A process parameter self-adaptive optimization method and system for battery shell cutting

Abstract

The application discloses a process parameter self-adaptive optimization method and system for battery shell cutting, and relates to the field of manufacturing control.The method comprises process parameter-quality mapping modeling based on a physical information deep neural network, adaptive parameter optimization based on a double-agent reinforcement learning, wear compensation, closed-loop continuous learning and knowledge sedimentation.The application embeds shear mechanics theory into a loss function through a physical information neural network to suppress prediction distortion, realizes real-time adjustment through a planning and regulation double-agent architecture, compresses batch adaptation time through meta-learning and transfer learning, and realizes cross-line knowledge sharing under the premise of protecting data privacy through federated learning.The application effectively improves cutting quality consistency, shortens process debugging time, prolongs die life, and enhances system self-adaptive capability.

Description

Technical Field

[0001] This invention relates to the field of manufacturing control, and in particular to an adaptive optimization method and system for process parameters of battery spring cutting. Background Technology

[0002] As a key precision structural component in the battery connection system, the cutting quality of battery springs directly affects the reliability of the electrical connection and the overall safety of the battery pack. Battery springs are mostly made of thin strip materials such as copper alloys and stainless steel, possessing characteristics of being thin, soft, and brittle. The cutting process must ensure micron-level dimensional accuracy while effectively controlling defects such as burrs and corner collapses. However, existing processes generally employ fixed parameters or manual experience-based parameter adjustment, making it difficult to cope with dynamic changes such as material batch fluctuations, mold wear, and operating condition drift, leading to unstable yields and long setup times.

[0003] Currently, the cutting of spring sheets for new energy batteries faces three major challenges: 1) Difficulty in controlling cutting quality; burrs, dimensional deviations, and other issues can easily lead to safety hazards, especially when cutting multi-layer materials simultaneously; 2) Process parameters rely on manual experience, and parameter tuning is time-consuming, affecting flexible production; 3) Lack of adaptive capabilities, unable to cope with mold wear, material batch differences, and environmental changes. The technical essence lies in the multi-parameter coupling and nonlinearity of this process, where traditional methods lack accurate modeling and dynamic decision-making capabilities. An adaptive optimization framework integrating multimodal perception, deep learning, and reinforcement learning needs to be constructed. Summary of the Invention

[0004] To address the technical problems in the prior art, this invention provides an adaptive optimization method and system for process parameters of battery spring sheet cutting.

[0005] This invention is achieved through the following technical solution: An adaptive optimization method for process parameters in battery spring sheet cutting includes: Multimodal process sensing and data acquisition; including the use of piezoelectric force sensors, accelerometers, acoustic emission sensors and infrared thermal imaging modules to measure force-displacement curves, three-dimensional vibration characteristics, high-frequency elastic wave signals and temperature distribution, as well as the use of industrial cameras for online detection of cutting quality; Process parameter-quality mapping modeling; constructed based on physical information deep neural network, inputting adjustable process parameter vectors and outputting quality indicators; and performing multi-fidelity transfer learning and uncertainty quantification; Adaptive parameter optimization; reinforcement learning based on a dual-agent system of planning and control agents, and the introduction of a meta-learning fast adaptation mechanism; Multi-objective optimization and parameter generation with wear compensation; including online identification of mold wear state and generation of wear compensation parameters for soft compensation and tool change warning; Closed-loop continuous learning and knowledge accumulation; including establishing a priority experience replay pool and constructing a process knowledge graph, conducting cross-production line collaboration based on federated learning, and handling human-machine collaborative anomalies.

[0006] Furthermore, the multi-fidelity transfer learning framework is constructed based on a global base model and local adapters; for new batches of materials, a small number of trial-cut samples are collected to fine-tune the last N layers of the global base model.

[0007] Furthermore, the planning agent adopts a Deep Q-Network architecture; the inputs are the spring feature vector and material properties, including material thickness, spring width, bending angle, material Brinell hardness, tensile strength, and material elongation; the output is an initial process parameter suggestion; the control agent uses a Soft Actor-Critic algorithm architecture to fine-tune the process parameters in real time.

[0008] Furthermore, the loss function of the reinforcement learning process, combined with the dense reward function, is expressed as follows: in, As a quality intermediate reward, For process rewards, For the final quality award, The Dirac function is activated only at the end of the cutting process; the intermediate quality reward is designed based on transient feature vectors, which include force drop slope, peak punching force, and root mean square acoustic emission; the process reward is based on the adjustment direction of parameters guided by action amplitude penalty and safety constraint violation penalty.

[0009] Furthermore, the meta-learning rapid adaptation mechanism introduces the MAML meta-learning framework into the regulatory agent. During the meta-training phase, an initial parameter is obtained by training on multiple historical batches of data, enabling the parameter to quickly adapt to new tasks after a small number of gradient updates.

[0010] Furthermore, the wear compensation parameters are based on the online identification results of the mold wear state, and the compensation objects include cutting pressure compensation, cutting speed compensation, gap compensation, clamping force compensation, and bite angle compensation.

