustment strategy is output.

[0033] Furthermore, the method for loading and optimizing the model adjustment strategy includes:

[0034] Set the expected value of the simulation effect. If the simulation effect is less than the expected value, optimize the parameter configuration of the model corresponding to the model adjustment strategy. Stop iterating when the simulation effect is greater than or equal to the expected value and output the optimized model strategy.

[0035] A deep learning optimization system for intelligent devices, used to implement a deep learning optimization method for intelligent devices, characterized in that it includes:

[0036] The data acquisition module is used to collect the operating status parameters of the target smart device and perform data cleaning on the operating status parameters to obtain accurate device parameters.

[0037] The energy consumption scheduling module is used to identify the resource level of the target intelligent device based on accurate equipment parameters; and to perform energy consumption scheduling on the target intelligent device based on the resource level to obtain the scheduled intelligent device.

[0038] The parameter filtering module is used to simulate disturbances to precise equipment parameters and output disturbance stability indicators; based on the disturbance stability indicators, it filters appropriate parameters for precise equipment parameters and outputs stable equipment parameters.

[0039] The risk assessment module is used to assess the risks of stable equipment parameters and obtain parameter risk indicators.

[0040] The strategy generation module is used to perform model matching on the scheduling intelligent devices and generate model adjustment strategies.

[0041] The strategy optimization module is used to simulate scenarios based on model adjustment strategies, load and optimize the model adjustment strategies based on the simulation results, output optimized model strategies, and apply the optimized model strategies to the target smart devices; the modules are connected to each other via wired and / or wireless means.

[0042] The technical effects and advantages of the deep learning optimization system and method for intelligent devices of this invention are as follows:

[0043] By constructing a closed loop from device status monitoring to model deployment strategy optimization, the stability and security of deep learning models running on intelligent devices in resource-constrained environments are improved. Compared with existing technologies, this method achieves energy consumption matching between the model and the device by identifying resource levels and regulating energy consumption, enabling synchronous perception of system resources during model deployment and ensuring the model can operate normally under resource constraints. Perturbation simulation and result evaluation of device status parameters effectively identify unstable behaviors that the model may exhibit in certain input scenarios or under high-perturbation environments, and filters corresponding parameters to avoid risks such as model performance crashes caused by certain parameters. Simultaneously, parameter correction is achieved by combining resource status and parameter risk assessment, and the model adjustment strategy is optimized in multiple rounds using optimization algorithms. Therefore, this deep learning optimization method for intelligent devices ensures model accuracy and robustness while maximizing adaptation to the computing power and power consumption requirements of the target intelligent device, improving the overall operational security and scheduling capabilities of the device. Attached Figure Description

[0044] Figure 1 This is a schematic diagram of a deep learning optimization method for a smart device according to the present invention;

[0045] Figure 2 This is a schematic diagram of a deep learning optimization system for an intelligent device according to the present invention. Detailed Implementation

[0046] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0047] Example 1

[0048] Please see Figure 1 As shown in this embodiment, a deep learning optimization method for intelligent devices includes:

[0049] S1. Collect the operating status parameters of the target intelligent device, and perform data cleaning on the operating status parameters to obtain accurate device parameters;

[0050] S2. Identify the resource level of the target intelligent device based on accurate device parameters; perform energy consumption scheduling on the target intelligent device based on the resource level to obtain the scheduled intelligent device;

[0051] S3. Perform disturbance simulation on precise equipment parameters and output disturbance stability index; based on the disturbance stability index, perform appropriate parameter selection on precise equipment parameters and output stable equipment parameters;

[0052] S4. Conduct a risk assessment of the stable equipment parameters to obtain parameter risk indicators;

[0053] S5. Based on parameter risk indicators, perform model matching on scheduling intelligent devices and generate model adjustment strategies;

[0054] S6. Simulate the scene based on the model adjustment strategy, load and optimize the model adjustment strategy based on the simulation effect, output the optimized model strategy, and apply the optimized model strategy to the target smart device.

[0055] The operating status parameters include parameters that reflect the operating status of the target smart device, such as power consumption, current value, operating temperature, CPU utilization, and operating load. In this embodiment, the data cleaning process is realized by filling in missing values, isolating outliers, and standardizing the operating status parameters, resulting in higher quality data that is easier to process in the future.

