Deep learning optimization system and method for intelligent equipment

Abstract

The invention belongs to the technical field of data processing, and discloses a deep learning optimization system and method for intelligent equipment. Comprising the steps of collecting data and performing data cleaning to obtain accurate equipment parameters; identifying a target intelligent device resource level based on the accurate device parameters; performing energy consumption scheduling on the target intelligent device based on the resource level to obtain a scheduled intelligent device; performing disturbance simulation on the accurate equipment parameters, and outputting a disturbance stability index; appropriate parameter screening is carried out on the accurate equipment parameters based on the disturbance stability index, and stable equipment parameters are output; performing risk assessment on the stable equipment parameters to obtain parameter risk indexes; performing model matching on the scheduling intelligent equipment based on the parameter risk index to generate a model adjustment strategy; performing scene simulation based on the model adjustment strategy, performing loading optimization on the model adjustment strategy based on a simulation effect, and outputting an optimization model strategy; the efficient deep learning optimization method for the intelligent equipment is realized.

Description

Technical Field

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

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

[0003] In the actual operating conditions of smart devices, the control resource-related parameters of the device are highly time-varying. For example, parameters such as power consumption, memory usage, and system temperature will change with the task type or task execution stage. At this time, the resource consumption of the deep learning model deployed in the device may no longer match the power consumption threshold preset by the device at a certain task stage. Existing technologies often lack real-time monitoring and adjustment of energy consumption, resulting in a mismatch between model operation and the actual device resource status. For example, continuously loading a high-precision and high-load model under high-temperature conditions may cause the device to enter frequency reduction or the model to stop running, thereby causing task interruption. situation; in addition, existing model deployment methods generally ignore the instability and high risk that the input data itself may exhibit during actual operation. For example, after a certain model receives certain feature input data, its performance may degrade or crash. At the same time, existing technologies lack the necessary risk assessment before deploying models for such unstable input data, and are unable to identify whether certain input data is more likely to trigger model output errors. For example, in a certain vehicle recognition system, when the car windows are heavily fogged on rainy days, the background texture loss intensity of the forward-facing image exceeds the system's tolerable 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 smart devices to solve the above problems. Summary of the Invention

[0005] In order to overcome the above-mentioned defects of the prior art and to achieve the above-mentioned objectives, the present invention provides the following technical solution: a deep learning optimization method for smart devices, comprising:

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

[0007] S2. Identify the resource level of the target smart device based on precise device parameters; schedule the energy consumption of the target smart device based on the resource level to obtain the scheduled smart device;

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

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

[0010] S5. Perform model matching on dispatching intelligent devices based on parameter risk indicators and generate model adjustment strategies;

[0011] S6. Perform scenario simulation based on the model adjustment strategy, load and optimize the model adjustment strategy based on the simulation results, output the optimized model strategy, and apply the optimized model strategy to the target smart device.

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

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

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

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

[0016] The resource control points of the target smart devices are marked based on the resource level, and the resource control point labels are output; the key control structures of the target smart devices are identified based on the resource control point labels, and the control fields of the key control structures are deployed; a structural control function is constructed, and the specific parameter values ​​of the control fields are calculated using the structural control function; the preset deep learning model library is screened based on the resource level, and the energy consumption of the screened models is matched with the target smart devices after control to obtain the scheduling smart devices.

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

[0018] Set the control fields of the corresponding functions for each key control structure, and limit the energy consumption boundary of the key control structure based on the control fields; deploy the control fields to the relevant configuration interface of the key control structure;

[0019] Methods for matching the energy consumption of the screened model with the regulated target smart device include:

[0020] The energy consumption of the target smart device after control is evaluated during each period, and a stable operating power consumption range is set based on the evaluation results; the load standard index of the filtered 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 period, the filtered model is deployed to the target smart device of the period; if the energy consumption of the target smart device changes with the period, the model is hot-swapped to load a model whose load standard index meets the stable operating power consumption range corresponding to the period.

[0021] Furthermore, the method of performing disturbance simulation on precise device parameters includes:

[0022] Build smart device operation scenario modeling based on precise device parameters; construct device disturbance templates, and inject dynamic disturbance events into smart device operation scenario modeling based on the device disturbance templates;

[0023] Collect the response change information and final simulation feedback data of the disturbance simulation process, and construct the disturbance response path of each precise device parameter from the time series dimension based on the response change information and final simulation feedback data; extract the local disturbance fragments in the disturbance response path, analyze the response characteristics of the local disturbance fragments, and output the local disturbance response feature vector;

[0024] The local stability of each local disturbance response eigenvector is calculated using the stability function; all local disturbance response eigenvectors are vector-concatenated based on the time sequence to output the overall disturbance eigenvector; the macroscopic stability of the overall disturbance eigenvector is calculated; the local stability is weightedly summed and added to the macroscopic stability to output the disturbance stability parameter; all disturbance stability parameters are integrated to construct the disturbance stability index.

