A Reinforcement Learning-Based Adaptive Adjustment Method and System for Lithium-ion Battery Disassembly Parameters

Abstract

This application provides a method and system for adaptive adjustment of lithium battery dismantling parameters based on reinforcement learning, comprising: S1, collecting force feedback data during the battery dismantling process, wherein the force feedback data originates from force sensors and strain gauges; S2, inputting the force feedback data as a state into a multi-objective reinforcement learning model; S3, the multi-objective reinforcement learning model calculates rewards for dismantling speed, safety, and energy consumption objectives and outputs dismantling parameter adjustment amounts; S4, updating the strategy of the multi-objective reinforcement learning model according to the rewards to adaptively adjust the dismantling parameters; S5, applying the dismantling parameter adjustment amounts to control the dismantling equipment to perform battery dismantling operations and inputting the new force feedback data after execution into the multi-objective reinforcement learning model. This application can significantly improve dismantling efficiency and safety, and reduce energy consumption and accident risks.

Description

Technical Field

[0001] This invention relates to the field of intelligent lithium battery disassembly technology, and in particular to a method and system for adaptive adjustment of lithium battery disassembly parameters based on reinforcement learning. Background Technology

[0002] The dismantling and recycling of used batteries is a key area for achieving resource recycling and environmental protection. Its importance lies in effectively recovering rare metals while preventing the leakage of hazardous substances, which is crucial for sustainable development. During battery dismantling, components such as the battery casing, electrodes, and electrolyte need to be safely separated through cutting and separation methods to ensure the smooth extraction of subsequent materials.

[0003] Current disassembly methods typically use preset, fixed parameters to control cutting power, pressure, and speed. These parameters are difficult to flexibly adapt to actual changes when dealing with different batteries. When battery types, aging levels, or material compositions differ, the mechanical resistance encountered during cutting will vary significantly. For example, the outer casing of some aged batteries becomes more brittle, while others experience increased local hardness due to internal structural deformation. If the same set of parameters is still used, the equipment is prone to applying excessive force in areas with higher hardness, causing tool wear or motor overload, while insufficient force in brittle areas will result in incomplete cutting.

[0004] These changes directly lead to complex fluctuations in the mechanical response, with the cutting force sometimes rising sharply and sometimes falling rapidly, accompanied by drastic changes in vibration and deformation. This real-time complexity of the mechanical response makes it difficult to maintain a stable state during the disassembly process. On the one hand, excessive local force may cause internal short circuits or thermal runaway in the battery; on the other hand, insufficient force will prolong the operation time and increase energy consumption. The equipment may also be damaged due to continuously bearing mismatched loads.

[0005] How to perceive these mechanical changes in real time during the disassembly process and adjust the cutting parameters accordingly to ensure operational safety, control energy consumption, and maintain a reasonable speed has become a key problem that urgently needs to be solved in the field of battery disassembly. Summary of the Invention

[0006] To address the technical problems mentioned in the background section, the first aspect of this invention provides an adaptive adjustment method for lithium battery disassembly parameters based on reinforcement learning. The method includes: S1, collecting force feedback data during the battery disassembly process, wherein the force feedback data originates from a force sensor and a strain gauge; S2, inputting the force feedback data as a state into a multi-objective reinforcement learning model.

[0007] S3, the multi-objective reinforcement learning model calculates rewards for disassembly speed, safety, and energy consumption targets and outputs disassembly parameter adjustment amounts; S4, the strategy of the multi-objective reinforcement learning model is updated according to the rewards to adaptively adjust the disassembly parameters; S5, the disassembly parameter adjustment amounts are applied to control the disassembly equipment to perform battery disassembly operations and the new force feedback data after execution is input into the multi-objective reinforcement learning model.

[0008] Furthermore, the force feedback data includes cutting force data and vibration deformation data.

[0009] Furthermore, step S1 includes:

[0010] Step S11: Calibrate the sensor;

[0011] Step S12: The disassembly operation is started with preset initial parameters.

[0012] Step S13: The force sensor continuously collects triaxial force data during the cutting process at a sampling frequency of not less than 100Hz;

[0013] In step S14, strain gauges collect strain values ​​at multiple measuring points at the same frequency.

[0014] Furthermore, step S2 includes: step S21, where a force sensor acquires cutting force data of the disassembled component in real time, and a strain gauge acquires vibration deformation data of the disassembled component in real time; step S22, where the cutting force data and the vibration deformation data are combined to form a current disassembly state vector; and step S23, where the current disassembly state vector is input into the state space of a multi-objective reinforcement learning model.

[0015] Further, step S3 includes: step S31, obtaining the disassembly speed index, safety index, and energy consumption index corresponding to the current disassembly action; step S32, weighting and summing the disassembly speed index, safety index, and energy consumption index according to preset weights to obtain a comprehensive reward value; step S33, inputting the comprehensive reward value into the reward function of the multi-objective reinforcement learning model; step S34, the multi-objective reinforcement learning model outputs the disassembly parameter adjustment amount for cutting power, pressure, and rotation speed according to the reward function.