[0011] Furthermore, the federated learning-based cross-production line collaboration is used for enterprises with multiple production lines. Each production line independently trains its model using local data and only uploads model parameter updates to the central server. The server aggregates these updates to generate a global model and then distributes it to each production line.

[0012] Furthermore, the human-machine collaborative anomaly handling involves automatically triggering a human-machine collaborative process when the system encounters an anomaly that exceeds the current knowledge boundary. The specific process includes recording the anomaly scene data when the anomaly detection score reaches a set threshold, and pushing a diagnostic request to the engineer's terminal. The engineer then provides handling opinions through the interactive interface.

[0013] The present invention also provides an adaptive optimization system for process parameters of battery spring sheet cutting, based on the adaptive optimization method for process parameters of battery spring sheet cutting as described above, comprising: The multimodal process sensing and data acquisition module includes a piezoelectric force sensor, an accelerometer, an acoustic emission sensor, and an infrared thermal imaging module, used to measure force-displacement curves, three-dimensional vibration characteristics, high-frequency elastic wave signals, and temperature distribution. The process parameter-quality mapping modeling module is built on a physical information deep neural network. It takes an adjustable process parameter vector as input and outputs quality indicators. It also performs multi-fidelity transfer learning and uncertainty quantification. The adaptive parameter optimization module is based on dual-agent reinforcement learning of planning agent and regulation agent, and introduces a meta-learning fast adaptation mechanism. The multi-objective optimization and parameter generation module with wear compensation is used to identify the wear state of the mold online and generate wear compensation parameters for soft compensation and tool change warning. The closed-loop continuous learning and knowledge accumulation module is used to establish a priority experience playback pool and construct a process knowledge graph. Based on federated learning, it enables cross-production line collaboration and realizes a closed loop for human-machine collaborative anomaly handling.

[0014] Furthermore, to achieve the above objectives, the present invention also provides a computer-readable storage medium storing program instructions for an adaptive optimization method of process parameters for cutting battery spring sheets. The program instructions for the adaptive optimization method of process parameters for cutting battery spring sheets can be executed by one or more processors to implement the steps of the adaptive optimization method of process parameters for cutting battery spring sheets as described above.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention combines planning intelligent agents with knowledge graph retrieval, freeing the determination of initial parameters from dependence on human experience, improving the accuracy of initial parameters, and significantly shortening the process debugging time.

[0016] 2. This invention constructs a process parameter-quality mapping model through a physical information neural network (PINN), embeds shear mechanics theory into the loss function, effectively suppresses prediction distortion in the edge regions of the parameter space, improves the prediction accuracy in nonlinear abrupt regions, and avoids batch quality accidents caused by parameter out-of-bounds errors.

[0017] 3. This invention achieves soft compensation for mold wear through online wear state identification and compensation parameter generator, enabling qualified products to be produced even under moderate wear conditions. It also provides scientific tool change warnings to reduce the risk of sudden downtime.

[0018] 4. This invention employs a dual-agent reinforcement learning architecture, where the planning agent provides initial values ​​and the regulating agent performs millisecond-level real-time adjustments based on transient feature vectors. Combined with a dense reward function design, this reduces the high volatility of burrs during the cutting process, significantly improving quality consistency.

[0019] 5. This invention uses meta-learning rapid adaptation mechanism and multi-fidelity transfer learning to reduce the adaptation time for material differences between batches from the traditional hours to minutes, greatly reducing the required number of trial cut samples and significantly improving the flexible response capability of the production line.

[0020] 6. This invention achieves cross-production line collaboration through a federated learning framework, enabling knowledge sharing while protecting the data privacy of each production line, accelerating model convergence and improving generalization ability; through a closed loop of human-machine collaborative anomaly handling, the system has continuous learning capabilities, and the system's capabilities are continuously enhanced as production data accumulates and engineers' experience is refined. Attached Figure Description

[0021] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments of this application and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings: Figure 1 This is a schematic flowchart of a method for adaptive optimization of process parameters for cutting battery spring contacts according to an embodiment of this application. Detailed Implementation

[0022] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0023] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. The present invention can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that, in the absence of conflict, the following embodiments and features in the embodiments can be combined with each other. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0024] It should also be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. The illustrations only show the components related to the present invention and are not drawn according to the actual number, shape and size of the components in the actual implementation. In the actual implementation, the shape, quantity and proportion of each component can be arbitrarily changed, and the layout of the components may also be more complex.

[0025] See Figure 1 An adaptive optimization method for process parameters in battery spring sheet cutting includes the following steps: S1: Multimodal process sensing and data acquisition; S11: Sensor network deployment; A piezoelectric force sensor is installed on the back of the punch in the cutting die, and the sampling frequency is set to [frequency value missing]. ≥ 50 kHz was used to acquire the complete force-displacement curve of the punching force over time. Three miniature accelerometers were arranged at equal angular intervals along the circumference of the die, and the sampling frequency was set to ≥ 50 kHz. ≥ 100 kHz, used to capture the three-dimensional vibration characteristics during the cutting process. An acoustic emission sensor is embedded 1 mm to 3 mm from the cutting line on the side of the die cutting edge, and the sampling frequency is set to 100 kHz. ≥ 1 MHz, used to monitor high-frequency elastic wave signals generated when materials fracture. A miniature infrared thermal imaging module with a resolution of 120×90 pixels and a frame rate of 50 Hz is deployed on the side of the cutting area to monitor the temperature distribution near the cutting edge and the area where the spring is cut in real time.

