Industrial panel power consumption optimization control method and system based on edge AI

Abstract

The invention relates to the technical field of industrial tablet computer calculation optimization control, and discloses an industrial tablet computer power consumption optimization control method and system based on edge AI.The industrial tablet computer power consumption optimization control method based on the edge AI comprises the steps that a task dependency graph of an industrial application program is constructed, task node hotspot scores are calculated, and the task node hotspot scores are calculated; identifying a high-energy-consumption calculation hotspot; training a calculation path heat generation model, predicting the temperature change trend of the hot task and classifying the temperature change trend; constructing a heterogeneous computing unit power consumption performance characteristic database, and computing energy efficiency ratio index matrixes of tasks on different execution units; the task execution sequence is optimized, and targeted heat dissipation measures are activated in advance; optimizing a data access path of a non-migratable task, and continuously optimizing a prediction model and a scheduling strategy through online learning; according to the invention, stability improvement, temperature peak value reduction and power consumption optimization of the industrial panel under a high-load condition are realized, and the technical defects of traditional resource scheduling in the aspect of power consumption management are overcome.

Description

Technical Field

[0001] The present invention relates to the technical field of industrial flat panel computing optimization control, and more specifically, it relates to a method and system for optimizing the power consumption control of industrial flat panels based on edge AI. Background Art

[0002] As an important computing terminal device in the industrial automation environment, industrial flat panels are usually equipped with a variety of heterogeneous computing resources and need to simultaneously execute diverse computing tasks such as real-time control, data analysis, and image processing. With the rapid development of industry and intelligent manufacturing, the computing load of industrial flat panels is continuously increasing, and the application scenarios are becoming more and more complex, making power consumption management a key technical challenge.

[0003] Currently, the power consumption management of industrial flat panels mainly adopts traditional resource scheduling methods, but these methods have obvious defects: the traditional methods lack sufficient consideration of task dependencies and data flows, resulting in additional power consumption overhead during task switching and data transmission; the calculation hot spot management usually adopts a passive response strategy and cannot predict and respond to hot spots in advance, causing the processor to frequently enter the high-temperature throttling cycle, affecting system stability; the energy efficiency performance of different computing tasks on different processing units varies significantly, but there is a lack of intelligent methods to select the best execution unit; the existing thermal management system fails to optimize for the specific computing modes of industrial applications, resulting in unreasonable allocation of heat dissipation resources and low energy utilization efficiency.

[0004] Therefore, a technical solution that can intelligently manage the power consumption of industrial flat panels, improve energy efficiency, and ensure system stability is needed to solve the problems in the prior art. Summary of the Invention

[0005] The present invention provides a method and system for optimizing the power consumption control of industrial flat panels based on edge AI, which solves the technical problems of low efficiency in power consumption management, insufficient hot spot prediction ability, and unreasonable resource scheduling in related technologies.

[0006] The present invention provides a method for optimizing the power consumption control of industrial flat panels based on edge AI, including the following steps: Construct a task dependency graph of industrial application programs through program static analysis and dynamic execution tracing, calculate the hot spot scores of task nodes, and identify high-power consumption calculation hot spots; Based on the identified hot spot tasks, collect multi-dimensional data, train a calculation path heat generation model, predict the temperature change trend of hot spot tasks and classify them; For the identified hot spot tasks and their thermal characteristic classifications, construct a heterogeneous computing unit power consumption performance characteristic database, extract the calculation feature vectors of hot spot tasks, and calculate the energy efficiency ratio index matrix of tasks on different execution units; Implement a differential pre-cooling strategy before the execution of hot tasks according to the thermal characteristic prediction results and the temperature budget model, optimize the task execution order, and activate targeted heat dissipation measures in advance; Based on the hot spot analysis, thermal characteristic prediction, and energy efficiency evaluation results, migrate high-energy-consuming hot tasks to the computing unit with the optimal energy efficiency for execution, optimize the data access path of non-migratable tasks, and continuously optimize the prediction model and scheduling strategy through online learning.

[0007] In a preferred embodiment, the steps of constructing the task dependency graph of the industrial application program include: Extract the function call relationship and data flow through static code analysis; Perform dynamic execution monitoring by inserting lightweight probes at key positions in the application program, and record the actual execution path and data interaction; Combine the static analysis and dynamic monitoring information to generate a complete task dependency graph.

[0008] In a preferred embodiment, the steps of calculating the hot spot score of the task node include: Collect the execution frequency of the task node and the data transfer volume between related task nodes; Calculate the hot spot score, which is the product of the task node execution frequency and the total data transfer volume of all related task nodes; Sort the task nodes according to the hot spot score, and mark the task nodes with hot spot scores exceeding the preset threshold as hot tasks.