[0016] Further, step S34 includes: step S341, obtaining the initial disassembly parameter range from the parameter adjustment rule library; step S342, combining the initial disassembly parameter range with the reward function to determine candidate adjustment strategies; step S343, selecting the disassembly parameter adjustment amount with the highest reward among the candidate adjustment strategies through the strategy network; and step S344, outputting the disassembly parameter adjustment amount to the disassembly equipment controller.

[0017] Further, step S341 includes: step S3411, storing cutting power thresholds, pressure thresholds, and rotation speed thresholds corresponding to various battery disassembly scenarios in the parameter adjustment rule library; step S3412, matching the corresponding scenario in the parameter adjustment rule library according to the current battery type; step S3413, extracting the cutting power threshold, pressure threshold, and rotation speed threshold under the matched scenario as the initial disassembly parameter range; and step S3414, dynamically updating the parameter adjustment rule library using new experience data during online learning.

[0018] Furthermore, step S4 includes: step S41, storing the current state, action, reward, and new state into the experience replay buffer; step S42, sampling batch experience data from the experience replay buffer; step S43, using the batch experience data to perform gradient updates on the policy network and value network; and step S44, the updated policy network outputs adaptive disassembly parameters for the next disassembly cycle.

[0019] Furthermore, step S5 includes: step S51, the disassembly equipment controller receives the disassembly parameter adjustment amount and adjusts the cutting power, pressure and rotation speed in real time; step S52, the adjusted disassembly equipment continues to perform battery disassembly operation; step S53, the force sensor and strain gauge collect the new force feedback data after adjustment; step S54, the new force feedback data is used as the input to the multi-objective reinforcement learning model for the next state to form a closed-loop adaptive adjustment.

[0020] A second aspect of the present invention provides a reinforcement learning-based adaptive adjustment system for lithium battery dismantling parameters, which adaptively adjusts lithium battery dismantling parameters using the method described above. The system further includes:

[0021] The acquisition module is used to acquire force feedback data during the battery disassembly process. This force feedback data originates from force sensors and strain gauges. The input module is used to input the force feedback data as state data into a multi-objective reinforcement learning model.

[0022] The calculation module is used to calculate rewards for the multi-objective reinforcement learning model based on disassembly speed, safety, and energy consumption objectives, and output disassembly parameter adjustment amounts; the update module is used to update the strategy of the multi-objective reinforcement learning model according to the rewards to adaptively adjust the disassembly parameters; the control module is used to apply the disassembly parameter adjustment amounts to control the disassembly equipment to perform battery disassembly operations and input the new force feedback data after execution into the multi-objective reinforcement learning model.

[0023] The technical solution provided by this invention has the following beneficial effects:

[0024] This invention discloses a reinforcement learning-based adaptive adjustment method and system for lithium battery disassembly parameters. Addressing the challenges of adapting fixed disassembly parameters to complex mechanical changes due to differences in battery type, aging level, and materials during disassembly, which can easily lead to equipment damage, battery short circuits, or thermal runaway, and the difficulty in balancing disassembly speed, energy consumption, and safety in unique business scenarios, this invention uses real-time acquisition of cutting force and vibration deformation data from force sensors and strain gauges to form a state vector. This vector is input into a multi-objective reinforcement learning model, which calculates a weighted comprehensive reward based on disassembly speed, safety, and energy consumption indicators. The model outputs precise adjustments to cutting power, pressure, and rotation speed, forming a closed-loop adaptive adjustment. Combined with a parameter adjustment rule library constraining the action space and an online experience playback update strategy, the system continuously learns the optimal strategy in a real disassembly environment. Ultimately, this achieves highly adaptable disassembly for different batteries, significantly improving disassembly efficiency and safety while reducing energy consumption and accident risks. Attached Figure Description

[0025] Figure 1 This is a flowchart of the reinforcement learning-based adaptive adjustment method for lithium battery disassembly parameters according to the present invention.

[0026] Figure 2 This is a schematic diagram of the adaptive adjustment system for lithium battery disassembly parameters based on reinforcement learning according to the present invention. Detailed Implementation

[0027] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0028] like Figure 1As shown, the first aspect of this invention provides an adaptive adjustment method for lithium battery disassembly parameters based on reinforcement learning. The method includes: S1, collecting force feedback data during the battery disassembly process, wherein the force feedback data originates from force sensors and strain gauges. During battery disassembly, the disassembly equipment directly contacts components such as the battery casing, tabs, or modules, generating various mechanical feedback information. This information reflects the actual state of the disassembly operation in real time and is an important basis for subsequent parameter adjustment. The force feedback data mainly includes two categories: cutting force data and vibration deformation data. The cutting force data is collected by a force sensor installed on the disassembly execution end, such as a cutting tool or a robotic arm end effector, while the vibration deformation data is collected by strain gauges attached to key disassembly components. Specifically, the force sensor can be a six-dimensional force / torque sensor, capable of simultaneously measuring force and torque in three directions. During disassembly, when the cutting tool contacts the battery casing, the force sensor records the magnitude and direction of the resistance experienced by the tool in real time. For example, when cutting the casing of a power battery, if a region with high material hardness is encountered, the cutting force will suddenly increase, and the sensor will immediately detect this change. Strain gauges are primarily used to monitor minute deformations of disassembled components. For example, strain gauges attached to the joints of a robotic arm or the tool holder can detect structural deformation caused by excessive pressure or vibration, thereby preventing equipment damage or the risk of internal battery short circuits. In one embodiment, the data acquisition process includes:

[0029] Step S11: First, calibrate the sensor to ensure accurate zero-point calibration before disassembly begins.