[0026] The signals collected by the aforementioned sensors constitute the transient feature vector Φ(t) of the cutting process: Where F(t) is the punching force at time t, and the unit is Newton; Δx is the acceleration vector in three directions, in meters per second squared; AE(t) is the root mean square amplitude of the acoustic emission signal, in volts; T(x,y,t) is the temperature field distribution of the cut region at coordinates (x,y), in Kelvin or degrees Celsius.

[0027] S12: Adaptive sampling and edge computing; Edge computing units are deployed at the acquisition end to execute an adaptive sampling strategy: a punch position function s(t) is defined, in millimeters. When the punch position satisfies s(t)... contact When the system determines that it is in an idle phase, it uses the reference sampling rate f. base = 0.1 ×f s Data is collected when the punch position satisfies s(t) ≥ s contact When the system determines that it is in the contact cutting stage, it automatically switches to the highest sampling rate f. s ​Data collection is performed, where s contact The initial contact position between the punch and the material is determined through offline calibration.

[0028] The edge computing unit performs real-time feature extraction on the acquired raw signal and calculates the following transient feature vector Ψ: in, The peak punching force is expressed in Newtons (N), and the calculation formula is as follows: = max{F(t)}, t∈[ , ], The moment of contact initiation. This is the moment when the materials are completely separated; The slope of the force drop is expressed in Newtons per second, and the formula is as follows: , The punching force at the moment of separation. This corresponds to the peak force time. The vibrational envelope energy is dimensionless and is calculated using the following formula: ; is the root mean square value of the acoustic emission signal, in volts; The highest temperature in the cutting area, expressed in Kelvin; The maximum rate of temperature change, expressed in Kelvin per second, characterizes the intensity of friction.

[0029] The edge computing unit only uploads the compressed transient feature vector Ψ to the central database, while the original signal is cached locally for 7 days for fault tracing.

[0030] S13: Online inspection of cutting quality; A high-resolution industrial camera is deployed at the discharge port of the cutting equipment, with a lens magnification adjustable from 5x to 20x. Combined with a ring LED light source, it captures images of the cut spring sheet. An edge detection algorithm is used to extract the spring sheet contour, and the following quality indicators are calculated: Burr height h, in micrometers, is defined as the maximum vertical distance between the material protrusions at the cut edge. in, Let be the burr height value of the i-th sampling point, and N be the total number of sampling points along the cutting edge.

[0031] Dimensional deviation Δd, in millimeters, is defined as the absolute value of the difference between the measured dimension and the design dimension. in, These are actual measured dimensions. For design dimensions.

[0032] Surface quality score s, with a value range of [0,1], where 1 indicates no visible surface defects: in, This refers to the total area of ​​defects such as scratches and indentations. The total surface area of ​​the shrapnel. The number of scratches is denoted by γ, which is the penalty coefficient for the number of scratches and is calibrated to a value between 0.01 and 0.05 through experiments.

[0033] The quality index vector Q is defined as: S2: Process parameter-quality mapping modeling based on physical information deep neural network; S21: Input / output space definition; Define the input space X as a vector of adjustable process parameters: Where p is the cutting pressure, in megapascals, and its value ranges from 1.5 to 3 times the yield strength of the material; v is the cutting speed, in millimeters per second, and its value ranges according to the equipment capacity; g is the die edge clearance, in millimeters, and its value ranges from 3% to 12% of the material thickness; α is the bite angle, in degrees, and its value ranges from 1° to 5°. The lateral clamping force is measured in Newtons and is 1.2 to 2 times the minimum clamping force required to prevent material slippage.

[0034] Define the output space Y to include quality metrics: S22: Establishment of a physical information neural network architecture; A physical information neural network (PINN) is constructed as the prediction model. The network structure includes: an input layer that receives a 5-dimensional process parameter vector; a hidden layer consisting of 4 fully connected layers with 128, 256, 128, and 64 neurons in each layer, and the activation function being the ReLU function, defined as ReLU(z) = max(0, z); and an output layer that simultaneously outputs the predicted values ​​of three quality indicators and the predicted value of the auxiliary physical quantity—the peak value of the cutting force.

[0035] The mathematical expression for a network is: in, For a neural network with parameters θ, where θ includes the weight matrices W of each layer... (l) and bias vector b (l) ; This is the prediction vector for the quality index. This is the predicted value of the peak cutting force.

[0036] S23: Design of composite loss function; Define the loss function as a weighted sum of data loss and physical loss: in, The physical loss weighting coefficient was determined to be between 0.1 and 0.5 through cross-validation.

