Intelligent disassembling and sorting system for recycling lithium batteries and control method

Abstract

The invention relates to the technical field of lithium battery recovery, in particular to an intelligent disassembling and sorting system for lithium battery recovery and a control method. The method comprises the following steps: taking first state estimation as an initial condition to drive forward deduction of the digital twin to obtain a baseline prediction trajectory; comparing the second state estimation with a baseline prediction value generated at the same moment of the baseline prediction trajectory, and calculating a difference; if the difference exceeds a preset threshold value, triggering diagnosis; sequentially deducing intervention hypotheses in the digital twin by taking the second state estimation as a starting point to obtain an intervention prediction trajectory set; and selecting the intervention hypothesis corresponding to one intervention prediction trajectory in the intervention prediction trajectory set to generate a final decision instruction. According to the method, the decision-making basis is changed from shallow signal correlation to deep physical and chemical causal understanding, so that the adaptability to complex working conditions and decision-making accuracy are remarkably improved in the non-linear and strong-coupling process of disassembly and sorting of the waste lithium batteries.

Description

Technical Field

[0001] This invention relates to the field of lithium battery recycling technology, and more specifically, to an intelligent dismantling and sorting system and control method for lithium battery recycling. Background Technology

[0002] With the rapid development of my country's new energy vehicle industry, power batteries have entered a stage of large-scale retirement. Against this backdrop, the recycling of power batteries will become a core requirement for the sustainable development of the industry.

[0003] The essence of power battery recycling lies in dismantling and sorting. Current technologies employ a rigid process of crushing followed by sorting, which, due to a lack of real-time feedback on the internal material distribution, state of charge, and binder characteristics of the battery cells, is essentially a blind process operating from a "black box." However, with the development and mature application of smart technologies, the power battery recycling process, through the infusion of intelligent technologies, gains the ability to perceive and make decisions.

[0004] After in-depth analysis of existing intelligent application scenarios for lithium battery recycling technology, we found that even with a closed-loop architecture of "perception-execution-re-perception-decision," its core effectiveness is still limited by a fundamental problem: the system's processing of feedback information remains at the signal correlation level, rather than the causal understanding level. Existing re-perception stages collect signals such as force, sound, and spectra, and make matching decisions through preset rules or models. This essentially treats complex physicochemical processes as a black box, adjusting inputs through external outputs. When facing nonlinear, strongly coupled, and state-variable crushing and sorting processes, the upper limit of this method's adaptability depends on the completeness of the preset rules, which is almost impossible to achieve in open real-world scenarios. Its creative bottleneck lies in the system's lack of deep understanding of why these signals are generated, thus preventing it from performing principle-based reasoning and autonomous strategy generation when encountering entirely new or complex operating conditions. Summary of the Invention

[0005] The purpose of this invention is to provide an intelligent dismantling and sorting system and control method for lithium battery recycling, in order to solve the problem that after the application of intelligent technology in the dismantling and sorting process of waste lithium batteries, the closed-loop architecture lacks a deep understanding of the cause of signal generation, resulting in low adaptability and decision-making accuracy.

[0006] To achieve the above objectives, a first technical solution is disclosed: a control method for lithium battery recycling, comprising: Obtain the first state estimate; The first state estimate is used as the initial condition to drive the forward extrapolation of the digital twin, and the baseline predicted trajectory is obtained; Obtain the second state estimate; The second state estimate is compared with the baseline prediction value generated at the same time as the baseline prediction trajectory, and the difference is calculated. If the difference exceeds a preset threshold, a diagnosis is triggered and the diagnosis result is output. An intervention hypothesis is generated based on the diagnosis result. If the difference does not exceed the preset threshold, an intervention hypothesis is generated directly. Starting with the second state estimate, the intervention hypothesis is sequentially deduced in the digital twin to obtain the set of intervention prediction trajectories; Select one intervention prediction trajectory from the set of intervention prediction trajectories, generate the final decision instruction based on the corresponding intervention hypothesis, and output it.

[0007] As a further improvement to this technical solution, obtaining the first state estimate includes: Obtain the perception data at the current moment; The current sensing data is reconstructed to obtain the virtual state of the lithium battery at the current moment; The first state estimate is generated based on the virtual state of the lithium battery at the current moment.

[0008] As a further improvement to this technical solution, obtaining the second state estimate includes: Acquire the perception data for the next moment; The sensing data of the next moment is reconstructed to obtain the virtual state of the lithium battery in the next moment; The second state estimate is generated based on the virtual state of the lithium battery at the next time step.