[0030] Step S12: Subsequently, the disassembly device starts the disassembly operation with preset initial parameters.

[0031] Step S13: Then the force sensor continuously collects triaxial force data during the cutting process at a sampling frequency of not less than 100Hz.

[0032] In step S14, the strain gauges collect strain values ​​at multiple measuring points at the same frequency.

[0033] After analog-to-digital conversion and preliminary filtering, the raw data forms a force feedback data sequence that can be used for subsequent analysis. It should be noted that the acquisition of force feedback data needs to be continuous throughout the entire operation of the dismantling equipment to ensure the real-time and continuous nature of the state information. This acquisition method provides rich and accurate state input for the subsequent multi-objective reinforcement learning model, thereby enabling adaptive adjustment of dismantling parameters. S2, the force feedback data is used as state input to the multi-objective reinforcement learning model. After acquiring the force feedback data, it is directly input into the pre-trained multi-objective reinforcement learning model as the state representation of the current dismantling environment. Based on the force feedback data, the model comprehensively considers three mutually constraining objectives: dismantling speed, safety, and energy consumption. It calculates the comprehensive reward of the current action through its internal policy network and reward mechanism, and outputs specific adjustments to the dismantling parameters accordingly. Specifically, the multi-objective reinforcement learning model adopts an actor-critic architecture framework, where the policy network is responsible for outputting the action, i.e., the dismantling parameter adjustment, and the value network is responsible for evaluating the long-term reward of the state-action pair. In this embodiment, the state space is composed of force feedback data, including the three components of the cutting force, the torque value, and multiple strain readings from the strain gauge. This data, after normalization, forms a high-dimensional state vector. For example, in one possible implementation, when the force feedback data shows a sudden increase in cutting force, the model determines that the current disassembly action may encounter hard materials or be at risk of jamming. In this case, it tends to adjust the cutting power or pressure to prioritize safety. Simultaneously, the model also comprehensively evaluates the current disassembly progress and energy consumption. If the speed is too slow and energy consumption is too high, it may appropriately increase the rotation speed within a safe range to improve efficiency. In another embodiment, the force feedback data undergoes feature enhancement processing before state input. For example, statistical features such as the mean, variance, and rate of change of force within a certain time window are calculated, and these features are combined with the original data to form the state vector. This approach allows the model to better capture the dynamic trends of force feedback, thereby making more accurate parameter adjustment decisions. By using force feedback data as state input, the multi-objective reinforcement learning model can autonomously learn the optimal strategy in a complex and ever-changing disassembly environment, avoiding the instability of traditional fixed-parameter schemes under different battery types or aging levels, thereby significantly improving the overall adaptability of the disassembly process.

[0034] Optionally, this step also includes: Step S21, where the force sensor acquires the cutting force data of the disassembled component in real time, and the strain gauge acquires the vibration deformation data of the disassembled component in real time. In the specific composition of the force feedback data, the acquisition of cutting force data and vibration deformation data are two parallel real-time processes. The force sensor is typically installed at a key location on the cutting actuator, enabling it to accurately sense the resistance generated when the tool contacts the battery material. For example, when disassembling a square battery casing, the force sensor continuously monitors the normal and tangential forces acting on the tool along the cutting path. As the tool cuts into the aluminum casing, the normal force gradually increases until it breaks through the casing surface. The strain gauge is mainly placed in areas prone to deformation, such as the arm or tool holder connection of the robotic arm. When the disassembly operation generates significant vibration or impact, the strain gauge detects micro-strain changes within the material. For example, when separating the adhesive structure connecting the battery module to the casing, excessive pressure may cause the robotic arm to bend slightly. The strain gauge can promptly capture this deformation signal, thus providing a basis for subsequent safety assessments. In one embodiment, the acquisition frequencies of both the force sensor and the strain gauge are set to above 200Hz to ensure that rapid mechanical changes that may occur during disassembly are captured. Simultaneously, the system performs synchronous processing on the acquired data to ensure that the cutting force and vibration deformation data at the same moment are strictly aligned on the time axis, facilitating subsequent state vector construction.