[0037] Data loss Using mean square error form: Where M is the batch size. Let j be the true quality index vector of the j-th sample. Let ||·||2 represent the corresponding predicted value, and let ||·||2 represent the Euclidean norm.

[0038] Physical loss Defined as: in, Let be the measured peak cutting force of the j-th sample, ε be a small constant of 1×10^{-6} to prevent division by zero errors, and μ be the consistency loss weight of 0.2. Let j be the predicted spur height of the j-th sample. Let the peak value of the predicted cutting force be the j-th sample. Let j be the die cutting edge clearance of the j-th sample. The physical consistency constraint is designed based on shear mechanics theory: in, The reference burr height threshold is set to 30μm; The peak cutting force is taken as the average value measured under standard working conditions; The optimal gap value is 5% to 8% of the material thickness; sign is the sign function. The physical meaning of this constraint is: when the gap deviates from the optimal value, the increase in burrs and the decrease in cutting force should show a consistent relationship. If the model prediction violates this physical law, a penalty is imposed.

[0039] The aforementioned loss function design, by introducing physical consistency constraints, forces the model output to satisfy the physical laws governing gaps, burrs, and peak forces, thus solving the prediction distortion problem in sparse data regions. The multi-task learning architecture simultaneously predicts burr height and peak force, sharing feature layers to improve generalization ability and suppress overfitting. The dimensionless relative bias form enhances numerical stability and avoids the influence of dimensions. An interpretable penalty mechanism embeds engineering semantics, enhancing model credibility. Compared to traditional purely data-driven methods, this design has significant advantages in sample efficiency, prediction accuracy, and engineering safety, providing a reliable prediction engine for process optimization.

[0040] S24: Multifidelity transfer learning; Construct a multi-fidelity transfer learning framework, including a global base model. and local adapter The global base model includes Large-scale datasets from historical batches Pre-train until convergence.

[0041] For new batches of materials, collect a small number of test cut samples. Quantity set to = 10 to 20. An adapter fine-tuning strategy is adopted: the parameters of the first L-2 layers of the global base model are frozen, and only the parameters of the last 2 layers and the output layer are fine-tuned. The fine-tuning loss function is: Where η is the fine-tuning regularization coefficient with a value of 0.01. The first term represents the initial adapter parameters, and the second term is the hold term, which prevents the adapter from overfitting to a small number of samples.

[0042] S25: Uncertainty Quantification The Monte Carlo Dropout method is used to estimate prediction uncertainty. During the inference phase, each input X undergoes K forward propagations, where K ranges from 50 to 100, and the proportion of random drops during each forward propagation is... The neurons. The predicted mean and variance are calculated as follows: in, This represents the prediction result of the k-th forward propagation. Define the confidence index. , This represents the range of values ​​for the quality indicator. When the confidence level C < 0.7, the system determines that the current parameter combination is in a sparse data region or close to the mutation boundary, refuses to execute, and triggers manual review.

[0043] S3: Adaptive parameter optimization based on dual-agent reinforcement learning; This invention employs a planning agent that takes the geometric features and material properties of the spring sheet as input and outputs suggested initial process parameters, replacing the traditional trial-and-error method and quickly providing reasonable initial values. A regulating agent is used, taking transient feature vectors collected in real-time during the cutting process as input and outputting millisecond-level parameter fine-tuning amounts, achieving on-the-fly adjustment. In other words, the planning agent determines the baseline, and the regulating agent dynamically corrects the settings. Combining a dense reward function and a meta-learning rapid adaptation mechanism, it achieves end-to-end adaptive optimization from initial value setting to real-time compensation.

[0044] S31: Planning Agent Design The planning agent adopts a Deep Q-Network architecture, and its inputs are the feature vector of the spring and the material performance parameters: Where t is the material thickness in millimeters; w is the spring width in millimeters; The bending angle is in degrees; HB is the Brinell hardness of the material, which is dimensionless. δ represents tensile strength, measured in megapascals (MPa); δ represents elongation of the material, measured as a percentage.

[0045] The output of the planning agent is a suggestion of initial process parameters: The agent is trained using the DQN algorithm, and the Q-value function is defined as follows: Where φ is the Q-network parameter, This is the cumulative reward after executing parameter X. During training, the agent learns from a database of historical success stories. The database contains learning mapping relationships. Each case study contains a feature vector of the shrapnel, initial parameters, and the corresponding final quality result.

[0046] The loss function for planning the agent is: Where Z is the state, i.e., the feature vector of the spring; X is the action, i.e., the initial process parameter suggestions output by the planning agent; R is the reward, the immediate reward obtained after executing action X; Z′ is the next state, the state to which the system transitions after executing action X; D is the experience replay pool, storing historical interaction samples; γ is the discount factor, used to balance the importance of current rewards and future rewards; X′ is all possible actions in the next state Z′. Let be the target Q-network, where ϕ− is a historical copy of ϕ used for stable training; Qϕ is the current Q-network used to estimate the value of the current state-action pair.