[0025] Furthermore, the method of screening appropriate parameters for precise equipment parameters based on the disturbance stability index includes:

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

[0027] Furthermore, the method of performing risk assessment on stable equipment parameters includes:

[0028] Identify abnormal response trajectories based on abnormal response fluctuation values ​​from each round of disturbance simulation of stable equipment parameters; perform anomaly analysis on the abnormal response trajectories to obtain abnormal fluctuation characteristics; and establish an abnormal mapping relationship between abnormal fluctuation characteristics and stable equipment parameters;

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

[0030] Furthermore, the method of performing model matching on the scheduling smart device includes:

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

[0032] Combined with the resource status parameters of the scheduling intelligent device in any period, determine whether the operating power consumption prediction value is consistent with the stable operating power consumption range of the scheduling intelligent device in the period; if it is consistent, match the parameter correction model with the scheduling intelligent device in the period, otherwise make secondary adjustments to the parameter correction model; perform power consumption analysis on the parameter correction model to identify high-energy consumption structures; perform hierarchical reconstruction on the high-energy consumption structure to generate a power-saving parameter model and match it with the corresponding scheduling intelligent device; structure the model matching process to output a model adjustment strategy.

[0033] Furthermore, the method of 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 until the simulation effect is greater than or equal to the expected value. Then stop the iteration and output the optimized model strategy.

[0035] A deep learning optimization system for smart devices, which is used to implement a deep learning optimization method for smart devices, is characterized by comprising:

[0036] The data acquisition module is used to collect the operating status parameters of the target intelligent 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 smart device based on the precise device parameters; schedule the energy consumption of the target smart device based on the resource level to obtain the scheduled smart device;

[0038] Parameter screening module, used to simulate disturbances on precise equipment parameters and output disturbance stability indicators; based on the disturbance stability indicators, it screens appropriate parameters for precise equipment parameters and outputs stable equipment parameters;

[0039] Risk assessment module, used to conduct risk assessment on 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 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; each module is connected by wired and / or wireless means.

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

[0043] By building a closed loop from device status monitoring to model deployment strategy optimization, the stability and security of deep learning models running on smart devices in resource-constrained environments are improved. Compared with existing technologies, the energy consumption level of the model and the device are matched by identifying resource levels and regulating energy consumption, and the synchronous perception of system resources when deploying the model is achieved, ensuring that the model can operate normally under device resource constraints. The device status parameters are subjected to disturbance simulation and result evaluation, effectively identifying the unstable behavior that is prone to occur in the model in certain specific input scenarios or high-disturbance environments, and screening the corresponding parameters to avoid the risk of certain parameters causing model performance crashes. At the same time, the resource status and parameter risk conditions are combined to realize the parameter correction of the model, and the model adjustment strategy is optimized for multiple rounds in combination with the optimization algorithm. Therefore, this deep learning optimization method for smart devices, while ensuring the accuracy and robustness of the model, adapts to the computing power and power consumption requirements of the current target smart device to the greatest extent, thereby improving the overall operational safety and scheduling capabilities of the device. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 A schematic diagram of a deep learning optimization method for an intelligent device according to the present invention;

[0045] Figure 2 Schematic diagram of a deep learning optimization system for intelligent devices of the present invention. DETAILED DESCRIPTION

[0046] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0047] Example 1

[0048] See also Figure 1 As shown, the deep learning optimization method for a smart device described in this embodiment includes:

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

[0050] S2. Identify the resource level of the target smart device based on precise device parameters; schedule the energy consumption of the target smart device based on the resource level to obtain the scheduled smart device;

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

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

[0053] S5. Perform model matching on dispatching intelligent devices based on parameter risk indicators and generate model adjustment strategies;

[0054] S6. Perform scenario simulation based on the model adjustment strategy, load and optimize the model adjustment strategy based on the simulation results, output the optimized model strategy, and apply the optimized model strategy to the target smart device.

[0055] The operating status parameters include parameters such as power, current value, operating temperature, CPU occupancy and operating load that can reflect the operating status of the target smart device; in this embodiment, the data cleaning process is achieved by filling missing values, isolating outliers and standardizing the operating status parameters to obtain higher quality data, which facilitates subsequent data processing.