[0035] In another implementation, when the disassembly enters the battery internal module separation stage, the force sensor additionally collects axial pull-out force data, and four additional strain gauges are placed on the module tray to specifically monitor the deformation during the overall lifting of the module. This additional data is also incorporated into the state vector, enabling the model to output differentiated strategies for different disassembly stages. For example, during the pull-out stage, the model significantly reduces the rotational speed while increasing the pressure to ensure the adhesive layer peels off slowly without damaging the cell tabs. Conversely, during the casing cutting stage, high rotational speed and low pressure are prioritized. The seamless switching between these two stages relies on the precise differentiation provided by the added dimension in the state vector. Step S22: The cutting force data and the vibration deformation data are combined to form the current disassembly state vector. The collected cutting force data and vibration deformation data need to be combined to form a unified current disassembly state vector, which serves as the input to the multi-objective reinforcement learning model. Specifically, the two types of data are first preprocessed, including denoising, normalization, and dimensionality reduction, and then concatenated into a high-dimensional vector according to a predetermined order. For example, in one implementation, the cutting force data contains six dimensions: three forces and three moments, and the vibration deformation data contains strain values ​​at eight measuring points, totaling fourteen original dimensions. These dimensions are standardized and then directly concatenated to form a state vector. Furthermore, time-series features, such as the force change trend over several steps, can be added to enhance the expressive power of the state vector. It should be noted that the design of the state vector's dimensions needs to strike a balance between expressive power and computational efficiency. In practical applications, the vector dimension is usually controlled to within 50 to ensure the model's real-time response performance. Through this combination method, the state vector can comprehensively reflect the mechanical characteristics of the current disassembly operation, providing sufficient information support for the model's decision-making. Step S23: The current disassembly state vector is input into the state space of the multi-objective reinforcement learning model for processing by the policy network. The combined current disassembly state vector is directly input into the state space of the multi-objective reinforcement learning model for forward propagation computation by the policy network. The policy network typically employs a multi-layer fully connected network or convolutional network structure, capable of extracting key features from high-dimensional states and outputting parameter adjustments in the continuous action space. In one embodiment, the policy network first performs a nonlinear transformation on the state vector through several hidden layers, and then outputs adjustment values ​​for three parameters: cutting power, pressure, and rotation speed. These adjustment values ​​are represented in increments, such as positive values ​​indicating an increase and negative values ​​indicating a decrease, thereby achieving fine-grained control of the dismantling equipment. For example, when the state vector shows that the cutting force is consistently high while vibration deformation is small, the policy network may output an adjustment combination of increasing the rotation speed and decreasing the pressure to improve the dismantling speed while maintaining safety. Through this processing mechanism, the model can autonomously explore the optimal parameter adjustment strategy in complex environments.

[0036] S3, the multi-objective reinforcement learning model calculates rewards for dismantling speed, safety and energy consumption objectives and outputs dismantling parameter adjustment amounts.

[0037] The described multi-objective reinforcement learning model specifically considers the important objectives of disassembly speed, safety, and energy consumption. Based on these objectives, it derives corresponding reward values ​​through a series of scientifically sound calculations. Furthermore, after calculating the reward, the model further outputs adjustments to the disassembly parameters based on these calculation results, thereby achieving more precise and effective disassembly operations. Optionally, this step also includes: Step S31, obtaining the disassembly speed index, safety index, and energy consumption index corresponding to the current disassembly action. After receiving the state, the multi-objective reinforcement learning model first needs to evaluate the performance of the current disassembly action across the three objective dimensions. Specifically, the disassembly speed index can be measured by the cutting length or the number of separated parts completed per unit time; the safety index is calculated based on the deviation of force feedback data from a preset safety threshold; and the energy consumption index mainly measures the power consumption of the drive motor and cutting mechanism under the current action. For example, in one implementation, the speed index is represented by the cutting progress in millimeters per second, the safety index is represented by a normalized force over-limit risk value, and the energy consumption index is represented by the cumulative consumption in watt-hours. The acquisition of these indicators relies on a real-time monitoring system for the dismantling equipment, enabling synchronous calculations within each decision-making cycle. In another embodiment, a moving average of historical data is also introduced as a reference for indicator calculation to smooth the impact of short-term fluctuations on reward evaluation. The comprehensive acquisition of multi-dimensional indicators provides an objective basis for subsequent reward calculations.

[0038] In another implementation, at the end of each decision cycle, the system extracts specific performance data of the completed action from the equipment log. The speed index is recorded as the actual number of millimeters cut from the start to the end of the action divided by the time taken. The safety index counts the number of times the force value exceeded the safety threshold and the maximum extent of exceeding the limit during the action. The energy consumption index is the sum of the real-time power integrals of the motor and air pump. For example, when cutting a battery base plate, if it takes 3.2 seconds to complete a 120 mm path, the speed is 37.5 mm / s. If the maximum cutting force is 385 Newtons, which does not exceed the 400 Newton threshold, the safety score is 0.96. If the average motor power is 210 watts, the energy consumption is 0.187 watt-hours. These values ​​are directly used for reward calculation.

[0039] Step S32: The disassembly speed indicator, safety indicator, and energy consumption indicator are weighted and summed according to preset weights to obtain a comprehensive reward value. After obtaining the three indicators, they are weighted and summed according to preset weight coefficients to form a single comprehensive reward value, which guides the model's learning direction. The weight coefficients reflect the degree of importance given to different objectives; for example, in scenarios with extremely high safety requirements, the weight of the safety indicator will be significantly higher than the other two. Specifically, the comprehensive reward value can be represented as the sum of the positive contribution of the speed indicator, the negative contribution of the energy consumption indicator, and the strong constraint contribution of the safety indicator. When the safety indicator exceeds the safety threshold, a larger negative reward is applied to force the model to quickly adjust its strategy to avoid risks. In one embodiment, the weight coefficients can be dynamically adjusted according to the actual disassembly task; for example, the safety weight is increased when processing aging batteries, and the speed weight is increased when disassembling in batches. This flexible weighting mechanism allows the model to adapt to different application needs. In one embodiment, the three indicators are multiplied by their corresponding weights and then summed. In the current task, the safety weight is set to 0.55, the speed weight to 0.3, and the energy consumption weight to 0.15. When the safety indicator is below 0.9, it is multiplied by a penalty coefficient of 5 to ensure that the model always prioritizes personal and equipment safety. In another implementation, when the safety indicator is below 0.95 three times consecutively, the weight is automatically increased to 0.7, forcing the model to prioritize learning conservative strategies. Once safety stabilizes, the weights are gradually relaxed to restore normal balance. Step S33: The comprehensive reward value is input into the reward function of the multi-objective reinforcement learning model. The calculated comprehensive reward value is directly input into the model's reward function as an immediate reward signal for the current state-action pair. The design of the reward function determines the model's learning direction and convergence characteristics. In this embodiment, a scalarized comprehensive reward form is used to ensure compatibility with mainstream reinforcement learning algorithms. For example, in one implementation, the reward function also includes an additional constraint term for target balance. When the variance of the three indicators is too large, a penalty is applied, thereby encouraging the model to find a balanced strategy among multiple objectives. Through this reward design, the model can gradually optimize its overall decomposition performance over long-term interactions.