[0047] S32: Control agent design; The control agent employs a Soft Actor-Critic algorithm architecture, with the input being the real-time transient feature vector Ψ and the output being the real-time fine-tuning of process parameters. The core of the SAC algorithm is to minimize the KL divergence between the policy and the target distribution, and its policy loss function is: Where ψ is the policy network parameter, s t = Ψ(t) represents the state at time t, a t = ΔX(t) represents the action at time t. The action distribution output by the policy network, For temperature coefficient, control exploration - utilizing equilibrium, ω is the soft Q function, and ω is the Q network parameter.

[0048] The loss function of the soft Q function is: in, The instantaneous reward at time t. The target Q-network parameters are denoted as .

[0049] S33: Dense Reward Function Design; The reward function for the immediate reward employs a dense reward design, continuously providing feedback signals during the pruning process: in, As a quality intermediate reward, For process rewards, For the final quality award, It is a Dirac function that is activated only at the end of the cut.

[0050] Intermediate Quality Bonus Design based on transient feature vectors: in, For the real-time calculation of the force drop slope, The slope of the target force reduction is obtained through historical best clipping learning; For real-time monitoring of the peak punching force, For reference peak force; This represents the root mean square value of real-time acoustic emission. This represents the maximum permissible value for acoustic emission signals. , , Let be the weighting coefficient, satisfying .

[0051] Process rewards Used to guide the direction of parameter adjustment: in, The penalty coefficient for the range of motion is set to 0.01. Let be the Euclidean norm of the action vector; The penalty coefficient for violating safety constraints is set to 10. (·) is an indicator function, which is used when the parameter X(t) exceeds the safety region. The value is 1 if it is true and 0 otherwise.

[0052] Final Quality Award Defined as: in, The maximum allowable burr height is set at 50 μm. The maximum allowable dimensional deviation is set at 0.05 mm. The surface quality weighting coefficient is set to 0.5.

[0053] S34: Meta-learning rapid adaptation mechanism The MAML meta-learning framework is introduced into the regulatory agent. During the meta-training phase, an initial set of parameters is obtained by training on multiple historical batches of data. This allows the parameter to quickly adapt to new tasks after a small number of gradient updates. The loss function for meta-learning is defined as: Where p(𝒯) is the task distribution, and each task corresponds to a material batch; For the i-th task; The meta-learning rate has a value of 0.001. For the task loss function on; For loss function Regarding initial parameters The gradient.

[0054] When encountering a new batch of materials, a small amount of trajectory data is used for rapid adaptation. The adaptation and update rules are as follows: in, These are the optimal initial parameters obtained during the meta-learning phase training. To accommodate the learning rate, a value of 0.01 is used. For the loss function on the new task, adaptation can be completed with only 1 to 2 trial trajectory cuts; For loss function Regarding parameters The gradient.

[0055] S4: Multi-objective optimization and parameter generation with wear compensation S41: Online identification of mold wear condition A one-dimensional convolutional neural network model is constructed for wear state recognition. The input is the time series of transient feature vector Ψ, and the output is the wear state level w ∈ {0, 1, 2, 3}, where 0 represents healthy state, 1 represents slight wear, 2 represents moderate wear, and 3 represents near-life state.

[0056] Extract wear-sensitive feature vectors from transient feature vectors : in, The difference between the current peak force and the initial peak force. This represents the peak force in the initial state of the mold. This refers to the change in the slope of the force drop. The energy of the high-frequency component of the vibration signal. The ratio of the two values ​​represents the total vibration energy, reflecting the proportion of high-frequency vibration; skewness(F(t)) is the skewness coefficient of the force curve, reflecting the asymmetry of the force distribution.

[0057] The wear state recognition model is a 1D-CNN, with the following structure: convolutional layers using 32 convolutional kernels of size 3, pooling layers using max pooling, fully connected layers containing 64 neurons, and an output layer using the softmax activation function to output the probability of each wear state. Wear condition is selected as the category with the highest probability: S42: Wear Compensation Parameter Generator Establish parameter compensation function Describe the relationship between wear condition and parameter compensation amount: in, This represents the optimal parameter vector under healthy mold conditions. The amount of compensation that needs to be applied.

[0058] The compensation function adopts an interpretable form derived from shear mechanics theory: For cutting pressure compensation : in, This is the initial value of the cutting pressure; The maximum value for the wear condition is 3. and The coefficients are obtained by fitting the experimental data, and their values ​​are 0.15 and 0.05, respectively.

[0059] For cutting speed compensation : The negative sign indicates that the speed decreases as wear increases. The coefficients are fitted using experimental data, and their value is 0.08. This is the initial value for the cutting speed.

[0060] For gap compensation amount : The negative sign indicates that the clearance decreases as wear increases to compensate for the dulling of the cutting edge. The coefficients are obtained by fitting the experimental data, with values ​​of 0.05 and 0.02. This is the initial value for the gap.

[0061] Compensation for clamping force : in, The coefficients are fitted using experimental data, and their value is 0.10. This is the initial value of the clamping force.