[0056] Methods for identifying the resource level of a target intelligent device based on accurate device parameters include:

[0057] The process involves acquiring task scenario information and then dividing precise device parameters into time-segmented parameters based on this information. Task scenario information refers to the type of task being executed corresponding to the precise device parameters. A sliding window with adjustable size is constructed, using the timestamp and task type of the precise device parameters as the basis for time-segmentation. In this embodiment, time-segmentation is marked by task switching of the smart device, ensuring that the time-segmented device parameters are representative of each stage and facilitating subsequent processing of parameters for each time segment. Statistical feature analysis is performed on the time-segmented device parameters for any given time segment, outputting state feature items. Statistical calculations are performed on the set of device parameters for each time segment to obtain features reflecting data distribution, such as mean, median, standard deviation, and extreme values. These features are then combined into state feature items, intuitively reflecting the load intensity of the target smart device in various resource dimensions, such as power, storage, and memory, during a given time segment.

[0058] The resource evolution amount between different time periods is calculated based on state feature terms, and change trend data is constructed based on the resource evolution amount. The state feature terms of multiple adjacent time periods are used as input to calculate the change amount information of each resource dimension, such as parameter change rate, average change magnitude, and change direction, which is the resource evolution amount. In this embodiment, the resource change curve is fitted using a regression model based on the resource evolution amount, and the parameters corresponding to each time stamp and the resource evolution amount of each segment are marked on the curve to obtain the change trend data.

[0059] The parameters belonging to each resource dimension in the trend data are weighted to obtain the resource status score vector for the corresponding time period. By assigning corresponding weights to the parameters of each resource dimension in the corresponding time period of the trend data, the mean of these parameters over the entire time period is calculated and weighted to obtain the resource status score vector. The weight settings are adjusted based on specific operating conditions. In this embodiment, key parameters that directly reflect resource changes, such as changes in CPU utilization and power consumption rate, are given greater weights, while other auxiliary parameters, such as changes in Bluetooth interface power consumption, which have a minimal impact on the smart target device, are given smaller weights. The weight values ​​range from [value missing]. .

[0060] The system queries the number of abnormal situations for a given time period, constructs an abnormal fluctuation coefficient based on this number, and multiplies the abnormal fluctuation coefficient with the resource status score vector to output the fluctuation status score vector. Abnormal situations include sudden temperature increases and sudden changes in CPU usage. The abnormal fluctuation coefficient is constructed based on the number of abnormal situations to represent the degree of instability during that time period. The abnormal fluctuation coefficient is set based on specific operating conditions. Multiplying the abnormal fluctuation coefficient with the original resource status score vector further highlights the instability of the corresponding time period, indirectly improving the accuracy of subsequent resource level assessments.

[0061] A resource level scoring matrix is ​​constructed, and the fluctuation state score vector is matched with the resource level scoring matrix to output the resource level. The resource level scoring matrix sets scoring categories such as processing power, memory capacity, power supply, and transmission scheduling according to different operating state scenarios, and each category can correspond to each dimension in the fluctuation state score vector, which also includes multiple preset resource level intervals. The fluctuation state score vector is matched with the resource level scoring matrix to determine the resource level interval of each category. The comprehensive matching level of each fluctuation state score vector is identified using a preset resource level matching template, which is the resource level of the target intelligent device in the corresponding time period, and is used for subsequent energy consumption scheduling of the target intelligent device.

[0062] Methods for scheduling energy consumption of target intelligent devices based on resource levels include:

[0063] Resource control points are marked on the target smart device based on resource level, and resource control point labels are output. The resource control points that consume significant resources in the target smart device during the corresponding time period are identified by the resource level. These points are used to indicate the trigger points when the device performs certain operations. Examples of resource control point labels include CPU frequency change points and interface occupancy detection points.

[0064] The key control structures of target intelligent devices are identified based on resource control point tags. At the same time, regulation fields are deployed on the key control structures. The key control structures are the specific physical structures responsible for performing device operations, which are associated with each resource control point tag and are identified by resource control point tags. Examples include frequency adjustment modules, transmission modules, and interface allocation modules. The regulation fields are used to limit the resource consumption of the key control structures in any given time period.