[0009] In a preferred embodiment, the steps of training the thermal generation model of the computing path include: Construct a temporal attention network including an encoder, an attention layer, and a decoder; Use multi-dimensional data to train the temporal attention network so that the temporal attention network can predict the temperature change trend of the execution path; The thermal generation model establishes a mapping relationship between the computing path and the temperature prediction result based on instruction characteristics, data access patterns, memory interaction characteristics, historical thermal behavior, and other parameters.

[0010] In a preferred embodiment, the classification of hot tasks includes: Rapid temperature rise type: The temperature rise rate exceeds the preset threshold; High temperature maintenance type: The maximum temperature is above a preset percentage of the processor temperature upper limit and the duration exceeds the preset threshold; Periodic fluctuation type: The difference between the temperature peak and valley exceeds the preset threshold and the period is less than the preset time.

[0011] In a preferred embodiment, the steps of the energy efficiency ratio index matrix of the computing task on different execution units include: Extract the computational feature vectors of hot tasks, including computational density, data parallelism, and memory access intensity; Calculate the energy efficiency ratio metric, which is the ratio of the average power of a task on a specific computing unit to the corresponding execution time.

[0012] In a preferred embodiment, the differential pre-cooling strategy implemented before the execution of hot tasks includes: For rapidly heating tasks, create a preset temperature buffer space; For high-temperature maintenance tasks, adopt a stepped frequency reduction to gradually reduce the frequency to the target value; For periodic fluctuation tasks, implement pulse frequency adjustment, adjust the processor frequency in the opposite phase to the temperature fluctuation period of the task, and actively smooth the temperature curve.

[0013] In a preferred embodiment, the step of optimizing the task execution order calculates the thermal impact matrix between tasks based on a heat conduction model, which considers task heat generation rate, system thermal resistance, heat conduction time constant, and other parameters, and is used to evaluate the temperature impact degree of one task on subsequent tasks.

[0014] In a preferred embodiment, the step of task migration decision includes: Calculate the task migration benefit, which is the energy consumption of the task on the original computing unit minus the energy consumption on the target computing unit minus the data migration energy consumption; When the migration benefit is positive, migrate the task from the original computing unit to the target computing unit for execution; For a hot task cluster with tight data dependencies, perform overall migration to reduce the cross-unit data transfer overhead.

[0015] In a preferred embodiment, an edge AI-based industrial tablet power consumption optimization control system for executing an edge AI-based industrial tablet power consumption optimization control method includes: A task analysis and hot spot identification module for constructing a task dependency graph of industrial applications, calculating the hot spot score of task nodes, and identifying high-energy-consuming computing hot spots; A thermal characteristic prediction and energy efficiency evaluation module for collecting multi-dimensional data, training a computational path heat generation model, predicting the temperature change trend of hot tasks, and constructing a power consumption performance characteristic database to calculate the energy efficiency ratio metric matrix of tasks on different execution units; A pre-cooling and scheduling optimization module for implementing a differential pre-cooling strategy according to the thermal characteristic prediction results, optimizing the task execution order, and activating targeted heat dissipation measures; A task migration execution module, which is used to migrate hot tasks to the computing unit with the optimal energy efficiency for execution and optimize the data access path of non-migratable tasks; An adaptive optimization module, which is used to collect actual energy consumption and temperature data, continuously optimize the prediction model and scheduling strategy, and realize the self-adjustment and performance improvement of the system.

[0016] The beneficial effects of the present invention are as follows: Through task dependency analysis and hot spot prediction algorithms, it is possible to predict the changing trend of the processor temperature in advance, avoid the processor from entering a high-temperature state, reduce the number of passive frequency drops, and improve the stability of the industrial tablet under high-load conditions.

[0017] The pre-cooling and scheduling collaborative optimization strategy reduces the temperature peak of the industrial tablet when executing high-load tasks, improves the thermal stability, avoids frequent triggering of thermal protection, and ensures the continuity and stability of industrial control tasks.

[0018] The intelligent allocation of heterogeneous resources and task migration optimization reduce the computing power consumption of the industrial tablet in typical industrial application scenarios; in the processing of compute-intensive tasks, while the processing speed is increased, the energy consumption is reduced, achieving a dual improvement in performance and energy efficiency.

[0019] The multi-level collaborative optimization strategy extends the battery usage time, improves the sustainability of industrial field operations, reduces the battery replacement or charging frequency, and improves industrial operation efficiency.

[0020] During the operation of the system, continuous self-optimization is carried out, with strong adaptability. It can automatically adjust the optimization strategy according to different industrial application characteristics without manual intervention and configuration, reducing the maintenance cost of industrial applications. Description of the Drawings

[0021] Figure 1 is a flowchart of a method for optimizing the power consumption control of an industrial tablet based on edge AI of the present invention. Detailed Embodiments

[0022] Now, the subject matter described herein will be discussed with reference to exemplary embodiments. It should be understood that discussing these embodiments is only to enable those skilled in the art to better understand and thus implement the subject matter described herein. Without departing from the protection scope of the content of this specification, changes can be made to the functions and arrangements of the elements discussed. Each example can omit, substitute, or add various processes or components as needed. Additionally, the features described in some examples can also be combined in other examples.