[0009] As a further improvement to this technical solution, the state reconstruction includes: Preprocess the sensed data; The preprocessed sensing data is input into the state reconstruction algorithm; The state reconstruction algorithm calculates a set of state variable values ​​that probabilistically best match all observations and physical constraints. The virtual state of the lithium battery is generated based on this set of state variable values.

[0010] As a further improvement to this technical solution, the state reconstruction algorithm includes: It includes a computable causal graph, where nodes define state variables and edges define physical or chemical constraint equations between state variables.

[0011] As a further improvement to this technical solution, the driving digital twin forward deduction includes: Load the currently set process parameters; Drive all the mechanistic equations in the digital twin and extrapolate them forward by a fixed time window; A baseline prediction trajectory is generated, which is used to describe the baseline prediction values ​​of each state variable at future time points.

[0012] As a further improvement to this technical solution, the trigger diagnosis includes: Lock the state variables that have deviated; Tracing the causal dependency path in reverse within the digital twin, we can pinpoint the root cause node with the greatest deviation between prediction and reality. Diagnostic results are output based on the root cause node.

[0013] As a further improvement to this technical solution, the selection of one intervention prediction trajectory from the set of intervention prediction trajectories includes: The final results of all intervention prediction trajectories in the intervention prediction trajectory set are evaluated; the evaluation criteria include the improvement of core target indicators and whether safety and quality constraints are violated. All feasible intervention prediction trajectories that meet the constraints are screened out, and the intervention prediction trajectory with the best optimization effect on the core objective is selected from them.

[0014] Second technical solution: An intelligent dismantling system for lithium battery recycling, including a processor and a memory, wherein the processor executes a program stored in the memory to implement the aforementioned control method for lithium battery recycling, and further includes: The baseline prediction trajectory generation module is used to receive the first state estimate, use the first state estimate as the initial condition, drive the digital twin to perform forward inference, and generate the baseline prediction trajectory. The comparison and diagnosis module is used to receive the second state estimate, compare the second state estimate with the baseline prediction value generated at the same time as the baseline prediction trajectory, and calculate the difference; if the difference exceeds a preset threshold, the diagnosis is triggered and the diagnosis result is output, and an intervention hypothesis is generated based on the diagnosis result; if the difference does not exceed the preset threshold, the intervention hypothesis is generated directly. The intervention prediction trajectory generation module is used to sequentially deduce intervention hypotheses and generate a set of intervention prediction trajectories. The decision determination module is used to select the intervention hypothesis corresponding to one intervention prediction trajectory in the set of intervention prediction trajectories and generate the final decision instruction; as well as, The disassembly decision instruction output module is used to transmit the final decision instruction to the execution terminal used for disassembly.

[0015] The third technical solution: an intelligent sorting system for lithium battery recycling, including a processor and a memory, wherein the processor executes a program stored in the memory to implement the aforementioned control method for lithium battery recycling, and further includes: The baseline prediction trajectory generation module is used to receive the first state estimate, use the first state estimate as the initial condition, drive the digital twin to perform forward inference, and generate the baseline prediction trajectory. The comparison and diagnosis module is used to receive the second state estimate, compare the second state estimate with the baseline prediction value generated at the same time as the baseline prediction trajectory, and calculate the difference; if the difference exceeds a preset threshold, the diagnosis is triggered and the diagnosis result is output, and an intervention hypothesis is generated based on the diagnosis result; if the difference does not exceed the preset threshold, the intervention hypothesis is generated directly. The intervention prediction trajectory generation module is used to sequentially deduce intervention hypotheses and generate a set of intervention prediction trajectories. The decision determination module is used to select the intervention hypothesis corresponding to one intervention prediction trajectory in the set of intervention prediction trajectories and generate the final decision instruction; as well as, The sorting decision instruction output module is used to transmit the final decision instruction to the execution terminal used for sorting.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This control method for lithium battery recycling uses a digital twin to simulate the internal dynamics of a lithium battery under mechanical and thermal influences in real time. By comparing the measured signals with the virtual signals generated by the twin simulation, the actual internal state of the lithium battery and the main physical mechanisms leading to that state can be inferred. This elevates the decision-making basis from superficial signal correlation to a deeper understanding of physicochemical causality, significantly improving adaptability and decision-making accuracy in the face of nonlinear and strongly coupled processes in the dismantling and sorting of waste lithium batteries.