[0056] Methods for identifying the resource level of target smart devices based on precise device parameters include:

[0057] Obtain task scenario information, and divide the precise device parameters into time periods based on the task scenario information to obtain time period device parameters, wherein the task scenario information refers to the type of task being executed corresponding to the precise device parameters, and construct a sliding window with an adjustable window size. The sliding window is used to realize time period division based on the timestamp and task type of the precise device parameters; wherein the time period division uses the task switching of the smart device as a mark point in this embodiment to ensure that the time period device parameters are representative of the stage, which is convenient for the subsequent processing of the parameters of each time period; perform statistical feature analysis on the time period device parameters of any time period, and output state feature items, wherein the device parameter set of each time period is statistically calculated to obtain characteristics that can reflect the data distribution in the time period, such as the mean, median, standard deviation and extreme value; these features are combined into state feature items to intuitively reflect the load intensity of the target smart device in various resource dimensions such as power, storage and memory in a certain time period.

[0058] The resource evolution amount between different time periods is calculated based on the state characteristic items, and the change trend data is constructed based on the resource evolution amount. The state characteristic items of multiple adjacent time periods are used as input, and the change amount information that changes over time in each resource dimension, such as the parameter change rate, average change amplitude, and change increase or decrease direction, is calculated, which is the resource evolution amount. In this embodiment, a regression model is used to fit the resource change curve based on the resource evolution amount, and the parameters corresponding to each time stamp and the resource evolution amount of each section are marked on the curve to obtain the change trend data.

[0059] The parameters belonging to each resource dimension in the change trend data are feature-weighted to obtain the resource status score vector of the corresponding time period. By setting corresponding weights for the parameters of each resource dimension in the corresponding time period of the change trend data, the mean of these parameters in the entire time period is calculated and weighted to obtain the resource status score vector; the setting of the weight is adjusted based on the specific working conditions. In this embodiment, the key parameters that can intuitively reflect the resource changes, such as the change in CPU utilization and the change rate of power consumption, are given greater weights, and other auxiliary parameters such as the change in Bluetooth interface power consumption, which have little impact on the intelligent target device, are given smaller weights. The value range of the weight is .

[0060] Query the number of abnormal conditions in the corresponding time period, construct an abnormal fluctuation coefficient based on the number of abnormal conditions, and multiply the abnormal fluctuation coefficient with the resource status score vector to output the fluctuation status score vector. Abnormal conditions include sudden temperature rise and sudden change in CPU usage. The abnormal fluctuation coefficient is constructed based on the number of abnormal conditions to indicate the degree of instability in the time period. The abnormal fluctuation coefficient is set based on specific working 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 judgments.

[0061] A resource level scoring matrix is ​​constructed, and the fluctuation state score vector is interval-matched with the resource level scoring matrix to output the resource level. The resource level scoring matrix sets scoring categories including 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, and the preset resource level matching template is used to identify the comprehensive matching level of each fluctuation state score vector, that is, the resource level of the target smart device in the corresponding time period, which is used for subsequent energy consumption scheduling of the target smart device.

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

[0063] The target smart device is marked with resource control points based on the resource level, and a resource control point label is output. The resource level is used to determine the resource control point where resource consumption is significant in the target smart device during the corresponding period. This point is used to reflect the trigger point when the device performs certain operations; resource control point labels include CPU frequency change points and interface occupancy detection points.

[0064] The key control structure of the target smart device is identified based on the resource control point tag, and the control field is deployed for the key control structure. The key control structure is the specific physical structure responsible for executing device operations, such as the frequency adjustment module, transmission module, and interface allocation module, which are associated with each resource control point tag and identified based on the resource control point tag. The control field is used to limit the resource consumption of the key control structure in any time period.

[0065] Construct a structural control function, and use the structural control function to calculate the specific parameter values ​​of the regulation fields. In this embodiment, the basic structure of the structural control function is a multi-factor weighted model. By taking the resource level label and the parameters of each dimension in the historical records of the target smart device as input data, the specific parameter values ​​of the regulation fields of the structure that needs to be regulated are calculated; based on the resource level, the preset deep learning model library is screened, and the energy consumption of the screened model is matched with the regulated target smart device to obtain a scheduling smart device, wherein the preset deep learning model library includes a variety of deep learning models, and the structural complexity of the deep learning model is matched with the load capacity reflected by the resource level, and the models whose structural complexity exceeds the adaptability of the current resource level are eliminated, and the preliminary screening of the deep learning model is completed to ensure that the screened model can run under the corresponding resource level conditions.

