A machine learning based power tool operating parameter optimization system

Abstract

The application discloses a machine learning-based electric tool operation parameter optimization system, and belongs to the technical field of computer-aided design.The system comprises a strategy setting unit, a data acquisition unit, a feature processing unit, an algorithm fusion unit, a decision control unit, an execution feedback unit and an application output unit.The strategy setting unit is used for generating strategy parameters.The data acquisition unit is used for collecting basic data, wherein the basic data comprises three-dimensional model data, material physical property data and microcontroller unit resource limitation parameters.The feature processing unit is used for receiving the basic data and encoding the basic data to generate a unified feature vector.The algorithm fusion unit is used for receiving the unified feature vector and performing multi-objective optimization calculation to generate a Pareto optimal front.The application realizes a full-automatic process from high-level strategy target setting to final deployable firmware parameter set generation, and significantly improves the design efficiency and quality of high-performance electric tool firmware.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of computer-aided design, in particular to an electric tool operation parameter optimization system based on machine learning. BACKGROUND

[0002] The traditional electric tool firmware development mode in the CAD stage relies on static experience model and physical prototype test, and has problems such as cognitive delay, lack of collaborative optimization, extensive control and model solidification. Especially in the face of the extreme challenge of designing a general firmware for an aviation-level riveting gun running on a resource-limited MCU, the existing technology cannot systematically balance the operation performance, resource consumption and model generalization ability, and it is difficult to effectively avoid the long-term reliability risk caused by the introduction of new materials in the design stage.

[0003] To solve the above problems, the present application provides an optimization system running in a CAD software environment.

[0004] The above information disclosed in the above background section is only used to strengthen the understanding of the background of the present disclosure, and therefore it can include information that does not constitute prior art known to those of ordinary skill in the art. SUMMARY

[0005] The purpose of the present application is to provide an electric tool operation parameter optimization system based on machine learning to solve the problems raised in the above background.

[0006] The technical solution of the present application is to generate strategy parameters by a strategy setting unit, acquire basic data by a data acquisition unit, process features by a feature processing unit, fuse algorithms by an algorithm fusion unit, make decisions and control by a decision control unit, perform feedback by an execution feedback unit, and output applications by an application output unit.

[0007] The strategy setting unit is used to generate strategy parameters.

[0008] The data acquisition unit is used to collect basic data, wherein the basic data includes three-dimensional model data, material physical property data and microcontroller unit resource limitation parameters.

[0009] The feature processing unit is used to receive the basic data and encode it to generate a unified feature vector.

[0010] The algorithm fusion unit is used to receive the unified feature vector and perform multi-objective optimization calculation to generate a Pareto optimal frontier.

[0011] The decision control unit is used to select from the Pareto optimal frontier to generate a to-be-verified execution parameter vector.

[0012] The execution feedback unit is used to perform high-precision simulation on the to-be-verified execution parameter vector to generate real target values and new labeled samples.

[0013] The algorithm fusion unit is also configured to receive the new labeled sample to update the agent model inside it.

[0014] The application output unit is configured to generate a firmware parameter set based on the optimal solution confirmed by the decision control unit.

[0015] Preferably, the feature processing unit generates the uniform feature vector in the following process:

[0016] The material physical property data is processed by a material attribute structural feature extractor to output a material feature vector;

[0017] The stress-strain field three-dimensional grid data obtained after simulation by the execution feedback unit is processed by a digital twin state tensor encoder to output a state tensor;

[0018] The material feature vector and the state tensor are spliced to generate the uniform feature vector.

[0019] Preferably, the algorithm fusion unit performs the multi-objective optimization calculation in the following process:

[0020] An agent model based on a Transformer architecture is constructed to predict the predicted values of three core optimization objectives, namely, job performance, resource consumption, and generalization ability;

[0021] A multi-objective evolutionary algorithm is adopted, and the agent model is used as a fitness function to perform evolutionary calculation on the execution parameter vector to generate the Pareto optimal front.

[0022] Preferably, the core optimization objectives are calculated in the following manner:

[0023] The job performance is calculated by the reciprocal of the single riveting time predicted by the agent model;

[0024] The resource consumption is calculated by weighted summation normalization of the memory occupation and CPU calculation period predicted by the agent model, combined with the upper limit of hardware resources obtained from the data acquisition unit and the weight coefficient defined by the strategy setting unit;

[0025] The generalization ability is characterized by calculating the mathematical expectation of a custom health degree function under the disturbance of the normal distribution of material yield strength.