[0017] 2. In this intelligent dismantling system for lithium battery recycling, the deeply integrated output module can directly and reliably transform such decisions based on causal cognition into executable instructions for the equipment control system. This enables the dismantling process to proactively avoid risks based on real-time understanding when facing battery packs with uncertain internal states, rather than passively responding to alarms afterward. This fundamentally improves the safety and predictive ability of high-risk dismantling operations.

[0018] 3. In this intelligent sorting system for lithium battery recycling, based on cross-process causal understanding (e.g., identifying that the root cause of high moisture content affecting sorting efficiency lies in insufficient upstream drying), structured optimization suggestions or collaborative adjustment requests are output to the upper-level production management system or upstream process controller through a standard interface. This makes the entire sorting system a quality optimization hub with a global perspective, capable of proactively adapting to fluctuations in incoming materials and driving collaboration between upstream and downstream processes through information flow. This stabilizes and improves the recovery rate and purity of the final product at the entire process level, solving the quality fluctuation problem caused by the lack of causal understanding and collaborative capabilities in traditional sorting. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the control method steps of the present invention, which is based on the improvement of the original closed-loop architecture. Figure 2 This is a schematic diagram of the state reconstruction steps of the present invention; Figure 3 This is a schematic diagram of the baseline prediction and consistency verification steps of the present invention; Figure 4 This is a schematic diagram illustrating the diagnostic and optimization deduction steps of the present invention; Figure 5 This is a schematic diagram of the decision generation and output steps of the present invention; Figure 6 This is a schematic diagram of the structure of the computable causal graph model of the present invention; Figure 7 This is a schematic diagram of the intelligent disassembly system of the present invention; Figure 8 This is a schematic diagram of the intelligent sorting system of the present invention. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] Figure 1 The steps of the control method based on the original closed-loop architecture are shown.

[0022] First embodiment, see Figure 1 This embodiment aims to disclose a control method for lithium battery recycling. This method, instead of directly using the multimodal raw signal streams (such as force curves, acoustic spectra, Raman shifts, and thermograms) acquired in the sensing stage of existing technologies for rule matching, inputs them into a digital twin constructed based on the fusion of first principles of physicochemistry and data-driven approaches. This allows the control method for lithium battery recycling to retain the original dynamic closed-loop architecture while also incorporating understanding and reasoning capabilities.

[0023] Specifically, a digital twin simulates in real time the dynamic processes inside the object being processed (such as a battery fragment or a localized area of ​​action) under mechanical and thermal influences, including stress propagation, crack propagation, phase transformation, interface debonding, and temperature rise conduction. By comparing and assimilating the measured signals with the virtual signals generated by the digital twin in real time, the true internal state of the material (such as the location of crack tips, the degree of binder softening, and the fatigue degree of the metal foil) and the main physical mechanisms leading to this state (such as whether brittle fracture or plastic shear is dominant) can be inferred.

[0024] See Figure 1Control methods for lithium battery recycling, including: S1.1 Obtain the current sensing data to get the first sensing data; Based on the first perception data, the state is reconstructed to generate a first state estimate; S1.2 Obtain the first state estimate. Using the first state estimate as the initial condition, drive the forward extrapolation of the computable causal graph (CCG) to generate the baseline predicted trajectory. It's important to note that the computable causal graph (CCG) plays a crucial role in the control method for lithium battery recycling. It explicitly encodes causal relationships within the domain graphically and possesses computability, thereby driving a closed loop from state awareness to autonomous decision-making. In this embodiment, there are several alternatives to the computable causal graph (CCG): When using the first-state estimate as the initial condition for dynamic extrapolation, the alternative must be transformed from a tool that is more static and more correlation-oriented to one that can digest and process a definite, multi-dimensional state variable that may contain uncertainty, and use this as a basis to extrapolate the future state evolution.

[0025] When Bayesian networks are used as an alternative, their adjustment direction is towards temporal and probabilistic state evolution. A typical static Bayesian network excels at conditional probability inference at a single point in time, but it lacks an inherent temporal dimension. To be capable of temporal extrapolation, it must be extended to a dynamic Bayesian network or a temporal network. In this case, each state variable and its uncertainty in the first-state estimate (e.g., the bearing temperature is 85°C with a variance of 2°C) are transformed into evidence inputs or prior probability distributions for the corresponding nodes in the initial time slice. The causal and probabilistic relationships of the system are then encoded in the transition network between time slices. The extrapolation process thus becomes forward probability propagation along the time axis based on initial evidence. This essentially endows the causal probabilistic model with state-driven dynamic sequence computation capabilities.