[0066] Methods for deploying control fields for key control structures include:

[0067] A control field of a corresponding function is set for each key control structure, and the energy consumption boundary of the key control structure is limited based on the control field, where the control field includes fields such as a maximum frequency field, a thread number upper limit field, a standby interval field, and a power allocation field, which are associated with the operations performed by the key control structure; the control field is used to constrain indicators affecting energy consumption, such as the scheduling upper limit and working cycle of the key control structure in the corresponding time period; 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 smart device in the form of computer instructions.

[0068] Methods for matching the energy consumption of the screened model with the regulated target smart device include:

[0069] The energy consumption of the target smart device after regulation is evaluated in each time period, 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 in the corresponding time period are collected, including parameters reflecting the device load such as power supply temperature, current fluctuation value and memory occupancy rate; by statistically analyzing the energy consumption parameters of each time period 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 used as the stable operating power consumption range.

[0070] Obtain the load standard index of the filtered model, compare the stable operation power consumption range with the load standard index, and if the load standard index is within the stable operation power consumption range of the period, deploy the filtered model to the target smart device of the period, where the load standard index includes parameters such as single operation power consumption, memory occupancy, and current peak corresponding to the model; compare the load standard index with the stable operation power consumption range item by item. If all parameters in the index fall within the stable operation power consumption range, it is considered that the model matches the power consumption capability of the target smart device of the current period, and deployment and operation are allowed.

[0071] If the energy consumption of the target smart device changes over time, the model is hot-switched, and the load standard index is loaded in accordance with the model of the stable operating power consumption range corresponding to this time period. Since the system power consumption will change over time when the target smart device is running under actual working conditions, the deployed model will no longer match the energy consumption requirements of the changed target smart device. Therefore, in this embodiment, the model hot-switching is implemented using an imperceptible replacement mechanism 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 smart device.

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

[0073] Based on precise device parameters, intelligent device operation scenario modeling is constructed. In this embodiment, precise device parameters are used as the basis to construct intelligent device operation scenario modeling in a virtual environment, including information such as processor configuration parameters, module scheduling and power consumption changes, where the scenario modeling refers to a simulation scenario in a certain period of time; a device disturbance template is constructed, and dynamic disturbance events are injected into the intelligent device operation scenario modeling based on the device disturbance template, where the device disturbance template is a parameter configuration scheme based on historical device operation records and existing data, which includes multiple parameters that can trigger model performance collapse or instability reactions; including disturbance types such as CPU frequency switching disturbance, power supply current fluctuation and memory write obstruction.

[0074] The response change information and final simulation feedback data of the disturbance simulation process are collected, and the disturbance response path of each precise device parameter is constructed from the timing dimension based on the response change information and the final simulation feedback data. The response change information includes data extracted during the disturbance simulation process that can reflect the parameter changes affected by the disturbance, such as the number of voltage callbacks, the number of thread migrations, and the temperature change rate. The final simulation feedback data refers to the result value of the parameter change obtained at the end of the simulation.

[0075] The local disturbance fragments in the disturbance response path are extracted, and the response characteristics of the local disturbance fragments are analyzed to output the local disturbance response feature vector. The complete disturbance response path is decomposed into several local disturbance fragments based on the timing dimension. The local disturbance response feature vector obtained by the response characteristic analysis includes elements such as response frequency, time delay and fluctuation amplitude.

[0076] The local stability value of each local disturbance response eigenvector is calculated using a stability function. In this embodiment, the stability function refers to a disturbance reconstruction residual function, wherein the calculation formula of the stability function is: ;in, represents the local stability amount; represents the average stability score, represents the maximum risk score, represents the cumulative volatility score; Indicates the deviation between the segment corresponding to the local perturbation response eigenvector and the benchmark segment, timestamp representing the local perturbation segment; ;in, represents the segment corresponding to the local perturbation response eigenvector; Indicates passing The benchmark segment obtained by smooth fitting; 、 and Represent the weights corresponding to the average stability score, maximum risk score, and cumulative volatility score, respectively, which are set based on expert experience. , , The local stability value calculated by this function is the stability score of the local disturbance segment, which reflects the stability of the segment under disturbance.