[0040] Step S34: The multi-objective reinforcement learning model outputs the disassembly parameter adjustment amount for cutting power, pressure and rotation speed according to the reward function.

[0041] Throughout the process, the multi-objective reinforcement learning model performs a series of calculations and analyses based on its own operating mechanism and the defined reward function. After a complex and precise internal processing, the multi-objective reinforcement learning model ultimately outputs adjustment values ​​for key parameters such as cutting power, pressure, and rotation speed. These adjustments provide important guidance for subsequent operations.

[0042] Optionally, this step also includes:

[0043] Step S341: Obtain the initial disassembly parameter range from the parameter adjustment rule library. Before the model outputs the adjustment amount, first query the rule library to obtain the upper and lower limits of cutting power, pressure, and rotation speed corresponding to the current battery type. For example, for square aluminum-cased batteries, the rule library provides a power range of 800 watts to 1600 watts, a pressure range of 200 Newtons to 500 Newtons, and a rotation speed range of 8000 rpm to 14000 rpm.

[0044] Optionally, this step also includes:

[0045] Step S3411: The parameter adjustment rule base stores the cutting power threshold, pressure threshold, and rotation speed threshold corresponding to various battery disassembly scenarios. The rule base is stored in tabular form and currently includes four mainstream types: ternary soft pack, lithium iron phosphate square, cylindrical 18650, and blade battery. Each type corresponds to five aging levels, with different threshold ranges for batteries ranging from brand new to those aged for 800 cycles. Step S3412: Match the corresponding scenario in the parameter adjustment rule base according to the current battery type. Before disassembly begins, the operator scans the battery's QR code or manually enters the model number. The system automatically matches the corresponding entry in the rule base. For example, entering "BYD blade battery 300 cycles" directly retrieves the thresholds for that scenario: power 900-1500 watts, pressure 250-450 Newtons, and rotation speed 9000-13000 rpm. Step S3413: Extract the cutting power threshold, pressure threshold, and rotation speed threshold under the matched scenario as the initial disassembly parameter range. After extraction, the threshold is immediately written into the model's action space constraint layer to ensure that all output adjustments do not exceed these hard limits, avoiding hardware overload or safety accidents. In step S3414, the parameter adjustment rule base is dynamically updated using new experience data during online learning. When 20 consecutive experiences show that safety remains excellent at the current threshold and speed can be further improved, the system automatically widens the upper limit by 5% and tightens the lower limit by 3%, for example, updating the power upper limit from 1500 watts to 1575 watts. This allows the rule base to continuously evolve with real disassembly data, adapting to actual on-site conditions. In step S342, the initial disassembly parameter range is combined with the reward function to determine candidate adjustment strategies. The model generates 20 candidate adjustment combinations within the rule base range. Each combination is substituted into the reward function to estimate the possible reward score, and the top 5 highest-scoring candidate strategies are selected for the next step. For example, when the current state indicates excessive vibration, the model will eliminate all candidates with pressure greater than 450 Newtons, retaining only 10 strategies with pressure below 300 Newtons, prioritizing safety. Step S343: Select the highest reward disassembly parameter adjustment amount from the candidate adjustment strategies through the policy network. The policy network scores the remaining candidates one by one and selects the one with the highest expected reward as the final output. For example, the final selection is the combination of power +10%, pressure -8%, and speed +12%, because this combination has the highest estimated reward in the current state and is completely within the rule base range.

[0046] Step S344: Output the disassembly parameter adjustment amount to the disassembly equipment controller.

[0047] S4. Update the policy of the multi-objective reinforcement learning model according to the reward to adaptively adjust the decomposition parameters. At the end of each decision cycle, after the model receives the comprehensive reward value, it immediately initiates the policy update process, using the current experience to adjust the parameters of the policy network and value network online, so that the adjustment amount output in the next cycle is closer to the multi-objective optimal solution. Specifically, the update process first packages the current state, the action taken, the reward obtained, and the new state after executing the action into a complete experience, and then stores it in the experience replay buffer. The buffer always maintains the most recent 5000 experiences; when the buffer is full, the oldest data is replaced using a first-in, first-out (FIFO) approach.