[0026] For physics-based mechanistic models, adjustments seem most straightforward, as differential equations inherently require initial values. However, the key adjustments lie in state alignment and uncertainty quantification. The state variables of mechanistic models are often based on idealized physical definitions, while the estimated state output by the upstream state estimator is an engineering approximation based on a fusion of finite and noisy observations. Both must be aligned in both physical and mathematical terms, often requiring model order reduction or state transformation. More importantly, if the initial state estimate itself is a probability distribution (e.g., including confidence intervals), simply substituting a single initial value will lead to a significant underestimation of future uncertainty. In this case, the model must be combined with uncertainty propagation methods (such as stochastic differential equations or Monte Carlo simulations) to achieve a complete chain from probabilistic initial conditions to probabilistic predictions.

[0027] Purely data-driven machine learning models, especially deep learning models, require a fundamental paradigm shift when facing this requirement. Many data-driven models are essentially complex input-output mapping functions, and their training data are usually independent sample points. The model does not explicitly maintain or understand the concept of system state internally. To enable such models to accept the first state estimate as an initial condition and perform multi-step derivations, there are generally two adjustment paths: one is to change the model structure, adopting architectures with temporal memory capabilities such as recurrent neural networks, long short-term memory networks, or state-space neural networks, and explicitly using historical state sequences as input and future states as outputs during training, so that it learns the dynamic evolution of states; the other is to change the way the model is used from direct mapping to iterative feedback, that is, using the state estimate as the first input to obtain the prediction of the next time step, and then using this prediction as part of the input, iterating cyclically.

[0028] In this preferred embodiment, a computable cause-effect graph (CCG) is used as the digital twin, which is based on a real-time, high-fidelity data stream. Whether it's vibration sensor data from the disassembly system or XRF spectral data from the sorting system, it is first aggregated into the digital twin to form its perception of the physical behavior of the lithium battery. This physical behavior refers to a series of dynamic processes and responses exhibited by the lithium battery throughout the entire recycling process, determined by its internal physical and chemical nature. Specifically, it encompasses at least the following interwoven levels: First, electrochemical behavior, namely the dynamic processes of ion migration, phase transition, side reactions, and gas production (such as hydrogen and hydrogen fluoride) of active materials inside the battery under conditions such as charging and discharging, breakage, and short circuits; second, thermal behavior, which is the energy manifestation of electrochemical behavior, including the conduction and evolution paths of heat generation, heat transfer, heat accumulation, and even potential thermal runaway under external field effects such as disassembly, breakage, and sorting; third, mechanical behavior, referring to the mechanical processes of deformation, rupture, and the initiation and propagation of internal short circuits in battery components (such as casings, electrode sheets, and separators) under mechanical stresses such as cutting, extrusion, and vibration; and fourth, apparent physical property behavior, such as the separation and mixing dynamics exhibited by different materials after breakage due to differences in density, magnetism, conductivity, and surface optical properties (responses under XRF, near-infrared, and other spectra). Then, the core of the digital twin is the set of models it contains that can describe and predict the behavior of physical entities. This model layer is not a single entity, but a fusion of multiple models, scales, and fidelity levels. It can be any one or more of the following: computable causal graphs, Bayesian networks, deep reinforcement learning, and data-driven approaches.

[0029] S1.3 Obtain the perception data at the next moment to obtain the second perception data; Based on the second sensing data, the state is reconstructed to generate a second state estimate; S1.4 Obtain the second state estimate, compare the second state estimate with the baseline prediction value generated at the same time as the baseline prediction trajectory, and calculate the difference; If the difference exceeds a preset threshold, a diagnosis is triggered and a diagnosis result is output. Multiple intervention hypotheses are generated based on the diagnosis result. If the difference does not exceed the preset threshold, multiple intervention hypotheses seeking proactive optimization are generated directly. S1.5 Starting with the second state estimate, the intervention hypothesis is sequentially deduced in the computable causal graph (CCG) to obtain the intervention prediction trajectory set in a forward deduction. The intervention prediction trajectory set has multiple intervention prediction trajectories corresponding to multiple intervention hypotheses. S1.6. Evaluate the final results of all intervention prediction trajectories in a set of intervention prediction trajectories to determine the optimal intervention hypothesis; Generate final decision instructions based on the optimal intervention assumption; S1.7. Reacquire the current sensing data and repeat steps S1.1-S1.7.