[0065] A structural control function is constructed, and the specific parameter values ​​of the control field are calculated using the structural control function. In this embodiment, the basic structure of the structural control function is a multi-factor weighted model. By using resource level labels and parameters of each dimension in the historical records of the target intelligent device as input data, the specific parameter values ​​of the control field of the structure to be controlled are calculated. Based on the resource level, a preset deep learning model library is screened, and the screened models are matched with the target intelligent device after control for energy consumption to obtain the scheduled intelligent device. The preset deep learning model library includes a variety of deep learning models. The structural complexity of the deep learning models is matched with the load capacity reflected by the resource level, and models whose structural complexity exceeds the adaptability of the current resource level are eliminated. This completes the initial screening of deep learning models and ensures that the screened models can run under the corresponding resource level conditions.

[0066] Methods for deploying control fields in critical control structures include:

[0067] For each key control structure, a corresponding control field is set, and the energy consumption boundary of the key control structure is limited based on the control field. The control field includes fields related to the operation performed by the key control structure, such as maximum frequency field, thread number limit field, standby interval field, and power allocation field. The control field is used to constrain the scheduling limit and working cycle of the key control structure in the corresponding time period, which affect energy consumption indicators. The control field is deployed to the relevant configuration interface of the key control structure. In this embodiment, the control field is written into the configuration interface corresponding to the key control structure in the target intelligent device in the form of computer instructions.

[0068] The methods for matching the energy consumption of the screened model with the adjusted target smart device include:

[0069] The target smart device after regulation is evaluated for energy consumption over a period of time, and a stable operating power consumption range is set based on the evaluation results. In this embodiment, the energy consumption parameters of the target smart device after regulation are collected in the corresponding period of time, including parameters that reflect the device load, such as power supply temperature, current fluctuation value and memory occupancy rate. By statistically analyzing the energy consumption parameters of each period of time and monitoring the operating status of the target smart device after regulation, the upper and lower limits of the energy consumption parameters corresponding to the stable operating status are taken as the stable operating power consumption range.

[0070] Obtain the load standard indicators of the filtered models, and compare the stable operating power consumption range with the load standard indicators. If the load standard indicators are within the stable operating power consumption range for that period, the filtered models are deployed to the target smart devices for that period. The load standard indicators include parameters such as the power consumption per run, memory usage, and peak current of the model. Compare the load standard indicators with the stable operating power consumption range item by item. If all parameters in the indicators fall within the stable operating power consumption range, the model is considered to match the power consumption capability of the target smart devices for the current period and is allowed to be deployed and run.

[0071] If the energy consumption of the target intelligent device changes over time, the model is hot-switched. The loaded load standard indicators conform to the model in the stable operating power consumption range corresponding to this time period. Since the system power consumption changes over time when the target intelligent device is running in actual working conditions, the deployed model no longer matches the energy consumption requirements of the changed target intelligent device. Therefore, in this embodiment, a seamless replacement mechanism is used to implement hot-switching of the model to ensure that the tasks executed by the system front end are not interrupted, and at the same time, the model and parameters are quickly migrated and switched, ensuring the sustainable and efficient operation capability of the target intelligent device.

[0072] Methods for simulating disturbances in precise equipment parameters include:

[0073] Based on precise device parameters, a smart device operation scenario model is constructed. In this embodiment, precise device parameters are used as the basis to construct a smart device operation scenario model in a virtual environment, including information such as processor configuration parameters, scheduling of each module, and power consumption changes. The scenario model refers to a simulated scenario in a certain period of time. A device disturbance template is constructed, and dynamic disturbance events are injected into the smart device operation scenario model based on the device disturbance template. The device disturbance template is a parameter configuration scheme based on historical device operation records and existing data, which contains multiple triggerable model performance crashes or instability reactions. These include disturbance types such as CPU frequency switching disturbances, power supply current fluctuations, and memory write obstruction.

[0074] The system collects response change information and final simulation feedback data during the disturbance simulation process. Based on the response change information and final simulation feedback data, it constructs the disturbance response path for each precise device parameter from a time-series perspective. The response change information includes data extracted during the disturbance simulation process that can reflect the changes in parameters affected by the disturbance, such as the number of voltage callbacks, the number of thread migrations, and the rate of temperature change. The final simulation feedback data refers to the result value of parameter changes obtained at the end of the simulation.