[0048] Optionally, this step also includes: Step S41, storing the current state, action, reward, and new state into the experience replay buffer. Each time a disassembly action is completed, the system automatically generates an experience record, including the state vector before adjustment, the output parameter adjustment amount, the calculated comprehensive reward, and the new state vector after adjustment. This four-tuple is completely stored in the buffer, providing material for subsequent batch learning. For example, when disassembling a ternary soft-pack battery, if an adjustment causes a slight exceedance of the safety index, resulting in a negative reward, this negative experience is stored in the buffer. When the model encounters a similar state next time, it will tend to choose a more conservative adjustment strategy, thus quickly learning to avoid dangerous actions. Step S42, sampling batch experience data from the experience replay buffer. Before the update begins, the controller randomly selects 64 experiences from the buffer to form a small batch. These experiences may come from different batteries, different cutting positions, or even different disassembly stages. Random sampling can break temporal correlation and improve learning stability. In one embodiment, experiences with larger absolute rewards are preferentially sampled; for example, difficult samples with rewards below -2 or above +3 are sampled with double the probability, allowing the model to converge to high-reward policies more quickly. Step S43: The batch of experience data is used to update the gradients of the policy network and the value network. The sampled batch of experiences is fed into the model. The value network first calculates the target Q-value for each experience, while the policy network calculates the action probability distribution based on the current policy. Both networks jointly calculate the loss function and update all network parameters through backpropagation. The learning rate is typically set to 0.0003, and the parameters take effect immediately after each update for the next decision. It should be noted that the update uses a dual-network mechanism: a main network calculates the current Q-value, and a target network calculates the target Q-value. The target network is soft-updated only every 200 updates, effectively reducing oscillations during the learning process. Step S44: The updated policy network outputs adaptive decomposition parameters for the next decomposition cycle. After the parameters are updated, the policy network immediately recalculates the actions using the previous state as input, outputting new disassembly parameter adjustments to the controller. The controller receives the adjustments and executes them immediately, initiating a new round of disassembly. In this way, the model continuously learns online in a real physical environment. For example, during the continuous disassembly of 20 batteries of the same model, the model was still exploring extensively with large fluctuations in adjustments for the first 5 batteries. From the 6th to the 15th battery, the model gradually stabilized, and after the 16th battery, the adjustment range changed by less than 3%. The disassembly speed increased by 27%, and the safety incident rate was zero, indicating that online updates have enabled the model to fully grasp the adaptive patterns of this battery type.

[0049] S5, the disassembly parameter adjustment amount is applied to control the disassembly equipment to perform battery disassembly operations and inputs the new force feedback data after execution into the multi-objective reinforcement learning model. The disassembly parameter adjustment amount is immediately sent to the disassembly equipment controller after being output from the model. The controller changes the three execution parameters—cutting power, pressure, and rotation speed—in real time according to the adjustment amount. The equipment immediately continues to cut the battery casing or separate module connectors with the new parameters. Simultaneously, the force sensor and strain gauge continue to collect the new force feedback data generated after this adjustment. This new data is directly used as the state for the next decision cycle and sent back to the model, forming a complete closed-loop adaptive process. In one embodiment, when the model outputs an adjustment combination that increases cutting power by 15%, decreases pressure by 10%, and increases rotation speed by 20%, the controller first sends a power command to the servo motor driver, a pressure command to the pneumatic or hydraulic actuator, and a rotation speed command to the spindle frequency converter. After all three take effect simultaneously, the cutter immediately accelerates to cut into the next section of the aluminum alloy casing. The new cutting force data rises instantaneously, but the vibration deformation remains stable. These changes are captured by the sensors and packaged into a new state vector, which is then sent back to the model. The model will output a more refined fine-tuning amount based on this new state in the next second. It should be noted that the response time of the entire application process is controlled within 50 milliseconds to ensure that there is almost no delay between the adjustment command and the actual execution. This allows for a rapid response when encountering sudden hard spots or thickening of the adhesive layer, preventing tool jamming or battery fire.