[0030] To facilitate understanding of the technical solution, the technical terms appearing in the above methods are explained as follows: State estimation refers to the process of reconstructing the state, specifically by fusing real-time sensor data with a mechanistic model to calculate the most likely virtual state of the lithium battery at the current moment. The sensors include visual and spectral sensors, force sensors, acoustic sensors, vibration sensors, and anemometers and wind pressure sensors (installed in the air ducts of the air classifier to monitor and provide real-time feedback on key airflow velocities and pressures for material separation, ensuring stable separation conditions). The mechanistic model includes a crushing particle size prediction model (based on crushing dynamics and population balance models, using current tool wear, material hardness, crusher speed, and feed rate as inputs). By solving a set of differential equations describing the repeated impact and shearing of materials until a certain particle size distribution is achieved, the expected particle size distribution of the crushed products is theoretically predicted. A gas-solid two-phase flow separation model (based on the coupling principle of computational fluid dynamics and discrete element method. This model uses the geometry of the air classifier, the real-time wind speed / pressure measured by sensors, and the particle size and density distribution of the material predicted by the crushing model as boundary conditions. Through numerical simulation, it calculates the trajectory and final landing point of particles of different densities and sizes, such as black powder, copper foil scraps, and aluminum foil scraps, in the airflow, thereby theoretically predicting the copper-aluminum separation efficiency and black powder recovery rate). In summary, sensors are responsible for capturing the massive amounts of real-time, variable physicochemical signals on the production line, providing direct or indirect observation data; the mechanism model mathematically encodes the inherent laws of lithium batteries during physical crushing, separation, and chemical transformation, providing the physicochemical laws of state evolution.

[0031] An intervention hypothesis refers to a virtual operating plan or parameter adjustment suggestion that is generated and can be tested and evaluated based on the current state estimate and understanding of the mechanism in order to achieve a specific goal (such as correcting deviations or proactive optimization).

[0032] The baseline prediction trajectory refers to the future state evolution trajectory obtained by forward extrapolation through a computable causal graph, based on the first state estimate as the initial condition and the current unchanged process parameters. It can be understood as what the lithium battery will be like in the future if the status quo is maintained.

[0033] Intervention prediction trajectory refers to the future state evolution trajectory obtained by simulating a change in a certain process parameter according to the intervention assumption in counterfactual inference, and deducing it through a computable causal graph. It can be understood as what will happen to lithium batteries in the future if a certain change is made.

[0034] The following is a detailed step-by-step plan: The first step is state reconstruction (obtaining the current facts), see [link to documentation]. Figure 2 : S2.1 Real-time acquisition of multi-source sensor data from each workstation on the dismantling line (i.e., sensing data, such as cell images from a vision camera, scanning of broken particle size by a laser sensor, capture of internal short-circuit explosions by an acoustic emission sensor, monitoring of local hot spots by a thermal imager, and measurement of residual charge by a current sensor).

[0035] S2.2. After preprocessing, the multi-source sensor data is input into the state reconstruction algorithm. This algorithm incorporates a computable causal graph (CCG), where nodes define state variables (such as "cell casing stress," "cobalt-nickel concentration in black powder," "crusher tool wear," and "electrolyte vapor concentration"), and edges define the physical / chemical constraint equations between them. The algorithm (e.g., sequence estimation based on Bayesian filtering) combines real-time observation data with the CCG model to calculate the set of state variable values ​​that probabilistically best matches all observations and physical constraints.

[0036] S2.3 Output a unified, high-confidence virtual state of the lithium battery at the current moment, i.e., the first state estimate.

[0037] Step 2: Baseline Prediction and Consistency Validation (Foreseeing the future and testing cognition) See Figure 3 : S3.1 Starting from the first state estimate, load the currently set process parameters, drive all the mechanism equations in CCG, and extrapolate forward (forward) a fixed time window (e.g., the next 30 seconds).

[0038] S3.2 Generate a baseline prediction trajectory, which details the baseline prediction values ​​of each state variable at future time points from the current moment. For example: "It is predicted that the purity of black powder will rise to 98.2% in 15 seconds; the temperature of the crusher bearing will rise to 72℃ in 25 seconds; and the copper-aluminum separation efficiency will reach 96.5% in 30 seconds."