[0062] Biting angle compensation : in, The coefficients are fitted using experimental data, and their value is 0.03. This is the initial value for the bite angle.

[0063] The final wear-compensated parameters are as follows: When the wear state w=2 (moderate wear), the compensation mechanism is automatically activated, and the compensated parameters are sent out for execution.

[0064] S43: Multi-objective Pareto optimization Define three optimization objective functions: Minimize burr height objective: Where h(X,w) is the predicted burr height under parameter X and wear state w, which is predicted by the PINN model in step S2.

[0065] The objective of maximizing cutting efficiency (converted to minimization): in, For the cutting cycle time, This is the reference cycle time.

[0066] Objective: Minimize mold wear rate in, The wear rate is calculated using the following empirical formula: in, The wear coefficient is determined experimentally; p is the cutting pressure. The pressure index is set to 1.2; v is the cutting speed; The speed exponent is set to 0.8; The activation energy is set to 20 kJ / mol; The gas constant is taken as 8.314 J / (mol·K); Temperature of the cutting contact area; The amplification factor for the wear condition is set to 0.3.

[0067] The multi-objective optimization problem is formulated as follows: The Pareto front is solved using the NSGA-III algorithm. The population size is initialized to N. pop =100, the number of iterations is N gen =200. In each generation, selection, crossover, and mutation operations are performed, with a crossover probability of 0.9 and a mutation probability of 0.1. The final output is the Pareto optimal solution set: Here, the symbol ≺ denotes the Pareto dominance relation: F(X') ≺ F(X) if and only if for all k ∈ {1,2,3} f k (X') ≤ f k (X) and there exists at least one k such that the strict inequality holds.

[0068] An interactive selection interface is provided, allowing engineers to choose the solution vector based on the current production priorities: in, , , The weighting coefficients specified by the user satisfy .

[0069] S44: Soft Compensation Execution and Tool Change Warning When the wear condition w=2, the system automatically calculates the parameters after compensation. The order was then issued and implemented. Simultaneously, the height of the burrs after compensation was monitored. : Define compensation effect indicators: in, This represents the burr height before compensation. When... A value ≥ 0.8 indicates that the compensation is effective; when A value less than 0.8 indicates insufficient compensation.

[0070] A tool change warning is triggered when any of the following conditions are met: Condition 1: w=3 and > This means that even after moderate wear compensation, the quality requirements still cannot be met. Condition 2: Predict the remaining available attempts Less than the safety threshold N threshold =500 times.

[0071] The remaining available attempts prediction model is as follows: in, The threshold value for the critical wear state is 3.5; Δw / ΔN is the wear increment per unit number of cuts, which is obtained by linear regression of historical data.

[0072] S5: Closed-loop continuous learning and knowledge accumulation S51: Priority Experience Replay Pool Establish a unified experience replay pool Each empirical sample contains complete trajectory information: in, For the feature vector of the shrapnel, Let Ψ(t) be the initial parameter, ΔX(t) be the transient feature vector sequence, ΔX(t) be the control action sequence, and r(t) be the reward sequence. For the final quality result, and These represent the wear status at the beginning and end of the cutting process, respectively.

[0073] Define sample priority : in, As the final reward, For average reward; ζ represents the set of boundary cases, including samples near the mutation region and samples with abnormal quality fluctuations; ζ is the boundary case reward coefficient with a value of 0.5; ξ is the gradient priority coefficient with a value of 0.3; ∇ℒ(θ) is the loss gradient of the current model with respect to this sample.

[0074] The sampling probability is directly proportional to the priority: Wherein, κ is the priority index with a value of 0.6.

[0075] S52: Construction of Process Knowledge Graph Building a process knowledge graph , where the set of nodes include: Material details include material grade, thickness, hardness, and tensile strength. The mold node has attributes including cutting edge material, initial clearance, and current wear status. The parameter node has the attribute of the process parameter vector X; A quality node, whose attribute is a quality metric vector Q; Environmental nodes, whose attributes include temperature and humidity.

[0076] edge set The edges represent the relationships between nodes, and their weights represent the strength of those relationships. For new tasks, similarity retrieval is performed within the knowledge graph. Where, φ i The importance weight of the i-th attribute is calculated using information gain. The most similar historical cases are retrieved, and their parameters are extracted as a reference for initial suggestions.

[0077] Knowledge graph update rules: After each successful pruning, the new case is added to the graph as a node subgraph; when the graph size exceeds N... max When there are 10,000 nodes, a pruning strategy based on information entropy is used to delete the node with the least information.

[0078] S53: Federated Learning for Cross-Production Line Collaboration For a company with M production sites, a federated learning framework is constructed. Let D be the local dataset of the m-th site. m The local model parameter is θ m The central server maintains the global model parameter θ. global .

[0079] The steps for each round of federated learning are as follows: (1) The central server issues the current global model parameters. To each venue; (2) Each site will conduct 5 rounds of local training on local data: in, The local learning rate is set to 0.01. For local loss functions; Upload model parameters updates for each site To the central server; (4) The central server performs federated aggregation: in, For the first Global model parameters after round of federated learning For the first Global model parameters at the start of a round of federated learning; For the first The number of model parameter updates uploaded for each site.