[0050] Optionally, this step also includes: Step S51, where the disassembly equipment controller receives the disassembly parameter adjustment amount and adjusts the cutting power, pressure, and speed in real time. After receiving the adjustment amount, the controller first parses out three independent incremental values ​​of power, pressure, and speed, and then maps them to the corresponding hardware control registers. For example, the cutting power adjustment amount is converted into a change in PWM duty cycle, the pressure adjustment amount is converted into a proportional valve opening command, and the speed adjustment amount is converted into a frequency converter setting value. The three commands are simultaneously sent to the actuator. For example, when disassembling a severely aged lithium iron phosphate prismatic battery, the model finds that the current cutting force is too high and the vibration has begun to increase. A combination of reducing the output power by 8%, reducing the pressure by 12%, and increasing the speed by 5% is immediately executed by the controller. The cutting sound of the blade changes from sharp to smooth, and the force curve becomes significantly flatter, indicating that the adjustment has successfully suppressed the potential risk. In another implementation, the controller also adds a limiting protection. When the adjustment amount exceeds the safe range, it is automatically cut off to the rated threshold. For example, the power will never exceed 120% of the equipment's rated value. This way, even if the model outputs extreme values ​​in the initial exploration phase, it will not cause hardware damage. Step S52: The adjusted disassembly equipment continues the battery disassembly operation. After adjustment, the disassembly equipment does not interrupt the operation but seamlessly continues cutting along the predetermined path, for example, from the top cover of the battery to the side reinforcing rib, and then to the position of the module fixing bolts. The tool path remains continuous throughout the process. The adjustment only changes the execution force and speed without changing the cutting trajectory, ensuring the integrity of the disassembly. Specifically, when separating the battery module from the adhesive layer of the base plate, the reduced pressure after adjustment allows the tool to advance slowly without tearing the adhesive layer, the increased speed makes the cut smoother, and the cutting power is moderately reduced to avoid overheating and thermal runaway of the internal cells. The operator can clearly see the force curve change from violent fluctuations to a steady rise on the monitoring screen, indicating that the current parameter combination is in a safe and efficient range. Step S53: The force sensor and strain gauge collect the new force feedback data after adjustment. After the adjustment takes effect, the sensor immediately begins to collect a new round of data. The force sensor records the reaction force experienced by the tool when cutting into the material under the adjusted parameters, and the strain gauge records the new deformation of the arm and tool holder due to the new pressure and speed. These data usually show smaller peaks and more uniform distribution compared to before adjustment. For example, when cutting the heat sink on the side of the battery, the peak force reached 420 Newtons before adjustment and was accompanied by obvious vibration. After adjustment, the peak force dropped to 280 Newtons, and the vibration deformation decreased by 60%. The newly collected data sequence showed a clear improvement trend, directly providing a positive learning signal for the model. Preferably, the acquisition window of the new force feedback data is strictly synchronized with the decision cycle. After each adjustment is completed, data within 100 milliseconds is immediately collected as a complete set of new states to avoid data misalignment leading to model misjudgment. In step S54, the new force feedback data is used as the input of the next state into the multi-objective reinforcement learning model to form a closed-loop adaptive adjustment.The newly acquired force feedback data, after undergoing the same preprocessing as in step S22, forms a new state vector, which is directly sent back to the model's state input. The model immediately starts a new round of forward calculation, outputting the next adjustment amount. This process repeats until the current battery disassembly task is completed. In one possible implementation, when the force feedback data shows that the cutting force is stable within a safe range and the speed reaches the expected level after 10 consecutive adjustments, the model automatically reduces exploration noise and outputs a more conservative fine-tuning amount, allowing the disassembly to enter a stable and efficient stage until the entire battery casing is completely peeled off.

[0051] like Figure 2 As shown, in a second aspect, the present invention provides a reinforcement learning-based adaptive adjustment system for lithium battery dismantling parameters, which adaptively adjusts lithium battery dismantling parameters using the method described above. The system further includes:

[0052] Acquisition module A1 is used to acquire force feedback data during the battery disassembly process. The force feedback data comes from force sensors and strain gauges. Input module A2 is used to input the force feedback data as a state input to a multi-objective reinforcement learning model.

[0053] The calculation module A3 is used to calculate rewards for the multi-objective reinforcement learning model for disassembly speed, safety and energy consumption objectives and output disassembly parameter adjustment amounts; the update module A4 is used to update the strategy of the multi-objective reinforcement learning model according to the rewards to adaptively adjust the disassembly parameters; the control module A5 is used to control the disassembly equipment to perform battery disassembly operations by applying the disassembly parameter adjustment amounts and input the new force feedback data after execution into the multi-objective reinforcement learning model.

[0054] Specifically, the acquisition module A1's main function is to collect force feedback data during the battery disassembly process. This force feedback data comes from two devices: force sensors and strain gauges. Force sensors accurately detect the magnitude and direction of various forces generated during battery disassembly, while strain gauges measure the strain of an object to reflect the effect of the force, thus providing comprehensive and accurate force feedback data for subsequent analysis.

[0055] Input module A2 takes the collected force feedback data as its state and inputs it into the multi-objective reinforcement learning model. During this process, input module A2 performs necessary format conversions and processing on the force feedback data to ensure that the data can be successfully recognized and utilized by the multi-objective reinforcement learning model, laying the foundation for subsequent calculations.

[0056] The computation module A3 is tasked with calculating rewards for the multi-objective reinforcement learning model based on three key objectives: disassembly speed, safety, and energy consumption. The disassembly speed objective focuses on completing battery disassembly as quickly as possible while maintaining quality; the safety objective emphasizes ensuring no dangerous situations arise during disassembly, guaranteeing the safety of operators and equipment; and the energy consumption objective aims to reduce energy consumption during disassembly, achieving energy conservation and emission reduction. After calculating the rewards, the computation module A3 outputs disassembly parameter adjustment values, which provide crucial guidance for subsequent disassembly operations.

[0057] The update module A4 updates the policy of the multi-objective reinforcement learning model based on the calculated reward. By continuously adjusting the policy according to the reward, the multi-objective reinforcement learning model can adaptively adjust the decomposition parameters, making the decomposition process more optimized. During the policy update process, the update module A4 comprehensively considers various factors to ensure that the model can make reasonable adjustments according to different situations to achieve the best decomposition effect.