[0039] S3.3 During actual production line operation, state reconstruction is continuously performed, periodically generating new state estimates. The newly obtained state estimate (representing the "new reality") is compared with the baseline prediction value generated at the same time previously. For example, when a state estimate of "current time + 10 seconds" is obtained, the prediction value for "current time + 10 seconds" on the baseline prediction trajectory is searched, and the difference is calculated.

[0040] Step 3: Diagnosis and Optimization Deduction (Analyzing Causes and Exploring Possible Solutions), see [link / reference] Figure 4 : S4.1 When consistency verification reveals a discrepancy exceeding a preset threshold (e.g., actual black powder purity is 97.0%, while the baseline prediction is 98.2%), a diagnosis is triggered. The state variable exhibiting the deviation (e.g., "black powder purity") is identified, and its causal dependency path is traced backward in the CCG to pinpoint the root cause node with the largest deviation between prediction and reality. For example, the tracing reveals that the actual value of "fragmentation particle size distribution" is more uneven than the predicted value, and the actual value of "tool wear status" (based on vibration signal reconstruction), which affects this node, is 15%, far exceeding the 12% model value used in the baseline prediction. The diagnostic conclusion is: "Accelerated tool wear is the root cause of the decrease in black powder purity."

[0041] S4.2 This step is performed periodically, regardless of whether a problem is diagnosed. Based on the optimization objective (e.g., "maximum cobalt recovery"), multiple process parameter adjustment assumptions are generated. For each assumption (e.g., "increase the wind classifier velocity to 10 m / s" or "increase the pre-discharge current by 20%), a copy of the current state estimate is used as a starting point, the parameter change is simulated in CCG, and then the forward inference is rerun.

[0042] S4.3 Obtain a set of intervention prediction trajectories. Each trajectory shows the evolution path of various state variables and key indicators under this specific intervention. For example, intervention hypothesis A shows that "at a wind speed of 10 m / s, the cobalt recovery rate is predicted to be 99.1% after 30 seconds, but the aluminum chip entrainment rate will rise to 3%"; intervention hypothesis B shows that "with an increase in discharge current, the cobalt recovery rate is predicted to be 98.8% after 30 seconds, but the overall energy consumption will increase by 15%".

[0043] Step 4: Decision Generation and Output (Selecting the Best Action): S5.1 Evaluate the final results of all intervention prediction trajectories. Evaluation criteria include: 1) improvement of core target indicators (such as cobalt recovery rate); 2) whether safety and quality constraints are violated (such as electrolyte vapor concentration must not exceed the lower explosive limit, stress must not exceed the material limit, and the content of key impurities must not exceed the standard).

[0044] S5.2. Filter out all feasible intervention prediction trajectories that meet the constraints, and select the intervention prediction trajectory that has the most significant optimization effect on the core objective. Generate the final decision instruction based on the intervention hypothesis corresponding to the intervention prediction trajectory.

[0045] S5.3 Convert the optimal solution into specific, executable instructions. For example, if the optimal solution is "adjust the wind speed of the air separator to 9.5 m / s and adjust the duct angle to 22 degrees", then directly issue these two precise adjustment instructions to the PLC. If the optimal solution is "immediately stop the machine and replace the No. 2 crusher blades", then issue a high-priority alarm to the operator along with a detailed diagnostic report.

[0046] S5.4 After the command is executed, the production line enters a new operating state. Sensor data streams continue to be input, starting again from the first step: state reconstruction. Based on the new reality, the state estimate is updated, initiating a new cycle of prediction, verification, deduction, and decision-making. The entire process thus forms an autonomous intelligent closed loop of perception, cognition, decision-making, and execution that continuously iterates and optimizes itself.

[0047] For details, see Figure 6 The computable causal graph (CCG) model comprises an ontology layer, a physical layer, an observation layer, and a computational layer, wherein: At the ontology layer, the model first establishes a semantic framework for all relevant entities, defining that "cutting tool" is a "component" of "crusher," that "vibration acceleration sensor" is "installed" on the crusher bearing housing to "monitor" its state, and that the "hardness" of "lithium battery" will "affect" the "wear" of "cutting tool," etc. These static, conceptual relationships provide logical constraints for the entire estimation task, ensuring that variables and equations in subsequent layers have consistent physical meanings and avoiding semantic ambiguity.