[0080] The global loss function for federated learning is: The FedAvg algorithm is used for aggregation, eliminating the need to exchange original data and protecting the data privacy of each production line.

[0081] S54: Closed-loop system for handling anomalies in human-machine collaboration When the system encounters an anomaly that exceeds its current knowledge boundaries, it automatically triggers a human-machine collaboration process. Define anomaly detection metrics: Abnormal scores: in, The standard deviation of uncertainty calculated for S25. This is the baseline value for the standard deviation under normal operating conditions; Historical average transient characteristics; The range of values ​​for the transient feature; This is within the acceptable quality range.

[0082] when When the threshold value is 2.0, the system records abnormal on-site data and pushes a diagnostic request to the engineer's terminal. The pushed content includes: Feature vectors of abnormal samples and Key segments of the transient feature vector sequence Ψ(t); quality inspection image and recognition results; uncertainty calculated in step S25.

[0083] Engineers provide handling suggestions through an interactive interface, including: classification of the anomaly cause, suggested parameter adjustments, and whether equipment maintenance is required. The system stores this anomaly handling record in the experience replay pool and marks it as a high-priority sample. Simultaneously, the anomaly case is added to the knowledge graph as a reference for similar scenarios in the future.

[0084] The playback weights for human-machine collaboration records are updated as follows: in, The basic priority set manually has a value of 0.8. The original priority remains.

[0085] In this implementation, the following is achieved: 1. This invention combines planning intelligent agents with knowledge graph retrieval, freeing the determination of initial parameters from dependence on human experience, improving the accuracy of initial parameters, and significantly shortening the process debugging time.

[0086] 2. This invention constructs a process parameter-quality mapping model through a physical information neural network (PINN), embeds shear mechanics theory into the loss function, effectively suppresses prediction distortion in the edge regions of the parameter space, improves the prediction accuracy in nonlinear abrupt regions, and avoids batch quality accidents caused by parameter out-of-bounds errors.

[0087] 3. This invention achieves soft compensation for mold wear through online wear state identification and compensation parameter generator, enabling qualified products to be produced even under moderate wear conditions. It also provides scientific tool change warnings to reduce the risk of sudden downtime.

[0088] 4. This invention employs a dual-agent reinforcement learning architecture, where the planning agent provides initial values ​​and the regulating agent performs millisecond-level real-time adjustments based on transient feature vectors. Combined with a dense reward function design, this reduces the high volatility of burrs during the cutting process, significantly improving quality consistency.

[0089] 5. This invention uses meta-learning rapid adaptation mechanism and multi-fidelity transfer learning to reduce the adaptation time for material differences between batches from the traditional hours to minutes, greatly reducing the required number of trial cut samples and significantly improving the flexible response capability of the production line.

[0090] 6. This invention achieves cross-production line collaboration through a federated learning framework, enabling knowledge sharing while protecting the data privacy of each production line, accelerating model convergence and improving generalization ability; through a closed loop of human-machine collaborative anomaly handling, the system has continuous learning capabilities, and the system's capabilities are continuously enhanced as production data accumulates and engineers' experience is refined.

[0091] This invention also proposes an adaptive optimization system for process parameters in battery spring sheet cutting, based on the adaptive optimization method for process parameters in battery spring sheet cutting described above, including: The multimodal process sensing and data acquisition module includes a piezoelectric force sensor, an accelerometer, an acoustic emission sensor, and an infrared thermal imaging module, used to measure force-displacement curves, three-dimensional vibration characteristics, high-frequency elastic wave signals, and temperature distribution. The process parameter-quality mapping modeling module is built on a physical information deep neural network. It takes an adjustable process parameter vector as input and outputs quality indicators. It also performs multi-fidelity transfer learning and uncertainty quantification. The adaptive parameter optimization module is based on dual-agent reinforcement learning of planning agent and regulation agent, and introduces a meta-learning fast adaptation mechanism. The multi-objective optimization and parameter generation module with wear compensation is used to identify the wear state of the mold online and generate wear compensation parameters for soft compensation and tool change warning. The closed-loop continuous learning and knowledge accumulation module is used to establish a priority experience playback pool and construct a process knowledge graph. Based on federated learning, it enables cross-production line collaboration and realizes a closed loop for human-machine collaborative anomaly handling.

[0092] Furthermore, this embodiment of the invention also proposes a computer-readable storage medium storing program instructions for an adaptive optimization method of process parameters for cutting battery spring sheets. The program instructions for the adaptive optimization method of process parameters for cutting battery spring sheets can be executed by one or more processors to implement the steps of the adaptive optimization method of process parameters for cutting battery spring sheets as described above.