[0023] At least one embodiment of the present invention discloses a method for optimizing the power consumption control of an industrial tablet based on edge AI, as Figure 1 shown, including the following steps: Step 1: Construct a task dependency graph for the industrial application through program static analysis and dynamic execution tracing, calculate the hot-spot score of task nodes, and identify high-energy-consuming calculation hot spots. Specifically, it includes the following sub-steps: Step 1.1: By performing static code analysis and dynamic execution monitoring on the industrial application, construct the execution control flow graph and data flow graph of the application, and identify the data dependency relationships between task nodes.

[0024] The static analysis part extracts the function call relationships and data flow directions by parsing the intermediate code representation of the industrial application. The dynamic monitoring part records the actual execution paths and data interactions by inserting lightweight probes at key positions in the application.

[0025] The combination of these two parts of information generates a complete task dependency graph, where the nodes in the graph represent calculation tasks, and the edges represent data dependency relationships or control dependency relationships.

[0026] In some embodiments, the static analysis can be extended to multi-level analysis, which not only analyzes function-level dependencies but also includes module-level and thread-level dependency relationships, forming a hierarchical dependency graph structure.

[0027] Step 1.2: For the identified task nodes, collect their execution frequencies and the data transfer volumes between related task nodes , where represents the currently analyzed task node, represents the related node that has data interaction with .

[0028] The execution frequency is obtained by counting the number of task executions, and the data transfer volume is obtained by monitoring memory access and cross-task data copy operations.

[0029] In some embodiments, the execution duration and average power of the task can also be collected for more accurate evaluation of the energy consumption characteristics of the task.

[0030] Step 1.3: Calculate the hot-spot score of each task node, and the scoring formula is: ; where is the hot-spot score of task node , indicating the importance of this node in terms of energy consumption; is the execution frequency, representing the number of times task node is called per unit time; represents node and node The amount of data transferred between them, measured in bytes or the number of data packets; For the nodes The set of related task nodes with data interaction, including all other nodes that have a data transfer relationship with the node ; Indicates the node The total amount of data transferred between and all related nodes. This formula combines the execution frequency of the task and the amount of data interaction, and can effectively identify the hot tasks that are both frequently executed and have a large amount of data interaction. Such tasks are usually the energy consumption bottlenecks.

[0031] In some embodiments, the hot spot scoring formula can be extended to: ; Wherein, Is the hot spot score of the task node , Is the execution frequency of the task node , Indicates the amount of data transferred between the node And the node , Indicates the total amount of data transferred between the node And all related nodes, Is the set of related task nodes with data interaction with the node , Is the average power of the task node , Is the execution duration of the task node , , Respectively represent the weight coefficients of data interaction energy consumption and computing energy consumption, and are used to balance the energy consumption caused by data interaction and the energy consumption caused by computing itself.

[0032] Step 1.4, sort the task nodes according to the hot spot score, and identify the high-energy consumption computing hot spots.

[0033] The task nodes with hot spot scores exceeding the preset threshold Are marked as hot tasks, and these tasks are the key objects for subsequent optimization.

[0034] In practical applications, the threshold Can be dynamically adjusted according to the characteristics of industrial applications and the performance parameters of the tablet device.

[0035] Step 2, based on the identified hot tasks, collect multi-dimensional data, train a computing path heat generation model, predict the temperature change trend of the hot tasks and classify them; Specifically, it includes the following sub-steps: Step 2.1, collect the instruction characteristics , data access patterns , memory interaction characteristics and historical thermal behavior and other multi-dimensional data.

[0036] Among them, include indicators such as instruction type distribution and computing density; include features such as data locality and access regularity; include indicators such as memory access frequency and cache hit rate; include the temperature curve data of the historical execution path. These data are collected through built-in performance counters and temperature sensors, and are aligned in time series and feature extracted.

[0037] In some embodiments, processor power consumption data , current distribution and other electrical characteristic information can also be collected to enhance the accuracy of thermal characteristic prediction.

[0038] Step 2.2, use these multi-dimensional data to train a computing path thermal generation model. This model adopts a time series attention network structure and can capture the time series thermal characteristics of the execution path. The model expression is: ; Among them, represents the temperature prediction result of the computing path , that is, the future temperature change of the execution path of a specific computing task; represents the mapping function of the thermal generation model, which is a mathematical function that converts input features into temperature prediction values; represents instruction characteristics, including indicators such as instruction type distribution and computing density, reflecting the instruction-level characteristics of the computing task; represents data access patterns, including features such as data locality and access regularity, describing the data processing method of the task; represents memory interaction characteristics, including indicators such as memory access frequency and cache hit rate, reflecting the interaction between the task and the memory system; represents historical thermal behavior, including the temperature curve data of the historical execution path, providing a historical reference for temperature changes.