[0048] At the physical layer, the model transforms ontological concepts into computable mechanisms. Taking "tool wear" as an example, based on impact wear theory, its evolution is described by differential equations: Formula 1 In Equation 1, and These are the current and previous wear values ​​(unit: micrometers); It is the material wear coefficient; It refers to the hardness of the material; It is the average impact force; It is the contact area of ​​the cutting tool; It is the impact frequency; It refers to the time step. Equation 1 forms a causal boundary, which quantitatively expresses how wear is affected by material properties (material hardness). ) and operating parameters ( and Driven by the accumulation of wear, the physical layer may also contain other algebraic relationships, such as "cumulative processing volume," which is highly correlated with wear. In other words, the physical layer provides a mechanistic prediction of state evolution, i.e., how wear will change according to known physical laws without external observation. It is the core source of the model prediction part of the entire estimation process.

[0049] At the observation layer, the model establishes a probabilistic bridge between the physical state and real-world sensor data. Since wear cannot be directly measured, we utilize its resulting effect—changes in vibration characteristics. Therefore, the observation layer defines the observation variable "vibration energy amplitude" and establishes an observation model relating it to "tool wear." Formula 2 In Equation 1, and These are calibration coefficients, representing the contributions of wear and foundation vibration to the observed energy, respectively. It is Gaussian observation noise with variance of Equation 2 clearly expresses the sensor reading. This is the actual wear and tear condition. A noisy, operating condition-dependent ( The mapping of interference. The role of the observation layer is to provide data evidence, quantifying the degree of consistency between model predictions and real-world measurements.

[0050] At the computational layer, an engine integrates and solves information from the first three layers. Specifically, it compiles the semantic constraints of the ontology layer, the mechanistic equations of the physical layer, and the probabilistic model of the observation layer, along with real-time sensor data, into an executable state estimation problem, and outputs numerical results through algorithms (such as Kalman filtering). The computation process is a recursive loop; the following is a specific time step. The deduction steps are as follows: Input preparation: The algorithm reads the posterior estimate of the wear amount from storage at the previous time step. Micrometers and their uncertainties Equation 1 and its parameters are obtained from the ontology layer and configuration (assuming...). =0.05, =120, , Equation 2 and its parameters are obtained from the observation layer (assuming...). , noise variance Real-time data streams provide current observations. and current working conditions .

[0051] Prediction step (based on physical layer): The algorithm executes the physical layer equations, performs prior predictions, and estimates the prior state. micrometer; Propagation of a priori uncertainty: +Q=1.0+4.0=5.0, where Q is the process noise variance, representing the model error.

[0052] Update step (based on observation layer): The algorithm incorporates actual observations to correct the predictions. Calculate the predicted observations: ; Calculate the innovation residual: ; Calculate the Kalman gain (balancing the confidence level of the model and the observations): ; Update state estimate (first state estimate): micrometer; Updated uncertainty: .

[0053] The final output of the computational layer is not an isolated number, but a structured estimate (i.e., a first-state estimate) with accompanying confidence information. This output must contain complete context for subsequent steps such as causal mechanism analysis. An example of its standard format is as follows: { “estimation_report”: { "timestamp": "2024-10-27T14:30:00Z", “target_variable”: { "name": "tool_wear_amount", "id": "TWA_001", "value": 8.9, "unit": "micrometers" }, "probabilistic_characterization": { "distribution_type": "Gaussian", "mean": 8.9, "variance": 0.833, "confidence_interval": { "lower": 7.1, "upper": 10.7, "confidence_level": 0.95 } }, "fusion_sources": { "physical_model": "discrete_wear_evolution", "observation_model": "vibration_energy_mapping", "key_observations": [ { “sensor”: “vib_001”, “type”: “energy_band”, “value”: 8.6, “unit”: “m² / s³”} ], "key_parameters": [ { “name”: “material_hardness”, “value”: 120, “unit”: “HB”}, { “name”: “impact_frequency”, “value”: 8200, “unit”: “counts / hour”} ] } } } This format explicitly includes: 1) timestamps for time-series alignment; 2) semantic identifiers and numerical values ​​of the target variable; 3) probabilistic representations (mean, variance, confidence interval), which are the core of uncertainty quantification; and 4) fusion sources, clearly listing the underlying physical models, observational models, specific observational data, and key parameters. This structured output provides directly usable input for the next step of causal mechanism analysis. Based on this high-confidence first-state estimate ("the current wear level is 8.9 micrometers, but still within a safe range"), the analysis module can automatically extrapolate future impacts and generate intervention hypotheses such as "it is recommended to schedule maintenance before the wear level reaches 15 micrometers," thus forming a closed loop from perception to decision-making. The entire process demonstrates CCG's powerful ability to transform raw data into actionable knowledge through multi-layered abstraction and computation.