[0026] Preferably, the custom health degree function is calculated based on an improved logistic function, combined with the maximum equivalent stress calculated by the execution feedback unit, the workpiece material yield strength read from the data acquisition unit, and the safety coefficient threshold and health degree curve steepness coefficient defined by the strategy setting unit.

[0027] Preferably, the decision control unit adopts a reinforcement learning decision agent, and the operation mechanism of the reinforcement learning decision agent is:

[0028] defining a state, the state including a number of Pareto frontier solutions, a target vector of a performance optimal solution, a target vector of a health optimal solution and a normalized state vector of a target space hyper-volume;

[0029] defining an action, the action being selecting a specific solution from the Pareto frontier as the to-be-verified execution parameter vector;

[0030] defining a reward, the reward being calculated according to a real simulation result of the to-be-verified execution parameter vector obtained by the execution feedback unit, in combination with the expected performance target and the strategy weight defined by the strategy setting unit, and including a penalty term;

[0031] wherein if the real simulation result corresponds to a real health degree lower than a preset health degree threshold, the penalty term is triggered; and if the real health degree is greater than or equal to the preset health degree threshold, the penalty term is not triggered.

[0032] Preferably, the execution feedback unit generates a verification result and a new labeled sample;

[0033] the verification result including an accurate stress field and a real health degree calculated by a high-precision simulation engine;

[0034] the new labeled sample being formed by pairing the input to-be-verified execution parameter vector and the real target value output by simulation.

[0035] Preferably, the firmware parameter set includes motor control parameters, process monitoring parameters and material adaptation parameters;

[0036] the motor control parameters including current loop PID parameters, speed loop feedforward coefficient, PWM frequency and time-torque anchor point sequence of a target rotating speed segmented curve;

[0037] the process monitoring parameters including impact force sensor trigger threshold, motor stall detection current threshold and displacement-force comprehensive logic threshold for judging completion of a task;

[0038] the material adaptation parameters including a plurality of parameter sets preset for different materials, and weight and bias parameters of a lightweight neural network for calculating a new material parameter correction amount.

[0039] The present application provides an electric tool running parameter optimization system based on machine learning by improvement, compared with the prior art, has the following improvements and advantages:

[0040] 1. By constructing a vertically integrated system composed of a strategic setting unit, a data acquisition unit, a feature processing unit, an algorithm fusion unit, a decision control unit, an execution feedback unit, and an application output unit, the technical bottlenecks of cognitive delay and lack of collaborative optimization in traditional development mode are overturned. The system realizes the full automation process from high-level strategic goal setting to the generation of the final deployable firmware parameter set, significantly improving the design efficiency and quality of high-performance power tool firmware;

[0041] 2. The feature processing unit of the system can efficiently encode heterogeneous physical world data, including material physical property data and simulation-generated stress-strain field three-dimensional grid data, into a unified feature vector through material attribute structured feature extractors and digital twin state tensor encoders. This processing method solves the fusion problem between different data sources and lays a solid data foundation for subsequent cross-domain multi-objective optimization. The algorithm fusion unit innovatively uses a proxy model based on the Transformer architecture to accurately predict the three core optimization objectives of job performance, resource consumption, and generalization ability. Combined with multi-objective evolutionary algorithms, the system can efficiently perform global optimization in high-dimensional parameter space, replacing the time-consuming high-precision simulations in traditional methods, thereby generating a Pareto optimal frontier representing different performance trade-offs in a very short time, greatly accelerating the optimization process;

[0042] 3. The system uniquely introduces a decision control unit that uses a reinforcement learning decision agent. This agent can perceive the global distribution pattern of the Pareto frontier and automatically select a to-be-verified execution parameter vector that best balances multiple conflicting objectives based on a reward mechanism that includes a risk-averse penalty term. This intelligent decision-making mechanism surpasses the subjectivity and limitations of manual selection, ensuring that the final solution achieves an optimal balance between performance, cost, and generalization ability, and actively avoids high-risk designs;

[0043] 4. Through the closed-loop learning path constructed between the execution feedback unit and the algorithm fusion unit, the system realizes continuous self-optimization. The execution feedback unit performs high-precision simulation on the selected solution, and the generated real target values and new labeled samples are used to iteratively update the proxy model within the algorithm fusion unit. This mechanism ensures that the prediction accuracy of the proxy model continuously improves with the number of iterations, ensuring the convergence of the entire optimization process and the physical reality of the final solution;