[0039] This time series attention network consists of an encoder and a decoder structure. The encoder contains multiple layers of bidirectional LSTM layers for extracting time series features; the attention layer is used to identify key influencing factors; the decoder then generates the temperature prediction sequence at future time points.

[0040] The model is first trained on historical data collected offline, and then continuously adapts to the thermal characteristics of a specific industrial tablet through an online learning mechanism.

[0041] In some embodiments, the heat generation model can be extended to a multi-region prediction model to separately predict the temperature changes of multiple physical regions of the processor (such as the core region, cache region, input / output region, etc.), thereby providing a more refined basis for thermal management.

[0042] Step 2.3, for the identified hot tasks, apply the trained heat generation model to predict the temperature rise curve during their execution, and calculate the temperature rise rate and the maximum temperature .

[0043] The prediction results include temperature estimation values at multiple time points, forming a complete temperature change curve, and calculating the temperature slope and peak value.

[0044] In some embodiments, it is also possible to predict the volatility index of the temperature, such as the temperature variance and the temperature change frequency , which are used to more accurately describe the thermal behavior characteristics of the tasks.

[0045] Step 2.4, according to the prediction results, classify the hot tasks into three categories: rapid temperature rise type (high temperature rise rate), high temperature maintenance type (high maximum temperature and long duration), and periodic fluctuation type (frequent temperature fluctuations), providing a basis for subsequent pre-cooling and scheduling optimization.

[0046] The classification uses a threshold-based rule method. For the rapid temperature rise type tasks, the temperature rise rate exceeds 1.5 °C / second. For the high temperature maintenance type tasks, the maximum temperature is above 85% of the processor temperature upper limit and the duration exceeds 5 seconds. For the periodic fluctuation type tasks, the difference between the temperature peak and valley exceeds 8 °C and the period is less than 10 seconds.

[0047] In some embodiments, machine learning methods can be used for hot task classification, and the classifier is trained to automatically learn the optimal classification boundary to adapt to the thermal characteristic differences of different hardware platforms.

[0048] Step 3, for the identified hot tasks and their thermal characteristic classifications, construct a heterogeneous computing unit power consumption and performance characteristic database, extract the computational feature vectors of the hot tasks, and calculate the energy efficiency ratio index matrix of the tasks on different execution units; Specifically, it includes the following sub-steps: Step 3.1, construct a power consumption and performance characteristic database for each heterogeneous computing unit in the industrial panel, and record the execution time and energy consumption of different types of tasks on each computing unit, where represents the task type, and represents the computing unit type.

[0049] This database is constructed by executing benchmark programs and typical industrial application task segments on different computing units, and records the energy consumption and execution time when each task runs on each computing unit.

[0050] In some embodiments, the execution time and energy consumption data under different frequency and voltage configurations can also be recorded to construct a multi-dimensional characteristic table, providing a basis for more refined energy efficiency optimization.

[0051] Step 3.2, for each hot task, extract its computational feature vector , including computational density, data parallelism, memory access intensity, etc.

[0052] The extraction of the computational feature vector is achieved through statistical analysis of the task instruction sequence and data access pattern. The main features include the number of computational operations per byte of data, the parallelism index within the task, the amount of memory access per second, etc.

[0053] In some embodiments, more fine-grained features such as the branch prediction characteristics, instruction-level parallelism, and cache affinity of the task can also be extracted to construct a more comprehensive task feature description.

[0054] Step 3.3, calculate the energy efficiency ratio index matrix of the task on different execution units , and the calculation formula is: ; Among them, represents the energy efficiency ratio index of task on computing unit , which is used to measure the energy efficiency performance of a specific task on a specific computing unit; is the average power of task on computing unit , representing the average energy consumption rate during the task execution; is the execution time of task on computing unit , representing the time required to complete the task. Each element of the energy efficiency ratio index matrix represents the energy efficiency performance of a specific task on a specific computing unit, and the smaller the value, the more energy-efficient the computing unit is for this task.

[0055] This matrix is generated by combining actual measurement and feature similarity interpolation. For tasks that are not directly measured, energy efficiency estimation is performed based on the similarity between their feature vectors and the measured tasks.

[0056] In some embodiments, the energy efficiency ratio calculation can consider the influence of the real-time system load and temperature status, and introduce a dynamic adjustment factor: ; Among them, Represents a task On the computing unit At time The dynamic energy efficiency ratio index; Represents a task On the computing unit The average power; Represents a task On the computing unit The execution time; Represents the computing unit At time The load level, reflecting the current resource occupancy; Represents the computing unit At time The temperature state; Represents an adjustment function, used to dynamically adjust the energy efficiency ratio calculation result according to the current load and temperature state, making the energy efficiency evaluation more in line with the real-time system state.

[0057] Step 3.4, according to the energy efficiency ratio index matrix, determine the computing unit with the optimal energy efficiency for each hot task, and calculate the overhead of task migration, including data transfer time and energy consumption.