[0075] Local disturbance segments are extracted from the disturbance response path, and response feature analysis is performed on the local disturbance segments to output local disturbance response feature vectors. The complete disturbance response path is decomposed into several local disturbance segments based on the time-series dimension. The local disturbance response feature vectors obtained by response feature analysis include elements such as response frequency, time delay, and fluctuation amplitude.

[0076] The local stability quantity of each local perturbation response eigenvector is calculated using a stability function. In this embodiment, the stability function refers to the perturbation reconstruction residual function, and the formula for calculating the stability function is as follows: ;in, Indicates a locally stable quantity; This represents the average stability score. This represents the maximum risk score. This represents the cumulative fluctuation score; This represents the deviation between the segment corresponding to the local perturbation response feature vector and the reference segment. The timestamp represents a local disturbance segment; ;in, This represents the segment corresponding to the feature vector of the local disturbance response; Indicates passage The baseline segment obtained through smooth fitting; , and These represent the weights corresponding to the average stability score, maximum risk score, and cumulative volatility score, respectively, and are set based on expert experience in this embodiment. , , The local stability quantity calculated using this function is the stability score of the local perturbation segment, reflecting the stability of the segment under perturbation conditions.

[0077] Based on the temporal sequence, all local disturbance response feature vectors are concatenated to output the overall disturbance feature vector. The overall disturbance feature vector refers to the vector corresponding to the complete disturbance response path within the complete cycle, representing the complete disturbance coverage. The macroscopic stability quantity of the overall disturbance feature vector is calculated. In this embodiment, the trajectory disturbance degree function is used to calculate the macroscopic stability quantity, and the formula for calculating the macroscopic stability quantity is as follows: ;in Indicates macroeconomic stable quantities; This represents the standard deviation of each parameter within the complete disturbance response path; This represents the average slope of the response curves for different local segments within the complete disturbance response path; This indicates the frequency at which intelligent devices self-adjust under disturbance conditions, such as voltage callback, in the complete disturbance response path; , and The weights for standard deviation, mean slope, and self-adjusting frequency are respectively set based on expert experience and are used in this embodiment. , , Macroscopic stability parameters are used to quantify the degree to which the overall response path is disturbed at the structural level. Local stability parameters are weighted and summed, then added to the macroscopic stability parameters to output perturbation stability parameters. All perturbation stability parameters are integrated to construct a perturbation stability index. These perturbation stability parameters quantify the stability of a particular parameter during the perturbation process, facilitating subsequent parameter selection based on stability strength. The perturbation stability index is a row matrix, where any element represents any perturbation stability parameter. Combining all perturbation stability parameters into the same matrix yields the perturbation stability index. For example, setting a perturbation stability parameter to... Furthermore, by adjusting the tail numbers to represent other disturbance stability parameters, the resulting disturbance stability index is... ,in This indicates the number of disturbance stability parameters.

[0078] Methods for selecting appropriate parameters for precise equipment based on disturbance stability indices include:

[0079] The abnormal response fluctuation values ​​of precise equipment parameters in any round of disturbance simulation are calculated based on the disturbance stability index. This is achieved by comparing the parameter changes before and after each round of disturbance simulation, and combining the disturbance stability parameters of each dimension parameter in the disturbance stability index to obtain the abnormal response fluctuation values ​​of each dimension parameter in a single disturbance simulation. The disturbance stability parameters are used to add weights; parameters with higher disturbance stability parameters have greater changes and therefore receive greater weights, and vice versa. The threshold values ​​for the disturbance stability parameters are set based on specific operating conditions. It should be noted that abnormal response fluctuation values ​​include, for example, predicted value offsets and time delay changes.

[0080] Parameters whose abnormal response fluctuation values ​​are less than a preset fluctuation threshold are considered stable device parameters. The preset fluctuation threshold is the allowable range of parameter fluctuations in the model deployed on the current target smart device, and this threshold is set based on the specific model type. The parameters obtained from this screening are those that have a small impact on the model under various disturbance conditions, which are considered stable device parameters. However, these parameters may still pose risks to certain structures of the model during model operation. Therefore, the subsequent risk assessment of stable device parameters is a second layer of insurance mechanism.