[0044] 5. The application output unit generates a firmware parameter set with high structure and adaptability. The parameter set not only contains precise motor control parameters and process monitoring parameters, but also integrates a lightweight neural network for calculating the correction amount of new material parameters. This design enables the final firmware to quickly adapt to unknown materials, significantly enhancing the generalization ability and application range of power tools in variable working conditions, especially for high-performance power tools such as aviation riveting guns with limited resources, providing an optimized solution that balances performance, resources, and generalization ability. BRIEF DESCRIPTION OF DRAWINGS

[0045] The application will be further explained in conjunction with the accompanying drawings and examples:

[0046] Figure 1 is a flow chart of the system of the application. DETAILED DESCRIPTION

[0047] To make the purpose, technical solutions and advantages of the application clearer, the application will be further described in conjunction with specific examples.

[0048] Example 1

[0049] Please refer to Figure 1 , the application provides a machine learning-based power tool running parameter optimization system, including a strategy setting unit, a data acquisition unit, a feature processing unit, an algorithm fusion unit, a decision control unit, an execution feedback unit and an application output unit;

[0050] The strategy setting unit is used to generate strategy parameters;

[0051] The data acquisition unit is used to collect basic data, wherein the basic data includes three-dimensional model data, material physical property data and microcontroller unit resource limitation parameters;

[0052] The feature processing unit is used to receive the basic data and encode it to generate a unified feature vector;

[0053] The algorithm fusion unit is used to receive the unified feature vector and perform multi-objective optimization calculation to generate a Pareto optimal frontier;

[0054] The decision control unit is used to select from the Pareto optimal frontier to generate a to-be-verified execution parameter vector;

[0055] The execution feedback unit is used to perform high-precision simulation on the to-be-verified execution parameter vector to generate real target values and new labeled samples;

[0056] The algorithm fusion unit is also used to receive the new labeled samples to update the internal surrogate model;

[0057] The application output unit is configured to generate a firmware parameter set based on the optimal solution confirmed by the decision control unit.

[0058] The application provides a technical solution: an electric tool operation parameter optimization system based on machine learning. A strategy setting unit generates strategy parameters to define the boundary and target of the optimization task; a data acquisition unit collects three-dimensional model data of the electric tool, material physical property data of various workpieces, and resource limitation parameters of the target microcontroller unit as basic data; a feature processing unit receives the basic data and encodes it into a uniform feature vector that can be processed by an algorithm; an algorithm fusion unit receives the uniform feature vector and performs multi-objective optimization calculation in a high-dimensional parameter space to generate a set of Pareto optimal frontiers representing different performance trade-offs; a decision control unit intelligently selects an execution parameter vector to be verified from the Pareto optimal frontiers according to the strategy parameters; an execution feedback unit performs high-precision digital twin simulation on the execution parameter vector to be verified to obtain its real target value and generates new labeled samples for model learning; the new labeled samples are fed back to the algorithm fusion unit for iterative updating of the internal surrogate model, forming a closed-loop learning path; and an application output unit finally generates a firmware parameter set that can be directly deployed on the electric tool hardware based on the optimal solution confirmed by the decision control unit; through the vertical integration and closed-loop feedback of the seven units, the system realizes an automated process from strategy definition to firmware generation, solves the technical problems of cognitive delay and lack of collaborative optimization in traditional development modes, and is especially suitable for designing firmware for high-performance electric tools such as aviation-level riveting guns that are resource-constrained.

[0059] Embodiment 2

[0060] The process of generating a uniform feature vector by the feature processing unit is as follows:

[0061] The material physical property data is processed by a material attribute structured feature extractor to output a material feature vector;

[0062] The stress-strain field three-dimensional grid data obtained after simulation by the execution feedback unit is processed by a digital twin state tensor encoder to output a state tensor;

[0063] The material feature vector and the state tensor are spliced to generate a uniform feature vector;

[0064] Considering that when the system performs the first iteration optimization, there is no simulation data generated by the execution feedback unit, at this time the state tensor is initialized as a zero vector with the same dimension as the specified dimension, for example, 128 dimensions The initialized state tensor is spliced with the material feature vector to start the first multi-objective optimization calculation.