[0058] The calculation of migration overhead considers the data volume, the bandwidth between the source and target computing units, and the energy consumption overhead during data transfer.

[0059] In some embodiments, the task allocation decision can also consider real-time constraints and resource utilization balance, and weigh the relationship among energy efficiency, latency, and resource utilization through a multi-objective optimization method.

[0060] Step 4, according to the thermal characteristics prediction result and the temperature budget model, implement a differential pre-cooling strategy before the execution of hot tasks, optimize the task execution order, and activate targeted heat dissipation measures in advance; Specifically, it includes the following sub-steps: Step 4.1, based on the thermal characteristics prediction result, for the fast-heating hot tasks, construct a temperature budget model: ; Among them, Represents the temperature budget at time point , that is, the additional temperature space that the processor can withstand at this time point; Is the temperature threshold, indicating the maximum working temperature allowed by the processor; Is the current temperature, indicating the actual temperature value of the processor at this moment; Is the predicted temperature rise, indicating the expected temperature increase from the current moment to time point Calculated according to the thermal characteristics prediction model.

[0061] The temperature budget model calculates the additional temperature space that the processor can tolerate at each time point, providing a quantitative basis for the pre-cooling strategy.

[0062] In some embodiments, the temperature budget model can be extended to a multi-region model to calculate the temperature budgets of different physical regions of the processor respectively, achieving more refined thermal management control.

[0063] Step 4.2, before the execution of the hot-spot task, according to the temperature budget model, temporarily adjust the processor frequency curve to implement the processor pre-cooling strategy.

[0064] For different hot-spot types, adopt a differentiated pre-cooling strategy: for tasks with rapid temperature rise, reduce the processor frequency 30 to 60 seconds in advance to create a preset temperature buffer space; For tasks with high-temperature maintenance, adopt a stepped frequency reduction, starting 5 to 10 seconds before the task starts and gradually reducing the frequency to the target value; For tasks with periodic fluctuations, implement a pulsed frequency adjustment, adjusting the processor frequency in the opposite phase to the temperature fluctuation period of the task to actively smooth the temperature curve.

[0065] In some embodiments, the pre-cooling strategy can also be combined with the Dynamic Voltage and Frequency Scaling (DVFS) technology. By reducing the frequency and voltage simultaneously, more obvious power consumption reduction effects can be obtained while maintaining performance.

[0066] Step 4.3, optimize the task execution order, calculate the thermal impact matrix between tasks based on the heat conduction model, and insert low-heat generation tasks before high-heat generation tasks to create better initial temperature conditions: ; where is the thermal impact of task on task , representing the temperature contribution of task to the execution environment of task after its execution; is the heat generation rate of task , representing the heat generated by task per unit time; is the thermal resistance, representing the impedance characteristics of the system's heat dissipation path and affecting the heat transfer efficiency; is the time interval between two tasks; is the thermal time constant, representing the characteristic time of the system temperature change and reflecting the magnitude of the system's thermal inertia.

[0067] In actual implementation, the task thermal impact matrix is obtained by combining offline thermal characteristic analysis and online observation, and is used to guide the task scheduler to optimize sorting. On the premise of ensuring task dependencies and real-time requirements, the task execution order is adjusted to optimize the overall temperature curve.

[0068] In some embodiments, the task scheduling optimization can also adopt a thermal-aware graph coloring algorithm. The hot tasks are regarded as nodes in the graph, and the thermal impact is used as the edge weight. The optimal task execution time arrangement is found through the graph coloring algorithm to minimize the peak temperature.

[0069] Step 4.4, according to the hot spot prediction result, activate the targeted heat dissipation measures in advance, adjust the fan speed curve and the heat dissipation channel configuration, and provide higher heat dissipation capacity for the predicted hot spot area.

[0070] The optimization of the heat dissipation system includes: adjusting the fan startup time, pre-starting 10 to 20 seconds before the execution of the hot task; Adjusting the fan speed curve according to the hot spot area to reach the required heat dissipation capacity in advance; Optimizing the heat dissipation channel configuration, such as adjusting the angle of the air grille blades to increase the air flow speed in the hot spot area.

[0071] In some embodiments, the heat dissipation control can also adopt the Model Predictive Control (MPC) method. Based on the thermal dynamics model, the future temperature changes are predicted, and the heat dissipation control sequence is optimized to ensure the heat dissipation effect and minimize the energy consumption and noise.

[0072] Step 5, based on the hot spot analysis, thermal characteristic prediction and energy efficiency evaluation results, migrate the high-energy consumption hot tasks to the computing unit with the optimal energy efficiency for execution, optimize the data access path of the non-migratable tasks, and continuously optimize the prediction model and scheduling strategy through online learning; Specifically, it includes the following sub-steps: Step 5.1, combining the hot spot analysis and energy efficiency evaluation results, construct a task migration decision model and calculate the migration benefit: ; Among them, represents the energy consumption benefit of migrating the task to the computing unit . A positive value indicates a decrease in energy consumption after migration, and a negative value indicates an increase in energy consumption after migration; represents the energy consumption of the task on the original computing unit, reflecting the total energy required for the task to be completed on the current execution unit; represents the energy consumption of the task on the target computing unit , reflecting the total energy required for the task to be completed on the new execution unit after migration; Represents the energy consumption of data migration, including communication energy consumption and additional processing overhead during data transmission.