[0093] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made by those skilled in the art to the technical solutions of the present invention without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. An adaptive optimization method for process parameters in battery spring sheet cutting, characterized in that, include: Multimodal process sensing and data acquisition; including the use of piezoelectric force sensors, accelerometers, acoustic emission sensors and infrared thermal imaging modules to measure force-displacement curves, three-dimensional vibration characteristics, high-frequency elastic wave signals and temperature distribution, as well as the use of industrial cameras for online detection of cutting quality; Process parameter-quality mapping modeling; constructed based on a physical information deep neural network, inputting an adjustable process parameter vector and outputting quality indicators; And perform multi-fidelity transfer learning and uncertainty quantification; Adaptive parameter optimization; Reinforcement learning is based on a dual-agent system consisting of a planning agent and a control agent, and a meta-learning rapid adaptation mechanism is introduced. Multi-objective optimization and parameter generation with wear compensation; This includes online identification of mold wear conditions, generation of wear compensation parameters, and soft compensation and tool change warning; Closed-loop continuous learning and knowledge accumulation; This includes establishing a priority experience replay pool and constructing a process knowledge graph, conducting cross-production line collaboration based on federated learning, and handling human-machine collaborative anomalies.

2. The adaptive optimization method for process parameters of battery spring sheet cutting according to claim 1, characterized in that, The multi-fidelity transfer learning framework is constructed based on a global base model and local adapters; for new batches of materials, a small number of trial cut samples are collected to fine-tune the last N layers of the global base model.

3. The adaptive optimization method for process parameters of battery spring sheet cutting according to claim 1, characterized in that, The planning agent adopts a Deep Q-Network architecture; the inputs are the feature vector of the spring sheet and material properties, including material thickness, spring sheet width, bending angle, material Brinell hardness, tensile strength, and material elongation; the output is an initial process parameter suggestion; the control agent uses a Soft Actor-Critic algorithm architecture to fine-tune the process parameters in real time.

4. The adaptive optimization method for process parameters of battery spring sheet cutting according to claim 3, characterized in that, The loss function used in the reinforcement learning process, combined with a dense reward function, is expressed as follows: in, As a quality interim reward, For process rewards, For the final quality award, The Dirac function is activated only at the end of the cutting process; the intermediate quality reward is designed based on transient feature vectors, which include force drop slope, peak punching force, and root mean square acoustic emission; the process reward is based on the adjustment direction of parameters guided by action amplitude penalty and safety constraint violation penalty.

5. The adaptive optimization method for process parameters of battery spring sheet cutting according to claim 4, characterized in that, The meta-learning rapid adaptation mechanism introduces the MAML meta-learning framework into the regulatory agent. During the meta-training phase, an initial parameter is obtained by training on multiple historical batches of data, enabling the parameter to quickly adapt to new tasks after a small number of gradient updates.

6. The adaptive optimization method for process parameters of battery spring sheet cutting according to claim 1, characterized in that, The wear compensation parameters are based on the online identification results of the mold wear state, and the compensation objects include cutting pressure compensation, cutting speed compensation, gap compensation, clamping force compensation, and bite angle compensation.

7. The adaptive optimization method for process parameters of battery spring sheet cutting according to claim 1, characterized in that, The federated learning-based cross-production line collaboration is used for enterprises with multiple production lines. Each production line independently trains its model using local data and only uploads model parameter updates to the central server. The server aggregates these updates to generate a global model and then distributes it to each production line.

8. The adaptive optimization method for process parameters of battery spring sheet cutting according to claim 6, characterized in that, The human-machine collaborative anomaly handling process involves automatically triggering a human-machine collaborative process when the system encounters an anomaly that exceeds the current knowledge boundary. The specific process includes recording the anomaly scene data when the anomaly detection score reaches a set threshold and pushing a diagnostic request to the engineer's terminal. The engineer then provides handling opinions through the interactive interface.

9. An adaptive optimization system for process parameters in battery spring sheet cutting, based on the adaptive optimization method for process parameters in battery spring sheet cutting as described in any one of claims 1 to 8, characterized in that, include: The multimodal process sensing and data acquisition module includes a piezoelectric force sensor, an accelerometer, an acoustic emission sensor, and an infrared thermal imaging module, used to measure force-displacement curves, three-dimensional vibration characteristics, high-frequency elastic wave signals, and temperature distribution. The process parameter-quality mapping modeling module is built based on a physical information deep neural network. It takes an adjustable process parameter vector as input and outputs quality indicators. And perform multi-fidelity transfer learning and uncertainty quantification; The adaptive parameter optimization module is based on dual-agent reinforcement learning of planning agent and regulation agent, and introduces a meta-learning fast adaptation mechanism. The multi-objective optimization and parameter generation module with wear compensation is used to identify the wear state of the mold online and generate wear compensation parameters for soft compensation and tool change warning. The closed-loop continuous learning and knowledge accumulation module is used to establish a priority experience playback pool and construct a process knowledge graph. Based on federated learning, it enables cross-production line collaboration and realizes a closed loop for human-machine collaborative anomaly handling.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores program instructions for an adaptive optimization method of process parameters for cutting battery spring contacts, which can be executed by one or more processors to implement the steps of the adaptive optimization method of process parameters for cutting battery spring contacts as described in any one of claims 1 to 8.

Images