[0065] The process of multi-objective optimization calculation by the algorithm fusion unit is as follows:

[0066] A proxy model based on the Transformer architecture is constructed to predict the predicted values of the three core optimization objectives of job performance, resource overhead, and generalization ability.

[0067] A multi-objective evolutionary algorithm is used, and the proxy model is used as the fitness function to perform evolutionary calculation on the execution parameter vector to generate the Pareto optimal front.

[0068] This embodiment is an explanation and description in embodiment 1. Specifically, the process of generating a unified feature vector by the feature processing unit is refined. The material attribute structural feature extractor inside the feature processing unit is configured as a graph neural network, which processes the input material physical property data, learns the deep correlation between material chemical composition and lattice type parameters, and finally outputs a fixed-dimensional material feature vector. At the same time, the digital twin state tensor encoder inside the feature processing unit is configured as a convolutional autoencoder, which processes the stress-strain field three-dimensional grid data obtained after simulation by the execution feedback unit, compresses the high-dimensional space distribution information, and outputs a state tensor containing key mechanical features. The material feature vector and the state tensor are spliced to generate a unified feature vector; the algorithm fusion unit receives the unified feature vector and constructs a proxy model based on the Transformer architecture. The proxy model is pre-trained to replace time-consuming high-precision simulation to quickly predict the predicted values of the three core optimization objectives of job performance, resource overhead, and generalization ability corresponding to any execution parameter vector; the algorithm fusion unit further uses the NSGA-II multi-objective evolutionary algorithm to perform efficient evolutionary calculation on the execution parameter vector using the proxy model as the fitness function, and explores and generates the Pareto optimal front.

[0069] In a preferred embodiment, the parameter configuration of the NSGA-II multi-objective evolutionary algorithm is as follows: the population size is set to 100, the iteration number is set to 200, the crossover probability is set to 0.9, and the mutation probability is set to 0.1.

[0070] By encoding heterogeneous physical data into a unified feature vector and using a proxy model to accelerate optimization, the system can efficiently perform global optimization under complex constraints.

[0071] To enable those skilled in the art to implement, a key model architecture is described as follows:

[0072] The graph neural network can specifically adopt a structure containing 3 graph convolution layers, each followed by a ReLU activation function, and finally output a 64-dimensional material feature vector through a global pooling layer.

[0073] Convolutional Autoencoder, whose encoder part can be composed of 4 3D convolutional layers with a step of 2, with the number of convolutional kernels being 16, 32, 64, and 128 respectively, for compressing the input three-dimensional grid data with a size of ; and whose decoder part adopts corresponding transpose convolutional layers for reconstruction. The 128-dimensional vector output by the encoder is the state tensor;

[0074] The agent model based on the Transformer architecture can adopt a structure containing only an encoder, which is stacked by 6 Transformer layers, each of which contains 8 attention heads. The input of the model is the 192-dimensional, 64-dimensional material feature + 128-dimensional state unified feature vector and the execution parameter vector to be predicted after splicing, and the output is the prediction value of the three targets of job performance, resource consumption, and generalization ability; the pre-training is to use a dataset containing 10,000 simulation label samples obtained by random sampling parameters to supervise the learning and training of the agent model, so that it has preliminary prediction ability.

[0075] Embodiment 3

[0076] The calculation method of the core optimization target is:

[0077] The job performance is calculated by the reciprocal of the single riveting time predicted by the agent model;

[0078] The resource consumption is calculated by weighted summation and normalization of the memory occupation and CPU calculation period predicted by the agent model, combined with the upper limit of hardware resources obtained from the data acquisition unit and the weight coefficient defined by the strategy setting unit;

[0079] The generalization ability is represented by calculating the mathematical expectation of the self-defined health function under the normal distribution disturbance of the material yield strength.

[0080] The self-defined health function is calculated based on the improved logistic function, combined with the maximum equivalent stress calculated by the execution feedback unit, the workpiece material yield strength read from the data acquisition unit, and the safety factor threshold and health function curve steepness coefficient defined by the strategy setting unit.