[0073] The migration decision-making model not only considers the migration benefits of single tasks, but also takes into account the mutual influence between tasks and the overall optimization goal of the system. Through the weighted multi-objective optimization method, it balances the processing performance, power consumption, and temperature control goals.

[0074] In some embodiments, the migration decision can also consider the urgency and priority of tasks. For high-priority tasks, the energy efficiency requirements can be appropriately reduced to ensure execution performance.

[0075] Step 5.2, for the hot tasks with positive migration benefits, migrate them from the original computing unit to the computing unit with the optimal energy efficiency for execution.

[0076] In addition, for the hot task clusters with tight data dependencies, perform overall migration to reduce the cross-unit data transmission overhead.

[0077] The identification of task clusters is based on the data dependency graph. The clustering algorithm is used to divide the tightly associated task subgraphs. When performing cluster migration, the overall migration benefits of the cluster and the cross-cluster data transmission cost will be considered.

[0078] In some embodiments, the task migration can also adopt an incremental migration strategy, decomposing large tasks into multiple subtasks and migrating and executing them step by step to reduce the amount of data and latency of a single migration.

[0079] Step 5.3, for the non-migratable hot tasks (such as due to architecture limitations or real-time requirements), optimize their data access paths and cache policies to reduce memory access hotspots and lower data transmission energy consumption.

[0080] Therefore, the strategies include data prefetching, cache partitioning, access pattern optimization, etc.

[0081] The data prefetching algorithm analyzes the access pattern of tasks and preloads the data into the cache before it is needed; The cache partitioning technology allocates dedicated cache areas for different tasks to reduce cache conflicts; The access pattern optimization improves the cache hit rate and reduces memory access by adjusting the data layout and access order.

[0082] In some embodiments, a collaborative optimization strategy can also be implemented, such as memory page migration, co-location of computing and storage, hierarchical utilization of heterogeneous memory, etc., to further reduce data access energy consumption.

[0083] Step 5.4, implement the monitoring of the task execution process, collect the actual energy consumption and temperature data, and continuously optimize the prediction model and scheduling strategy by applying the online learning algorithm to adapt to the changes in industrial application loads and environmental conditions.

[0084] The online learning system adopts the incremental learning method to continuously update the thermal characteristic prediction model and energy efficiency evaluation model with the newly collected data. As the system running time increases, the prediction accuracy and optimization effect are continuously improved.

[0085] Meanwhile, the system also maintains a reinforcement learning model to learn the optimal control parameters under different workload conditions, including the pre-cooling timing, pre-cooling amplitude, task scheduling strategy, etc.

[0086] In some embodiments, the online learning can also adopt the federated learning framework to integrate the optimization experiences of multiple industrial tablets on the premise of protecting data privacy, accelerate the model convergence and improve the generalization ability.

[0087] Application example of this embodiment: This embodiment is implemented and applied on the industrial tablet devices in a smart manufacturing factory. The factory uses industrial tablets to run various industrial applications, including tasks such as real-time production monitoring, machine vision quality inspection, and production data analysis. Taking the actual application of this factory as an example, the implementation process and effect of this embodiment are demonstrated below.

[0088] The industrial tablets used in this factory are equipped with an 8-core ARM processor, a 4-core GPU, and a dedicated NPU accelerator, and run an industrial control system and various application software. Before applying this embodiment, the following problems existed in the tablets: when running the machine vision quality inspection task, the processor temperature quickly rose to the critical value, triggering the frequency reduction protection, resulting in task execution delay; the battery life was significantly insufficient when running the data analysis task and could not meet the requirements of a complete work shift; the performance was unstable when multiple tasks were executed in parallel, affecting the production management efficiency.

[0089] Implementation process example: Task dependency analysis and hot spot identification example: Analyze the industrial application programs used in the factory and construct a dependency graph containing 137 task nodes.

[0090] Through the monitoring of the actual production process within a week, the execution frequency and data interaction volume of each task node are collected.

[0091] Taking the machine vision quality inspection application as an example, 4 main hot spot tasks are identified: image preprocessing, feature extraction, defect identification, and data storage, and their hot spot scores are 378, 524, 621, and 247 respectively.

[0092] Among them, the defect recognition task has the highest hot-spot score, mainly due to its high computational volume and a large amount of data interaction with the feature extraction task.

[0093] Implementation example of the thermal characteristic prediction model: Collected data on the temperature change of the flat under different workloads and trained the thermal characteristic prediction model.