[0081] Methods for risk assessment of stable equipment parameters include:

[0082] Abnormal response trajectories are identified based on the abnormal response fluctuation values ​​of the stable equipment parameters in each round of disturbance simulation. By reverse tracing the corresponding abnormal response fluctuation values ​​output by the stable equipment parameters in each round of disturbance simulation, the abrupt change points of the fluctuation values ​​of the stable equipment parameters at certain moments in the corresponding time period are taken as trajectory change points, and the trajectory change points are integrated to construct the abnormal response trajectory.

[0083] Anomaly analysis is performed on the abnormal response trajectory to obtain abnormal fluctuation characteristics. These abnormal fluctuation characteristics refer to features extracted from the abnormal response trajectory using feature extraction algorithms that can reflect abnormal patterns. These features include, for example, output error, average lag response delay, and confidence level abrupt changes. An abnormal mapping relationship between abnormal fluctuation characteristics and stable equipment parameters is established. By mapping stable equipment parameters to their corresponding abnormal fluctuation characteristics, a causal relationship is formed between stable equipment parameters and their potential risk-causing features, providing a basis for subsequent operations.

[0084] The corresponding model for stable equipment parameters is determined, and a performance sensitivity map of this model under stable equipment parameter conditions is constructed. It should be noted that by using stable equipment parameters as the data foundation, it is necessary to detect whether there are certain models that are extremely sensitive to changes in these stable equipment parameters, and match the model with the highest sensitivity with the corresponding stable equipment parameters. The performance sensitivity map uses each substructure of the model as a node, and the changes in stable equipment parameters between different substructures as edges. Anomaly correlation subgraphs are extracted from the performance sensitivity map based on anomaly mapping relationships. The anomaly correlation subgraphs are extracted from the performance sensitivity map by combining the anomaly mapping relationships between the parameters in the corresponding edges of several related model substructures. For example, a sudden drop in CPU frequency connects the CPU module and several other related modules. These modules and edge relationships are integrated into an anomaly correlation subgraph, which reflects that several substructures in the model have a strong response correlation to certain disturbances.

[0085] The offset degree of each abnormal correlation subgraph is quantified to obtain an offset mapping vector. In this embodiment, the offset value is obtained by calculating the relative change between the perturbation scenario corresponding to the abnormal correlation subgraph and the features of each dimension under the preset undisturbed state. Normalization is used to eliminate the dimensional differences of the offset values ​​of all dimensions, and the offset value is used as an element of the vector to form the offset mapping vector. The offset mapping vector is used to reflect the risk degree of each substructure in the corresponding abnormal correlation subgraph under certain perturbation conditions.

[0086] A risk propagation model is constructed, using the offset mapping vector as the basis for its operation. The risk propagation model is used to calculate the risk score of the stable equipment parameters, and the parameter risk index is constructed based on the risk score. In this embodiment, the risk propagation model is constructed based on a graph neural network model. The offset mapping vector is used as the input feature of the risk propagation model. The risk propagation model is used to evaluate the influence strength of the corresponding parameter on the local area based on the risk level of each local anomaly association subgraph, and outputs the risk score of each stable equipment parameter. The risk scores are then aggregated by local substructure to obtain the parameter risk index of the corresponding substructure.

[0087] The methods for model matching of scheduling intelligent devices include:

[0088] Based on the parameter risk index, affected structures in the current deployment model are identified. Since the parameter risk index reflects the influence strength of several parameters on a certain substructure, the substructures corresponding to the parameters whose risk scores are higher than the preset risk score threshold are identified as affected structures. The functional association information of the affected structures is obtained, and the parameters of the affected structures are adjusted based on the functional association information to obtain the parameter correction model. The functional association information of the affected structures refers to the relevant parameters of the upstream and downstream directly related structures of the identified affected structures, and these parameters can reflect the dependencies and operating logic between several substructures. Within the coverage of the functional association information, in order to ensure that the affected structures do not affect other substructures, the parameters of the affected structures are adjusted using techniques such as channel pruning and parameter fine-tuning to obtain the parameter correction model.