[0058] The control module A5 functions by applying calculated disassembly parameter adjustments to control the disassembly equipment to perform battery disassembly operations. During the control process, the control module A5 precisely transmits the disassembly parameter adjustments to the disassembly equipment, ensuring that the equipment operates as required. Furthermore, after the disassembly equipment completes its operation, the control module A5 inputs the newly generated force feedback data back into the multi-objective reinforcement learning model, forming a closed-loop feedback system to continuously optimize the disassembly process and model performance. The above embodiments are merely illustrative of the technical solutions of this application and are not intended to limit it. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A method for adaptive adjustment of lithium battery disassembly parameters based on reinforcement learning, characterized in that, include: S1, Collect force feedback data during the battery disassembly process, the force feedback data comes from force sensors and strain gauges; S2, Input the force feedback data as a state into a multi-objective reinforcement learning model; S3, the multi-objective reinforcement learning model calculates rewards for disassembly speed, safety, and energy consumption targets and outputs disassembly parameter adjustment amounts; S4, the strategy of the multi-objective reinforcement learning model is updated according to the rewards to adaptively adjust the disassembly parameters; S5, the disassembly parameter adjustment amounts are applied to control the disassembly equipment to perform battery disassembly operations and the new force feedback data after execution is input into the multi-objective reinforcement learning model.

2. The method as described in claim 1, characterized in that, The force feedback data includes cutting force data and vibration deformation data.

3. The method as described in claim 2, characterized in that, Step S1 includes: Step S11: Calibrate the sensor; Step S12: The disassembly operation is started with preset initial parameters. Step S13: The force sensor continuously collects triaxial force data during the cutting process at a sampling frequency of not less than 100Hz; In step S14, strain gauges collect strain values ​​at multiple measuring points at the same frequency.

4. The method as described in claim 2, characterized in that, Step S2 includes: Step S21, the force sensor acquires the cutting force data of the disassembled component in real time, and the strain gauge acquires the vibration deformation data of the disassembled component in real time; Step S22, the cutting force data and the vibration deformation data are combined to form the current disassembly state vector; Step S23, the current disassembly state vector is input into the state space of the multi-objective reinforcement learning model.

5. The method as described in claim 1, characterized in that, Step S3 includes: Step S31, obtaining the disassembly speed index, safety index, and energy consumption index corresponding to the current disassembly action; Step S32, weighting and summing the disassembly speed index, safety index, and energy consumption index according to preset weights to obtain a comprehensive reward value; Step S33, inputting the comprehensive reward value into the reward function of the multi-objective reinforcement learning model; Step S34, the multi-objective reinforcement learning model outputting the disassembly parameter adjustment amount for cutting power, pressure, and rotation speed according to the reward function.

6. The method as described in claim 5, characterized in that, Step S34 includes: Step S341, obtaining the initial disassembly parameter range from the parameter adjustment rule base; Step S342, combining the initial disassembly parameter range with the reward function to determine candidate adjustment strategies; Step S343, selecting the disassembly parameter adjustment amount with the highest reward among the candidate adjustment strategies through the strategy network; Step S344, outputting the disassembly parameter adjustment amount to the disassembly equipment controller.

7. The method as described in claim 6, characterized in that, Step S341 includes: Step S3411, storing cutting power thresholds, pressure thresholds, and rotation speed thresholds corresponding to various battery disassembly scenarios in the parameter adjustment rule library; Step S3412, matching the corresponding scenario in the parameter adjustment rule library according to the current battery type; Step S3413, extracting the cutting power threshold, pressure threshold, and rotation speed threshold under the matched scenario as the initial disassembly parameter range; Step S3414, dynamically updating the parameter adjustment rule library using new experience data during online learning.

8. The method as described in claim 1, characterized in that, Step S4 includes: Step S41, storing the current state, action, reward and new state into the experience replay buffer; Step S42, sampling batch experience data from the experience replay buffer; Step S43, using the batch experience data to perform gradient updates on the policy network and value network; Step S44, the updated policy network outputs adaptive disassembly parameters for the next disassembly cycle.

9. The method as described in claim 1, characterized in that, Step S5 includes: Step S51, the disassembly equipment controller receives the disassembly parameter adjustment amount and adjusts the cutting power, pressure and rotation speed in real time; Step S52, the adjusted disassembly equipment continues to perform battery disassembly operation; Step S53, the force sensor and strain gauge collect the new force feedback data after adjustment; Step S54, the new force feedback data is used as the input of the next state to the multi-objective reinforcement learning model to form a closed-loop adaptive adjustment.

10. A reinforcement learning-based adaptive adjustment system for lithium battery dismantling parameters, wherein the lithium battery dismantling parameters are adaptively adjusted using the method described in any one of claims 1-9, characterized in that, The system also includes: The acquisition module (A1) is used to acquire force feedback data during the battery disassembly process. The force feedback data comes from force sensors and strain gauges. The input module (A2) is used to input the force feedback data as a state input to the multi-objective reinforcement learning model. The calculation module (A3) is used to calculate rewards for the multi-objective reinforcement learning model for disassembly speed, safety and energy consumption objectives and output disassembly parameter adjustment amounts; the update module (A4) is used to update the strategy of the multi-objective reinforcement learning model according to the rewards to adaptively adjust the disassembly parameters; the control module (A5) is used to apply the disassembly parameter adjustment amounts to control the disassembly equipment to perform battery disassembly operations and input the new force feedback data after execution into the multi-objective reinforcement learning model.

Images