[0077] All local disturbance response feature vectors are concatenated based on the time sequence to output the overall disturbance feature vector, where the overall disturbance feature vector refers to a complete cycle and a vector corresponding to the complete disturbance response path, which is used to reflect the complete disturbance coverage; the macroscopic stability of the overall disturbance feature vector is calculated. In this embodiment, the trajectory disturbance degree function is used to calculate the macroscopic stability, where the calculation formula of the macroscopic stability is: ;in represents the macroscopic stability quantity; represents the standard deviation of each parameter within the complete disturbance response path; It represents the mean slope of the response curves of different local segments in the complete disturbance response path; Indicates the frequency at which intelligent devices such as voltage callback adjust themselves in the event of disturbances in the complete disturbance response path; 、 and Represent the weights corresponding to the standard deviation, slope mean, and self-adjustment frequency, which are set based on expert experience. , , ; The macroscopic stability quantity is used to quantify the degree to which the overall response path is disturbed at the structural level; the local stability quantities are weighted and summed and added to the macroscopic stability quantity to output the disturbance stability parameter, and all disturbance stability parameters are integrated to construct the disturbance stability index, where the disturbance stability parameter quantifies the stability of a certain parameter during the disturbance process, which facilitates the subsequent parameter screening based on the stability; the disturbance stability index is a row matrix, and any element in the matrix is ​​any disturbance stability parameter. The disturbance stability index is obtained by combining all disturbance stability parameters in the same matrix; for example, a disturbance stability parameter is set to , and by adjusting the tail number to represent other disturbance stability parameters, the obtained disturbance stability index is ,in represents the number of disturbance stabilizing parameters.

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

[0079] Based on the disturbance stability index, the abnormal response fluctuation value of the precise equipment parameters in any round of disturbance simulation is calculated. By comparing the parameter changes before and after each round of disturbance simulation, and combining the disturbance stability parameters of each dimensional parameter in the disturbance stability index, the abnormal response fluctuation value of each dimensional parameter in a single disturbance simulation is obtained. The disturbance stability parameter is used to add weights. The higher the disturbance stability parameter, the greater the degree of change of the corresponding parameter, so the additional weight is greater, and vice versa. The value threshold of the disturbance stability parameter is set based on the specific working conditions. It should be noted that the abnormal response fluctuation value includes, for example, the predicted value offset and the delay change.

[0080] Parameters with abnormal response fluctuation values ​​less than the preset fluctuation threshold are regarded as stable device parameters, where the preset fluctuation threshold is the allowable parameter fluctuation range in the model deployed on the current target smart device, and the threshold is set based on the specific model type. The parameters obtained through this screening have a smaller impact on the model under various disturbance conditions, which are stable device parameters. However, these parameters may still generate risks with certain structures of the model when the model is running. Therefore, the subsequent risk assessment of the stable device parameters belongs to the second-level insurance mechanism.

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

[0082] The abnormal response trajectory is identified based on the abnormal response fluctuation values ​​of the stable device parameters in each round of disturbance simulation. By reversely tracing the corresponding abnormal response fluctuation values ​​output by the stable device parameters during each round of disturbance simulation, the fluctuation value mutation points of the stable device parameters at certain moments in the corresponding period are used 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, where abnormal fluctuation characteristics refer to features that can reflect abnormal patterns extracted from the abnormal response trajectory using feature extraction algorithms; including features such as output error, average lag response delay and confidence mutation; an abnormal mapping relationship between abnormal fluctuation characteristics and stable equipment parameters is established, and a mapping is established between stable equipment parameters and corresponding abnormal fluctuation characteristics to form a causal relationship between stable equipment parameters and their potential risk-inducing features, providing a basis for subsequent operations.

[0084] Determine the corresponding model of the stable device parameters, and construct a performance sensitivity map of the model under the conditions of stable device parameters. It should be noted that by taking the stable device parameters as the data basis, detect whether there are some models that are extremely sensitive to changes in these stable device parameters, and match the most sensitive model with the corresponding stable device parameters; the performance sensitivity map uses each substructure of the model as a node, and the changes in stable device parameters between different substructures as edges; based on the abnormal mapping relationship, extract the abnormal correlation subgraph from the performance sensitivity map, wherein the abnormal correlation subgraph is extracted from the performance sensitivity map by combining several related model substructures with the abnormal mapping relationship between the parameters in the corresponding edges; for example, a sudden drop in CPU frequency connects the CPU module and several other related modules, then these modules and edge relationships are integrated into an abnormal correlation subgraph, reflecting that several substructures in the model have 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 disturbance scenario corresponding to the abnormal correlation subgraph and the characteristics of each dimension in the preset non-disturbance state. Normalization is used to eliminate the dimensional difference of the offset values ​​of all dimensions, and the offset values ​​are used as elements of the vector to form an offset mapping vector. The offset mapping vector is used to reflect the risk level of each substructure in the corresponding abnormal correlation subgraph under certain disturbance conditions.