[0089] The power consumption prediction of the parameter correction model is calculated. In this embodiment, the upper and lower limits of the power consumption change of the parameter correction model during the time period in which the intelligent device is located are estimated using a power consumption simulation function model. Combined with the resource status parameters of the intelligent device at any time period, it is determined whether the predicted power consumption value conforms to the preset power consumption range of the intelligent device in that time period. The resource status parameters refer to the real-time status parameters of the intelligent device in a certain time period, such as the current fluctuation amplitude and heat dissipation status. Based on these parameters, a preset power consumption range for the intelligent device in that time period is constructed, and it is determined whether the predicted power consumption value falls within the preset power consumption range.

[0090] If the model matches the specified parameters, it is matched with the scheduling intelligent device for that time period. Otherwise, a secondary adjustment is performed on the parameter correction model. Since the energy consumption requirement may change after parameter adjustment and structural path changes, energy consumption matching between the model and the scheduling intelligent device is still required. Power consumption analysis is performed on the parameter correction model to identify high-energy-consuming structures. In this embodiment, structural heatmap analysis is used to perform power consumption analysis on the parameter correction model to identify structures whose power consumption is higher than a preset threshold, which are high-energy-consuming structures. Hierarchical reconstruction is performed on the high-energy-consuming structures to generate a power-saving parameter model and match it with the corresponding scheduling intelligent device. It should be noted that the hierarchical reconstruction of the high-energy-consuming structure is achieved by performing module-level reorganization or hierarchical pruning of each level. After each hierarchical reconstruction, the power consumption is evaluated. If it is lower than the preset power consumption threshold, it can be matched with the corresponding scheduling intelligent device. The model matching process is structured and the model adjustment strategy is output.

[0091] Methods for optimizing model tuning strategies include:

[0092] A simulated effect expectation value is set. If the simulated effect is less than the expected value, the parameters of the model corresponding to the model adjustment strategy are optimized. The iteration stops when the simulated effect is greater than or equal to the expected value, and the optimized model strategy is output. In this embodiment, the optimization algorithm is used to optimize the parameter combination included in the model adjustment strategy, and the effect of each optimized parameter combination is scored and compared with the simulated effect expectation value. The iteration of the optimization algorithm stops when the simulated effect expectation value is reached. The scoring dimensions for the effect score include running time, strategy accuracy, and power fluctuation.

[0093] This embodiment improves the stability and security of deep learning models running on intelligent devices in resource-constrained environments by constructing a closed loop from device status monitoring to model deployment strategy optimization. Compared with existing technologies, it achieves energy consumption matching between the model and the device by identifying resource levels and regulating energy consumption, enabling synchronous perception of system resources during model deployment and ensuring the model can operate normally under device resource constraints. It simulates disturbances and evaluates results for device status parameters, effectively identifying unstable behaviors that the model may exhibit in certain input scenarios or under high-disturbance environments, and filters corresponding parameters to avoid risks such as model performance crashes caused by certain parameters. Simultaneously, it combines resource status and parameter risk assessment to correct model parameters, and performs multiple rounds of optimization of the model adjustment strategy using optimization algorithms. Therefore, this deep learning optimization method for intelligent devices ensures model accuracy and robustness while maximizing adaptation to the computing power and power consumption requirements of the target intelligent device, improving the overall operational security and scheduling capabilities of the device.

[0094] Example 2

[0095] Please see Figure 2 As shown, parts not described in detail in this embodiment are described in Embodiment 1. A deep learning optimization system for intelligent devices is provided, including:

[0096] The data acquisition module is used to collect the operating status parameters of the target smart device and perform data cleaning on the operating status parameters to obtain accurate device parameters.

[0097] The energy consumption scheduling module is used to identify the resource level of the target intelligent device based on accurate equipment parameters; and to perform energy consumption scheduling on the target intelligent device based on the resource level to obtain the scheduled intelligent device.

[0098] The parameter filtering module is used to simulate disturbances to precise equipment parameters and output disturbance stability indicators; based on the disturbance stability indicators, it filters appropriate parameters for precise equipment parameters and outputs stable equipment parameters.

[0099] The risk assessment module is used to assess the risks of stable equipment parameters and obtain parameter risk indicators.

[0100] The strategy generation module is used to perform model matching on the scheduling intelligent devices and generate model adjustment strategies.

[0101] The strategy optimization module is used to simulate scenarios based on model adjustment strategies, load and optimize the model adjustment strategies based on the simulation results, output optimized model strategies, and apply the optimized model strategies to the target smart devices; the modules are connected to each other via wired and / or wireless means.