[0094] Used an encoder with 2 BiLSTM layers (each with 128 hidden units) and a decoder with 1 LSTM layer (128 hidden units), with a 4-head attention mechanism layer in the middle.

[0095] For the defect recognition task in the quality inspection application, the model can predict its temperature change curve: Under the standard workload, it takes about 78 seconds to reach the temperature peak from the start of execution, the temperature rise rate is 0.4 °C / second, and the highest temperature is expected to reach 83 °C, belonging to a hot-spot task with high-temperature maintenance.

[0096] In actual tests, the average absolute error between the model prediction and the actual temperature curve is 1.7 °C, and it can accurately predict the time when the temperature peak appears 53 seconds in advance.

[0097] Implementation example of heterogeneous resource energy efficiency optimization: Through energy efficiency evaluation, an energy efficiency ratio index matrix of factory applications on different computing units was generated.

[0098] Taking the defect recognition task as an example, the energy efficiency ratios on the CPU, GPU, and NPU are 19.7, 12.6, and 8.3 respectively, indicating that it is most energy-efficient to execute on the NPU.

[0099] The energy efficiency ratios of the data warehousing task on the CPU, GPU, and NPU are 10.2, 25.3, and 22.7 respectively, indicating that it is optimal to execute on the CPU.

[0100] Task allocation based on these data improves the overall energy efficiency: migrating the defect recognition task from the original CPU to the NPU for execution, the energy efficiency is increased by 57.9%; at the same time, migrating the data processing task in the data analysis application from the GPU to the CPU, the energy efficiency is increased by 59.7%.

[0101] Implementation example of pre-cooling and scheduling optimization: For the high-temperature maintenance type defect recognition task in the quality inspection application, a pre-cooling strategy was implemented.

[0102] 45 seconds before the task starts, the system gradually reduces the processor frequency from the standard 2.4 GHz to 2.0 GHz to create a temperature buffer; At the same time, change the task execution order, insert a low-heat generation data sorting task before the defect recognition to optimize the overall temperature curve.

[0103] These strategies reduce the temperature peak during the execution of the quality inspection application from the original 85°C to 76°C, avoiding the frequency adjustment triggered by thermal protection and ensuring the stable operation of the application.

[0104] In actual tests, under the condition of maintaining the same quality inspection accuracy, the pre-cooling strategy reduces the number of times the processor enters the thermal protection state from an average of 4.2 times per hour to 0.8 times, improving the overall system stability.

[0105] Task migration and online optimization example: The system continuously monitors the actual energy consumption and performance of each task during operation and continuously optimizes the resource allocation strategy through online learning.

[0106] For example, when the production line speed increases, resulting in an increase in the frequency of quality inspection tasks, the system automatically adjusts the priority and resource allocation ratio of quality inspection tasks and data analysis tasks to ensure the real-time nature of quality inspection tasks while maintaining the stable execution of data analysis tasks.

[0107] Through online learning optimization, the accuracy of the system's prediction model has been improved from the initial 78% to 93%, the average response time of hot tasks has been reduced by 32%, and the average working temperature of the processor has been reduced by 7.6°C.

[0108] Technical effect verification: Effect of improving the temperature stability of the processor: Table 1: Comparison data of processor temperature changes before and after implementing this method;

[0109] Effect of power consumption optimization and battery life improvement: Table 2: Improvement data of power consumption and battery life under different working scenarios;

[0110] From the above actual application examples, it can be seen that the application effect of this implementation method in the real industrial environment is obvious. It not only significantly improves the thermal stability and energy efficiency of industrial tablets, but also improves the overall system performance through intelligent resource scheduling, providing strong support for efficient production in industrial sites.

[0111] The above describes the embodiments of the present invention, but these embodiments are not limited to the above specific implementation manners. The above specific implementation manners are only illustrative and not restrictive. Under the inspiration of this embodiment, those of ordinary skill in the art can also make more equivalent embodiments in various forms, all of which fall within the protection scope of this embodiment.

Claims

1. An industrial tablet power consumption optimization control method based on edge AI, characterized in that, It includes the following steps: Construct a task dependency graph of industrial application programs through program static analysis and dynamic execution tracing, calculate the hot-spot scores of task nodes, and identify high-energy-consuming computing hotspots; Based on the identified hot-spot tasks, collect multi-dimensional data, train a computing path heat generation model, predict the temperature change trend of hot-spot tasks and classify them; For the identified hot-spot tasks and their heat characteristic classifications, construct a database of power consumption and performance characteristics of heterogeneous computing units, extract the computing feature vectors of hot-spot tasks, and calculate the energy efficiency ratio index matrix of tasks on different execution units; According to the heat characteristic prediction results and the temperature budget model, implement a differential pre-cooling strategy before the execution of hot-spot tasks, optimize the task execution order, and activate targeted heat dissipation measures in advance; Based on the hot-spot analysis, heat characteristic prediction and energy efficiency evaluation results, migrate high-energy-consuming hot-spot tasks to the computing unit with the optimal energy efficiency for execution, optimize the data access path of non-migratable tasks, and continuously optimize the prediction model and scheduling strategy through online learning.