[0086] A risk propagation model is constructed, and the offset mapping vector is used as the operating basis of the risk propagation model. The risk score of the stable equipment parameter is calculated using the risk propagation model, and the parameter risk index is constructed based on the risk score. In this embodiment, the risk propagation model is constructed based on the graph neural network model, and the offset mapping vector is used as the input feature of the risk propagation model. The risk propagation model is used to evaluate the impact intensity of the corresponding parameter on the local area based on the risk degree of each local abnormal correlation subgraph, and the risk score of each stable equipment parameter is output. The risk score is then aggregated in units of local substructures to obtain the parameter risk index of the corresponding substructure.

[0087] Methods for model matching for dispatching intelligent devices include:

[0088] Based on the parameter risk index, the affected structures in the current deployment model are identified. Since the parameter risk index reflects the influence intensity of several parameters on a substructure, the substructure corresponding to the relevant parameters in the parameter risk index whose risk scores are higher than the preset risk score threshold is determined to be the affected structure; the functional association information of the affected structure is obtained, and the parameters of the affected structure are adjusted based on the functional association information to obtain a parameter correction model, where the functional association information of the affected structure refers to the relevant parameters of the upstream and downstream directly associated structures of the identified affected structure, and such parameters can reflect the dependency relationship and operation logic between several substructures; within the coverage of the functional association information, in order to ensure that the affected structure does not affect other substructures, the parameters of the affected structure are adjusted using technical means such as channel pruning and parameter fine-tuning to obtain a parameter correction model.

[0089] Calculate the operating power consumption prediction value of the parameter correction model. In this embodiment, the power consumption simulation function model is used to estimate the upper and lower limits of the power consumption change of the parameter correction model in the time period in which the scheduled smart device is located at this time; combine the resource status parameters of any time period of the scheduled smart device to determine whether the operating power consumption prediction value meets the preset power consumption range of the scheduled smart device in the time period, where the resource status parameters refer to the real-time status parameters of the scheduled smart device in a certain time period, such as the current fluctuation amplitude and heat dissipation status; based on these parameters, construct the preset power consumption range of the scheduled smart device belonging to the time period, and determine whether the operating power consumption prediction value falls within the preset power consumption range.

[0090] If the conditions are met, the parameter correction model is matched with the scheduling intelligent device of the time period. Otherwise, the parameter correction model is adjusted at a secondary level. Since the energy consumption demand may change after the model has undergone operations such as parameter adjustment and structural path change, it is still necessary to perform energy consumption matching between the model and the scheduling intelligent device. The parameter correction model is subjected to power consumption analysis to identify high-energy consumption structures. In this embodiment, the parameter correction model is subjected to power consumption analysis using structural heat map analysis to identify structures whose power consumption is higher than a preset threshold, i.e., high-energy consumption structures. Hierarchical reconstruction is performed on the high-energy consumption structure 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 structure is achieved by performing module-level reorganization on the high-energy consumption structure or performing hierarchical trimming on each level therein. The power consumption is evaluated after each hierarchical reconstruction. 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 to output a model adjustment strategy.

[0091] Ways to optimize loading of model adjustment strategies include:

[0092] An expected value of the simulation effect is set. If the simulation effect is less than the expected value of the simulation effect, the parameter configuration of the model corresponding to the model adjustment strategy is optimized until the simulation effect is greater than or equal to the expected value of the simulation effect, then the iteration is stopped and the optimized model strategy is output. In this embodiment, an optimization algorithm is used to optimize the parameter combination included in the model adjustment strategy, and an effect score is performed on each optimized parameter combination, and it is compared with the expected value of the simulation effect until the expected value of the simulation effect is reached, and then the iteration of the optimization algorithm is stopped; the scoring dimensions for the effect scoring include running time, strategy accuracy, and power fluctuation.

[0093] This embodiment improves the stability and security of deep learning models running on smart devices in resource-constrained environments by constructing a closed loop from device status monitoring to model deployment strategy optimization. Compared with the existing technology, the model and device are matched at the energy consumption level by identifying resource levels and regulating energy consumption, and the synchronous perception of system resources when deploying the model is achieved, ensuring that the model can operate normally under device resource constraints. The device status parameters are subjected to disturbance simulation and result evaluation, effectively identifying the unstable behavior that is prone to occur in the model in certain specific input scenarios or high-disturbance environments, and screening the corresponding parameters to avoid the risk of certain parameters causing model performance crashes. At the same time, the model parameters are corrected in combination with the resource status and parameter risk conditions, and the model adjustment strategy is optimized for multiple rounds in combination with the optimization algorithm. Therefore, this deep learning optimization method for smart devices, while ensuring the accuracy and robustness of the model, adapts to the computing power and power consumption requirements of the current target smart device to the greatest extent, thereby improving the overall operational safety and scheduling capabilities of the device.