[0081] This embodiment is an interpretation and explanation in embodiment 2, specifically, the calculation method of the core optimization target is refined; the job performance is calculated by the reciprocal of the single riveting time predicted by the agent model , and the calculation formula is:

[0082]

[0083] Wherein, : execution parameter vector, Single riveting time

[0084] Resource overhead is a dimensionless cost component, memory occupation predicted by proxy model and CPU computing cycles , combined with the upper limit of hardware resources obtained from the data acquisition unit and , and the weight coefficient defined by the strategy setting unit and , weighted sum normalization calculation, the calculation formula is:

[0085]

[0086] Among them, : predicted memory occupation, : predicted computing cycle, : memory upper limit, : cycle upper limit, : memory weight, : cycle weight;

[0087] Generalization ability By calculating the custom health function The mathematical expectation of the material yield strength under normal distribution disturbance is characterized, and the calculation formula is:

[0088]

[0089] Among them, : mathematical expectation, : yield strength after disturbance, : normal distribution, : original yield strength, : disturbance standard deviation;

[0090] Custom health function is used to quantify the safety relationship between the design point and the failure boundary, which is calculated based on the improved logistic function, and the calculation formula is:

[0091]

[0092] Among them, : maximum equivalent stress, : material yield strength, : steepness coefficient, a dimensionless parameter, used to adjust the steepness of the function curve, : safety factor threshold; : natural exponential function;

[0093] Among them, the maximum equivalent stress Yield strength of workpiece material calculated from simulation by execution feedback unit Read from material library of data acquisition unit, safety factor threshold Steepness coefficient of health degree curve Defined by strategy setting unit; this series of formulas converts abstract engineering targets into precise, continuous and differentiable mathematical expressions, providing a solid foundation for multi-objective optimization algorithms.

[0094] The decision control unit adopts a reinforcement learning decision agent, and the operation mechanism of the reinforcement learning decision agent is as follows:

[0095] The state is defined, including the number of Pareto frontier solutions, the target vector of the performance optimal solution, the target vector of the health optimal solution, and the normalized state vector of the target space hyper-volume;

[0096] The action is defined as selecting a specific solution from the Pareto frontier as the to-be-verified execution parameter vector;

[0097] The reward is calculated according to the real simulation result of the to-be-verified execution parameter vector obtained by the execution feedback unit, combined with the expected performance target and strategy weight defined by the strategy setting unit, and includes a penalty term;

[0098] If the real health degree corresponding to the real simulation result is lower than the preset health degree threshold, the penalty term is triggered; if the real health degree is greater than or equal to the preset health degree threshold, the penalty term is not triggered.

[0099] This embodiment is an explanation and description in embodiment 1. Specifically, the decision control unit adopts an offline trained reinforcement learning decision agent to realize automatic optimal selection from the Pareto optimal frontier; the operation mechanism of the reinforcement learning decision agent is precisely defined, and its state is defined as containing the number of Pareto frontier solutions, the target vector of the performance optimal solution, the target vector of the health optimal solution, and the normalized state vector of the target space hyper-volume, to perceive the global distribution form of the Pareto frontier; its action is defined as selecting a specific solution from the Pareto frontier as the to-be-verified execution parameter vector ; its reward is calculated according to the real simulation result of the to-be-verified execution parameter vector obtained by the execution feedback unit, combined with the expected performance target and strategy weight defined by the strategy setting unit, and the reward function is:

[0100]

[0101] wherein, : real performance, resource, generalization target value, : expected performance target, : strategy weight, : penalty function, : real health degree;

[0102] The reward function includes a penalty term , which is defined as a conditional function, when the real health degree corresponding to the real simulation result is lower than a preset health degree threshold , a larger negative penalty is triggered, thereby forcing the decision-making process to always avoid high-risk designs; the reinforcement learning decision-making agent can automatically balance multiple conflicting goals through learning and make intelligent trade-offs in line with long-term risk-averse strategies; the preset health degree threshold and the penalty value are defined by the strategy setting unit according to specific risk-averse strategies and performance requirements;

[0103] For further clarification, the reinforcement learning decision-making agent can use a proximal policy optimization algorithm for training. The offline training process is to perform 500 Latin hypercube sampling on the execution parameter vector space, run a complete high-precision simulation, and thereby construct an initial data set containing 500 groups of Pareto frontiers-decision choices-real returns. Subsequently, the PPO agent is trained using this data set, so that it learns the strategy of selecting the optimal solution from different forms of Pareto frontiers;

[0104] The strategy setting unit, in one embodiment, is a graphical user interface provided for users, allowing design engineers to manually input and adjust the above strategy parameters according to specific project requirements, such as prioritizing performance or prioritizing low power consumption; for example, for a high-performance scenario, the engineer can set a higher performance weight ; for an aviation-level application requiring high safety, a higher safety factor threshold T and health degree threshold can be set; all parameters have their physical or engineering meanings, and their value ranges can be set according to empirical data or industry standards.