2. The industrial tablet power consumption optimization control method based on edge AI according to claim 1, characterized in that, The steps of constructing the task dependency graph of industrial application programs include: Extract function call relationships and data flow directions through static code analysis; Perform dynamic execution monitoring by inserting lightweight probes at key positions in the application program, and record the actual execution path and data interaction; Combine static analysis and dynamic monitoring information to generate a complete task dependency graph.

3. The industrial tablet power consumption optimization control method based on edge AI according to claim 1, characterized in that, The steps of calculating the hot-spot scores of task nodes include: Collect the execution frequency of task nodes and the data transmission volume between related task nodes; Calculate the hot-spot score, which is the product of the execution frequency of the task node and the total data transmission volume of all related task nodes; Sort the task nodes according to the hot-spot scores, and mark the task nodes with hot-spot scores exceeding the preset threshold as hot-spot tasks.

4. A method for optimizing and controlling the power consumption of an industrial tablet based on edge AI according to claim 1, characterized in that, The steps of training the computing path heat generation model include: Construct a temporal attention network including an encoder, an attention layer and a decoder; Use multi-dimensional data to train the temporal attention network so that the temporal attention network can predict the temperature change trend of the execution path; The heat generation model establishes a mapping relationship between the computing path and the temperature prediction result based on instruction characteristics, data access patterns, memory interaction characteristics, historical heat behavior and other parameters.

5. The industrial tablet power consumption optimization control method based on edge AI according to claim 1, wherein, The classification of hot-spot tasks includes: Rapid temperature rise type: The temperature rise rate exceeds the preset threshold; High temperature maintenance type: The maximum temperature is above a preset percentage of the processor temperature upper limit and the duration exceeds the preset threshold; Periodic fluctuation type: The difference between the temperature peak and valley exceeds the preset threshold and the period is less than the preset time.

6. The industrial tablet power consumption optimization control method based on edge AI according to claim 1, wherein, The steps of calculating the energy efficiency ratio index matrix of tasks on different execution units include: Extract the computing feature vectors of hot-spot tasks, including computing density, data parallelism and memory access intensity; Calculate the energy efficiency ratio index, which is the ratio of the average power of the task on a specific computing unit to the corresponding execution time.

7. A method for optimizing the power consumption control of an industrial tablet based on edge AI according to claim 1, characterized in that, The differential pre-cooling strategy in implementing the differential pre-cooling strategy before the execution of hot-spot tasks includes: For tasks of the rapid temperature rise type, create a preset temperature buffer space; For tasks of the high temperature maintenance type, adopt a stepped frequency reduction to gradually reduce the frequency to the target value; For periodic fluctuation tasks, implement pulse - type frequency adjustment, adjust the processor frequency with a phase opposite to the temperature fluctuation period of the task, and actively smooth the temperature curve.

8. A method for optimizing and controlling the power consumption of an industrial tablet based on edge AI according to claim 1, characterized in that, The step of optimizing the task execution order calculates the thermal impact matrix between tasks based on the heat conduction model. The heat conduction model considers parameters such as task heat generation rate, system thermal resistance, heat conduction time constant, etc., and is used to evaluate the degree of temperature impact of one task on subsequent tasks.

9. The industrial tablet power consumption optimization control method based on edge AI according to claim 1, characterized in that The step of task migration decision includes: Calculate the task migration benefit, which is the energy consumption of the task on the original computing unit minus the energy consumption on the target computing unit minus the data migration energy consumption; When the migration benefit is positive, migrate the task from the original computing unit to the target computing unit for execution; For hot - spot task clusters with tight data - dependency relationships, perform overall migration to reduce the cross - unit data transfer overhead.

10. An industrial tablet power consumption optimization control system based on edge AI, which is used to execute an industrial tablet power consumption optimization control method according to any one of claims 1-9, characterized in that, Include: Task analysis and hot - spot identification module, which is used to construct the task dependency graph of industrial applications, calculate the hot - spot score of task nodes, and identify high - energy - consumption computing hot - spots; Thermal characteristic prediction and energy - efficiency evaluation module, which is used to collect multi - dimensional data, train the heat generation model of the computing path, predict the temperature change trend of hot - spot tasks, and construct a power consumption - performance characteristic database, and calculate the energy - efficiency ratio index matrix of tasks on different execution units; Precooling and scheduling optimization module, which is used to implement a differential precooling strategy according to the thermal characteristic prediction results, optimize the task execution order, and activate targeted heat dissipation measures; Task migration execution module, which is used to migrate hot - spot tasks to the computing unit with the optimal energy - efficiency for execution, and optimize the data access path of non - migratable tasks; Adaptive optimization module, which is used to collect actual energy consumption and temperature data, continuously optimize the prediction model and scheduling strategy, and achieve self - adjustment and performance improvement of the system.

Images