[0094] Example 2

[0095] See also Figure 2 As shown, for the parts not described in detail in this embodiment, please refer to the description of Example 1. A deep learning optimization system for smart devices is provided, including:

[0096] The data acquisition module is used to collect the operating status parameters of the target intelligent 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 smart device based on the precise device parameters; schedule the energy consumption of the target smart device based on the resource level to obtain the scheduled smart device;

[0098] Parameter screening module, used to simulate disturbances on precise equipment parameters and output disturbance stability indicators; based on the disturbance stability indicators, it screens appropriate parameters for precise equipment parameters and outputs stable equipment parameters;

[0099] Risk assessment module, used to conduct risk assessment on 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 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; each module is connected by wired and / or wireless means.

[0102] The foregoing 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 will be able to modify the technical solutions described in the foregoing embodiments or to substitute equivalents for some of the technical features. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention shall be included within the scope of protection of the present invention.

[0103] It should be noted that, in this document, the terms "comprises," "includes," or any other variations thereof are intended to encompass non-exclusive inclusion, such that a process, method, article, or apparatus comprising a series of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such process, method, article, or apparatus. In the absence of further limitations, an element defined by the phrase "comprising a ..." does not exclude the presence of other identical elements in the process, method, article, or apparatus comprising the element.

[0104] In the description of the present invention, it should be understood that the terms "first", "second", etc. are only used to distinguish the descriptions and cannot be understood as indicating or implying relative importance.

[0105] In the description of the present invention, unless otherwise specified, "plurality" means two or more.

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

[0107] Throughout this specification, reference to terms such as "one embodiment," "some embodiments," "examples," "specific examples," or "some examples" means that a specific feature, structure, material, or characteristic described in conjunction with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, schematic representations 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 any one or more embodiments or examples.

[0108] The formulas in this manual are all dimensionless and calculated using numerical values. The formulas are obtained by collecting a large amount of data and performing software simulation to obtain the most recent real situation. The preset parameters and thresholds in the formulas are set by technicians in this field based on actual conditions.

[0109] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to the embodiments without departing from the principles and spirit of the invention, and that the scope of the invention 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 smart device and perform data cleaning on the operating status parameters to obtain accurate device parameters; S2. Identify the resource level of the target smart device based on precise device parameters; schedule the energy consumption of the target smart device based on the resource level to obtain the scheduled smart device; S3. Perform disturbance simulation on the precise device parameters and output disturbance stability index; based on the disturbance stability index, perform appropriate parameter screening on the precise device parameters and output stable device parameters; S4. Conduct risk assessment on stable equipment parameters to obtain parameter risk indicators; S5. Perform model matching on dispatching intelligent devices based on parameter risk indicators and generate model adjustment strategies; S6. Perform scenario simulation based on the model adjustment strategy, load and optimize the model adjustment strategy based on the simulation results, output the optimized model strategy, and apply the optimized model strategy to the target smart device.

2. The deep learning optimization method for smart devices according to claim 1, characterized in that: The method of identifying the resource level of the target smart device based on precise device parameters includes: Obtain mission scenario information, divide precise equipment parameters into time periods based on the mission scenario information, and obtain time-divided equipment parameters; perform statistical feature analysis on the time-divided equipment parameters of any time period and output state feature items; calculate the resource evolution between different time periods based on the state feature items, and construct change trend data based on the resource evolution; The parameters belonging to each resource dimension in the change trend data are feature-weighted to obtain the resource status score vector for the corresponding time period; the number of abnormal conditions in the corresponding time period is queried, and the abnormal fluctuation coefficient is constructed based on the number of abnormal conditions. The abnormal fluctuation coefficient is multiplied by the resource status score vector to output the fluctuation status score vector; a resource grade scoring matrix is ​​constructed, and the fluctuation status score vector is interval-matched with the resource grade scoring matrix to output the resource grade.

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

4. The deep learning optimization method for smart devices according to claim 3, characterized in that: The method of deploying the control fields of the key control structure includes: Set the control fields of the corresponding functions for each key control structure, and limit the energy consumption boundary of the key control structure based on the control fields; deploy the control fields to the relevant configuration interface of the key control structure; Methods for matching the energy consumption of the screened model with the regulated target smart device include: The energy consumption of the target smart device after control is evaluated during each period, and a stable operating power consumption range is set based on the evaluation results; the load standard index of the filtered 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 period, the filtered model is deployed to the target smart device of the period; if the energy consumption of the target smart device changes with the period, the model is hot-swapped to load a model whose load standard index meets the stable operating power consumption range corresponding to the period.