[0102] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

[0103] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0104] In the description of this invention, it should be understood that the terms "first," "second," etc., are used only for distinguishing descriptions and should not be construed as indicating or implying relative importance.

[0105] In the description of this invention, unless otherwise stated, "a plurality of" means two or more.

[0106] In the description of this invention, "several" means one or more, and "a large number" means two or more.

[0107] In the description of this specification, references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0108] All formulas in this manual are dimensionless and calculated numerically. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters and thresholds in the formulas are set by those skilled in the art according to the actual situation.

[0109] Although embodiments of the invention have been shown and described, those skilled in the art will understand that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the claims and their equivalents.

Claims

1. A deep learning optimization method for intelligent devices, characterized in that, include: S1. Collect the operating status parameters of the target intelligent device, and perform data cleaning on the operating status parameters to obtain accurate device parameters; S2. Identify the resource level of the target intelligent device based on precise device parameters, including: acquiring task scenario information; dividing the precise device parameters into time periods based on the task scenario information to obtain time-segmented device parameters; performing statistical feature analysis on the time-segmented device parameters of any time period and outputting state feature items; calculating the resource evolution amount between different time periods based on the state feature items, and constructing change trend data based on the resource evolution amount; The parameters belonging to each resource dimension in the trend data are weighted by features to obtain the resource status score vector for the corresponding time period; the number of abnormal situations in the corresponding time period is queried, an abnormal fluctuation coefficient is constructed based on the number of abnormal situations, and the abnormal fluctuation coefficient is multiplied with the resource status score vector to output the fluctuation status score vector; a resource level scoring matrix is ​​constructed, and the fluctuation status score vector is matched with the resource level scoring matrix to output the resource level; Energy consumption scheduling of target intelligent devices is performed based on resource levels to obtain the scheduled intelligent devices; S3. Perform disturbance simulation on precise equipment parameters and output disturbance stability index; based on the disturbance stability index, perform appropriate parameter selection on precise equipment parameters and output stable equipment parameters; S4. Conduct a risk assessment of the stable equipment parameters to obtain parameter risk indicators; S5. Based on parameter risk indicators, perform model matching on the scheduling intelligent devices and generate model adjustment strategies, including: Identify affected structures in the current deployment model based on parameter risk indicators; obtain functional association information of the affected structures, and adjust the parameters of the affected structures based on the functional association information to obtain a parameter correction model; calculate the predicted power consumption of the parameter correction model. Based on the resource status parameters of the scheduling intelligent device for any given time period, determine whether the predicted operating power consumption value conforms to the stable operating power consumption range of the scheduling intelligent device within that time period. If it does, match the parameter correction model with the scheduling intelligent device for that time period; otherwise, perform secondary adjustments to the parameter correction model. Perform power consumption analysis on the parameter correction model to identify high-energy-consuming structures. Perform hierarchical reconstruction on the high-energy-consuming structures to generate a power-reduction parameter model and match it with the corresponding scheduling intelligent device. Structurally output the model adjustment strategy from the model matching process. S6. Simulate the scene based on the model adjustment strategy, load and optimize the model adjustment strategy based on the simulation effect, output the optimized model strategy, and apply the optimized model strategy to the target smart device.

2. The deep learning optimization method for intelligent devices according to claim 1, characterized in that, The method of energy consumption scheduling for target intelligent devices based on resource levels includes: Based on resource level, resource control points are marked for target intelligent devices, and resource control point labels are output. Based on the resource control point labels, the key control structures of the target intelligent devices are identified, and the regulation fields of the key control structures are deployed. The structure control function is constructed, and the specific parameter values ​​of the regulation fields are calculated using the structure control function. Based on resource level, a preset deep learning model library is screened, and the screened models are matched with the regulated target intelligent devices for energy consumption to obtain the scheduled intelligent devices.

3. The deep learning optimization method for intelligent devices according to claim 2, characterized in that, The methods for deploying control fields in key control structures include: For each critical control structure, define corresponding control fields and limit the energy consumption boundary of the critical control structure based on the control fields; deploy the control fields to the relevant configuration interfaces of the critical control structure. The methods for matching the energy consumption of the screened model with the adjusted target smart device include: The energy consumption of the target smart devices after regulation is assessed over time periods, and a stable operating power consumption range is set based on the assessment results. The load standard index of the screened model is obtained, and the stable operating power consumption range is compared with the load standard index. If the load standard index is within the stable operating power consumption range of the time period, the screened model is deployed to the target smart devices during this time period. If the energy consumption of the target smart devices changes over time, the model is hot-switched, and a model whose load standard index conforms to the corresponding stable operating power consumption range of this time period is loaded.