[0105] The execution feedback unit generates verification results and new labeled samples;

[0106] The verification results include the accurate stress field and real health degree calculated by the high-precision simulation engine;

[0107] The new labeled samples are formed by pairing the input execution parameter vector to be verified with the real target value output by the simulation.

[0108] This embodiment is an explanation and illustration in embodiment 1, specifically, the execution feedback unit is executed as a key to high-precision verification and data closed loop; the execution feedback unit receives the execution parameter vector to be verified selected by the decision control unit, and calls the high-precision simulation engine based on the extended finite element method to perform detailed physical process simulation;

[0109] The simulation setting includes: applying a fixed constraint to the bottom of the workpiece three-dimensional model as a boundary condition; generating a corresponding tool head dynamic load according to the execution parameter vector to be verified, and applying it to the riveting position; using hexahedral elements to divide the workpiece into a grid, with a total number of about 50000, and local encryption in the contact area; the mechanical behavior of the workpiece material is described by using an elastic-plastic constitutive model containing isotropic hardening;

[0110] After the simulation is completed, the execution-feedback unit generates two types of core outputs; the first type is the verification result, which includes the accurate stress field distribution in the simulated workpiece and the real health degree calculated according to the stress field, and the verification result is used to finally judge whether the current parameter vector meets the safety constraints defined in the strategic setting unit; the second type is a new labeled sample, which is formed by accurately pairing the input execution parameter vector to be verified with the real target value output by the simulation. The new labeled sample is directly fed back to the algorithm fusion unit for incremental updating and correction of the internal proxy model; through this mechanism, the prediction accuracy of the proxy model is continuously improved with the increase of the number of optimization iterations, ensuring the convergence of the entire system learning and the physical reality of the final solution.

[0111] The firmware parameter set includes motor control parameters, process monitoring parameters and material adaptation parameters;

[0112] The motor control parameters include current loop PID parameters, speed loop feedforward coefficient, PWM frequency and time-torque anchor point sequence of target speed segmented curve;

[0113] The process monitoring parameters include impact force sensor trigger threshold, motor stall detection current threshold and displacement-force comprehensive logic threshold for job completion judgment;

[0114] The material adaptation parameters include multiple sets of parameter sets preset for different materials, and the weight and bias parameters of a lightweight neural network for calculating the correction amount of new material parameters.

[0115] The embodiment is an explanation of the embodiment 1. Specifically, the application output unit is responsible for generating the final firmware parameter set after receiving the optimal solution confirmed by the decision control unit and the execution feedback unit; the firmware parameter set is structured into three modules; the motor control parameter module includes the current loop PID parameter, the speed loop feedforward coefficient, the PWM frequency, and the time-torque anchor point sequence defining the target speed segmented curve of the work process for precise motor motion control; the process monitoring parameter module includes the impact force sensor trigger threshold, the motor locked-rotor detection current threshold, and the displacement-force comprehensive logic threshold for judging whether the work is completed. The material adaptation parameter module not only includes multiple sets of the above-mentioned parameter sets preset for different materials in the material library, but also innovatively contains the weight and bias parameters of the lightweight neural network, which is used to receive the material feature vector of the new material and calculate a set of correction amounts for the reference parameters in real time, thereby realizing the rapid adaptation to unknown materials.

[0116] In the embodiment, the lightweight neural network is a multi-layer perceptron, whose structure is: the input layer receives the material feature vector of the new material, for example, 64 dimensions, connects two hidden layers containing 32 and 16 neurons respectively, the activation function is ReLU, and the output layer generates a set of correction amounts with the same dimension as the number of parameters to be corrected .

[0117] The finally generated firmware parameter set is output in the header file or JSON format, which provides an optimal solution for high-performance electric tools such as aviation riveting guns, considering performance, resources, and generalization ability.

[0118] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present application and not to limit it. Although the present application has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present application can be modified or replaced equivalently without departing from the spirit and scope of the technical solutions of the present application.