5. The deep learning optimization method for smart devices according to claim 4, characterized in that: The method of performing disturbance simulation on precise device parameters includes: Build smart device operation scenario modeling based on precise device parameters; construct device disturbance templates, and inject dynamic disturbance events into smart device operation scenario modeling based on the device disturbance templates; Collect the response change information and final simulation feedback data of the disturbance simulation process, and construct the disturbance response path of each precise device parameter from the time series dimension based on the response change information and final simulation feedback data; extract the local disturbance fragments in the disturbance response path, analyze the response characteristics of the local disturbance fragments, and output the local disturbance response feature vector; The local stability of each local disturbance response eigenvector is calculated using the stability function; all local disturbance response eigenvectors are vector-concatenated based on the time sequence to output the overall disturbance eigenvector; the macroscopic stability of the overall disturbance eigenvector is calculated; the local stability is weightedly summed and added to the macroscopic stability to output the disturbance stability parameter; all disturbance stability parameters are integrated to construct the disturbance stability index.

6. The deep learning optimization method for smart devices according to claim 5, characterized in that: The method of screening appropriate parameters for precise equipment parameters based on the disturbance stability index includes: Based on the disturbance stability index, the abnormal response fluctuation value of the precise equipment parameters in any round of disturbance simulation is calculated; the parameters with abnormal response fluctuation values ​​less than the preset fluctuation threshold are regarded as stable equipment parameters.

7. The deep learning optimization method for smart devices according to claim 6, characterized in that: The method of performing risk assessment on stable equipment parameters includes: Identify abnormal response trajectories based on abnormal response fluctuation values ​​from each round of disturbance simulation of stable equipment parameters; perform anomaly analysis on the abnormal response trajectories to obtain abnormal fluctuation characteristics; and establish an abnormal mapping relationship between abnormal fluctuation characteristics and stable equipment parameters; Determine the corresponding model of stable equipment parameters and construct a performance sensitivity map of the model under stable equipment parameter conditions; extract abnormal correlation subgraphs from the performance sensitivity map based on the abnormal mapping relationship; quantify the offset degree of each abnormal correlation subgraph to obtain the offset mapping vector; construct a risk propagation model, use the offset mapping vector as the operation basis of the risk propagation model, use the risk propagation model to calculate the risk score of the stable equipment parameters, and construct the parameter risk index based on the risk score.

8. The deep learning optimization method for smart devices according to claim 7, characterized in that: The method of performing model matching on the scheduling intelligent device includes: Identify the affected structures in the current deployment model based on the parameter risk indicators; obtain functional correlation information of the affected structures, and adjust the parameters of the affected structures based on the functional correlation information to obtain a parameter correction model; and calculate the operating power consumption prediction value of the parameter correction model; Combined with the resource status parameters of the scheduling intelligent device in any period, determine whether the operating power consumption prediction value is consistent with the stable operating power consumption range of the scheduling intelligent device in the period; if it is consistent, match the parameter correction model with the scheduling intelligent device in the period, otherwise make secondary adjustments to the parameter correction model; perform power consumption analysis on the parameter correction model to identify high-energy consumption structures; perform hierarchical reconstruction on the high-energy consumption structure to generate a power-saving parameter model and match it with the corresponding scheduling intelligent device; structure the model matching process to output a model adjustment strategy.

9. The deep learning optimization method for smart devices according to claim 8, characterized in that: The method of loading and optimizing the model adjustment strategy includes: 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 until the simulation effect is greater than or equal to the expected value. Then stop the iteration and output the optimized model strategy.

10. A deep learning optimization system for smart devices, used to implement a deep learning optimization method for smart devices according to any one of claims 1 to 9, characterized in that: include: The data acquisition module is used to collect the operating status parameters of the target intelligent device and perform data cleaning on the operating status parameters to obtain accurate device parameters; Energy consumption scheduling module, used to identify the resource level of target smart devices based on precise device parameters; Perform energy consumption scheduling on target smart devices based on resource levels to obtain scheduled smart devices; Parameter screening module, used to simulate disturbances on precise equipment parameters and output disturbance stability indicators; based on the disturbance stability indicators, it screens appropriate parameters for precise equipment parameters and outputs stable equipment parameters; Risk assessment module, used to conduct risk assessment on 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 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; each module is connected by wired and / or wireless means.

Images