4. The deep learning optimization method for intelligent devices according to claim 3, characterized in that, The methods for simulating disturbances to precise equipment parameters include: Modeling of intelligent device operation scenarios based on accurate device parameters; constructing device disturbance templates and injecting dynamic disturbance events into the intelligent device operation scenario modeling based on the device disturbance templates; Collect response change information and final simulation feedback data during the disturbance simulation process, and construct the disturbance response path for each precise device parameter from the time series dimension based on the response change information and final simulation feedback data; extract local disturbance segments in the disturbance response path, perform response feature analysis on the local disturbance segments, and output local disturbance response feature vectors; The local stability quantity of each local disturbance response feature vector is calculated using a stability function; all local disturbance response feature vectors are concatenated according to the time sequence to output the overall disturbance feature vector; the macroscopic stability quantity of the overall disturbance feature vector is calculated; the local stability quantities are weighted and summed and added to the macroscopic stability quantity to output the disturbance stability parameters; and all disturbance stability parameters are integrated to construct the disturbance stability index.

5. The deep learning optimization method for intelligent devices according to claim 4, characterized in that, The method for selecting appropriate parameters for precise equipment parameters based on disturbance stability indices includes: The abnormal response fluctuation values ​​of precise equipment parameters in any round of disturbance simulation are calculated based on the disturbance stability index; parameters with abnormal response fluctuation values ​​less than the preset fluctuation threshold are taken as stable equipment parameters.

6. The deep learning optimization method for intelligent devices according to claim 5, characterized in that, The methods for risk assessment of stable equipment parameters include: Abnormal response trajectories are identified based on the abnormal response fluctuation values ​​of each round of disturbance simulation using stable equipment parameters; abnormal response trajectories are analyzed to obtain abnormal fluctuation characteristics; and an abnormal mapping relationship between abnormal fluctuation characteristics and stable equipment parameters is established. Determine the corresponding model for stable equipment parameters and construct a performance sensitivity map of the model under stable equipment parameter conditions; extract anomaly correlation subgraphs from the performance sensitivity map based on anomaly mapping relationships; quantify the offset degree of each anomaly correlation subgraph to obtain an offset mapping vector; construct a risk propagation model, use the offset mapping vector as the basis for the operation of the risk propagation model, calculate the risk score of stable equipment parameters using the risk propagation model, and construct parameter risk indicators based on the risk score.

7. The deep learning optimization method for intelligent devices according to claim 6, characterized in that, The methods for loading and optimizing the model adjustment strategy include: Set the expected value of the simulation effect. If the simulation effect is less than the expected value, optimize the parameter configuration of the model corresponding to the model adjustment strategy. Stop iterating when the simulation effect is greater than or equal to the expected value and output the optimized model strategy.

8. A deep learning optimization system for an intelligent device, used to implement the deep learning optimization method for an intelligent device according to any one of claims 1-7, characterized in that, include: The data acquisition module is used to collect the operating status parameters of the target smart device and perform data cleaning on the operating status parameters to obtain accurate device parameters. The energy consumption scheduling module is used to identify the resource level of target intelligent devices based on accurate equipment parameters; Energy consumption scheduling of target intelligent devices is performed based on resource levels to obtain the scheduled intelligent devices; The parameter filtering module is used to simulate disturbances to precise equipment parameters and output disturbance stability indicators; based on the disturbance stability indicators, it filters appropriate parameters for precise equipment parameters and outputs stable equipment parameters. The risk assessment module is used to assess the risks of stable equipment parameters and obtain parameter risk indicators. The strategy generation module is used to perform model matching on the scheduling intelligent devices and generate model adjustment strategies. The strategy optimization module is used to simulate scenarios based on model adjustment strategies, load and optimize the model adjustment strategies based on the simulation results, output optimized model strategies, and apply the optimized model strategies to the target smart devices; the modules are connected to each other via wired and / or wireless means.

Images