Claims

1. A machine learning based power tool operating parameter optimization system, characterized by, The system comprises a strategy setting unit, a data acquisition unit, a feature processing unit, an algorithm fusion unit, a decision control unit, an execution feedback unit and an application output unit; The strategy setting unit is configured to generate strategy parameters; The data acquisition unit is configured to collect basic data, wherein the basic data comprises three-dimensional model data, material physical property data and microcontroller unit resource limitation parameters; The feature processing unit is configured to receive the basic data and encode the basic data to generate a unified feature vector; The algorithm fusion unit is configured to receive the unified feature vector and perform multi-objective optimization calculation to generate a Pareto optimal front; The decision control unit is configured to select from the Pareto optimal front to generate a to-be-verified execution parameter vector; The execution feedback unit is configured to perform high-precision simulation on the to-be-verified execution parameter vector to generate real target values and new labeled samples; The algorithm fusion unit is further configured to receive the new labeled samples to update an internal surrogate model; The application output unit is configured to generate a firmware parameter set based on an optimal solution confirmed by the decision control unit; The algorithm fusion unit performs the multi-objective optimization calculation in the following process: A surrogate model based on a Transformer architecture is constructed to predict the predicted values of three core optimization targets, namely, job performance, resource consumption and generalization capability; A multi-objective evolutionary algorithm is adopted, and the surrogate model is used as a fitness function to perform evolutionary calculation on the execution parameter vector to generate the Pareto optimal front; The decision control unit adopts a reinforcement learning decision agent, and the operation mechanism of the reinforcement learning decision agent is as follows: States are defined, including the number of Pareto front solutions, the target vector of the optimal performance solution, the target vector of the optimal health degree solution and the normalized state vector of the target space hyper-volume; Actions are defined, which are to select a specific solution from the Pareto front as the to-be-verified execution parameter vector; Rewards are defined, which are calculated according to the real simulation results of the to-be-verified execution parameter vector obtained by the execution feedback unit, combined with the expected performance targets and strategy weights defined by the strategy setting unit, and include a penalty term; If the real health degree corresponding to the real simulation results is lower than a preset health degree threshold, the penalty term is triggered; if the real health degree is greater than or equal to the preset health degree threshold, the penalty term is not triggered; The execution feedback unit generates verification results and new labeled samples; The verification results include accurate stress fields and real health degrees, real job performance, real resource consumption and real generalization capability calculated by a high-precision simulation engine; wherein the real job performance, real resource consumption and real generalization capability jointly constitute the real target values; The new labeled samples are formed by pairing the input to-be-verified execution parameter vector and the real target values output by simulation.

2. The machine learning based power tool operating parameter optimization system of claim 1, wherein, The feature processing unit generates the unified feature vector in the following process: The material physical property data is processed by a material attribute structural feature extractor to output a material feature vector; The stress-strain field three-dimensional grid data obtained after simulation by the execution feedback unit is processed by a digital twin state tensor encoder to output a state tensor; The material feature vector and the state tensor are spliced to generate the unified feature vector.

3. The machine learning based power tool operating parameter optimization system of claim 1, wherein, The core optimization objective is calculated in the following manner: The job performance is calculated by the reciprocal of the single riveting time predicted by the proxy model; The resource overhead is calculated by weighted summation and normalization of the memory occupation and CPU computation period predicted by the proxy model, combined with the upper limit of hardware resources obtained from the data acquisition unit and the weight coefficient defined by the strategy setting unit; The generalization ability is characterized by calculating the mathematical expectation of the self-defined health function under the normal distribution disturbance of the material yield strength.

4. The machine learning based power tool operating parameter optimization system of claim 3, wherein, The self-defined health function is calculated based on the improved logistic function, combined with the maximum equivalent stress calculated by the execution feedback unit, the workpiece material yield strength read from the data acquisition unit, and the safety factor threshold and health degree curve steepness coefficient defined by the strategy setting unit.

5. The machine learning based power tool operating parameter optimization system of claim 1, wherein, The firmware parameter set includes motor control parameters, process monitoring parameters, and material adaptation parameters; The motor control parameters include current loop PID parameters, speed loop feedforward coefficients, PWM frequency, and time-torque anchor point sequences of the target speed segmented curve; The process monitoring parameters include impact force sensor trigger threshold, motor stall detection current threshold, and displacement-force comprehensive logic threshold for job completion judgment; The material adaptation parameters include multiple sets of parameter collections preset for different materials, and the weight and bias parameters of a lightweight neural network for calculating the correction amount of new material parameters.

Images