[0054] The first-state estimate is the only reliable initial condition for all subsequent deductions and analyses.

[0055] Furthermore, the first state estimate (tool wear) output by the state reconstruction is used. micrometers, its variance (etc.) serve as the initial conditions for all relevant state variables in the entire CCG. Simultaneously, the currently set, adjustable process parameters (set as vectors) are loaded. This includes things like crusher speed. rpm, air separator wind speed (m / s, etc.). Subsequently, all mechanism equations defined in the CCG physical layer, including state evolution equations and algebraic constraint equations (through numerical integration), are extrapolated forward over a fixed time window. ( (seconds). This forward simulation process, mathematically speaking, is the solution of a system of differential-algebraic equations: for continuous state variables... (such as wear amount) Bearing temperature Black powder purity (etc.), its evolution follows: Formula 3 In Equation 3, It is a vector function that describes the causal mechanism. These are fixed parameters.

[0056] For algebraic variables (such as separation efficiency) ),satisfy: Formula 4.

[0057] By solving, we obtain a path from the current moment. arrive The baseline predicted trajectory, i.e., the set: This trajectory quantifies the expected future behavior under existing cognitive models and current operations, namely: "prediction". In seconds, the purity of black powder It will rise to 98.2%; forecast The temperature of the crusher bearing at 1000 sec. It will rise to 72℃.

[0058] However, the actual operation of the production line is dynamic. While advancing physical time, the first step of state reconstruction is continuously executed in parallel, periodically generating a sequence of state estimates based on the latest sensor data. This represents a new reality (i.e., a second-state estimate). The core of consistency verification is comparing the newly obtained second-state estimate with the baseline prediction previously generated at the same physical time point. When the actual time arrives... At the specified time, obtain the second state estimate for this moment. Simultaneously, the corresponding time point is read from the stored baseline prediction trajectory. Predicted value For each state variable of interest Calculate its deviation: ( )= This deviation is compared with the preset allowable threshold of the variable. Compare them.

[0059] This process essentially uses real-time observation data as the truth to continuously correct and verify the model-based forward simulation, which is a key step in determining whether the cognitive model can still accurately reflect physical reality.

[0060] Once consistency verification discovers a deviation in one or more state variables... Continuously exceeding the threshold This automatically triggers the third step of diagnosis and optimization deduction, which involves analyzing the causes and exploring possible optimizations. Once the diagnostic process is activated, it first identifies the state variable nodes exhibiting significant deviations (e.g., black powder purity). Then, using the directional causal dependency network defined by the CCG ontology and physical layers, reverse tracing is performed. The diagnostic algorithm backtracks layer by layer along the directed edges entering the node (i.e., edges from "cause" to "effect"), evaluating the contribution of the deviation between the predicted and actual values ​​of each parent node (causal variable). Source tracing reveals purity. Mainly affected by the "uniformity index of crushed particle size distribution" "Influence, relationship" And currently Actual estimated value Much lower than its baseline forecast ;Continue to trace back, impact The key parent node is "tool wear". "and discovered" (here) (Indicating wear level, in percentage) significantly higher than the baseline prediction used. ,Right now: ( )= ( Based on this, the diagnostic conclusion is generated: "Accelerated tool wear is the root cause of uneven particle size in crushing, which in turn leads to a decrease in the purity of black powder." This process formally utilizes conditional probability or sensitivity analysis in the graphical model to locate the root cause node with the largest deviation between the prediction and the actual result.

[0061] Based on this diagnostic result, the generated intervention hypothesis will directly target the identified root cause, "accelerated tool wear," and its direct impact chain. This includes: Intervention Hypothesis A (Mitigating Wear Effect): Reducing Crusher Speed From the current rpm (revolutions per minute) adjusted to rpm, designed to reduce impact frequency ( and (positive correlation), thereby slowing down the rate of wear accumulation. ( and (Positive correlation).

[0062] Intervention Hypothesis B (Directly Addressing the Root Cause): Implement a preventative tool change procedure to reset the state variable "effective tool wear" to a lower initial value in the model.

[0063] Intervention Hypothesis C (Assistive Measures): Increase coolant flow rate ,from L / min (liters per minute) increased to L / min, indirectly mitigating the wear rate by improving heat dissipation.

[0064] If the difference does not exceed the threshold ( If the fault is not detected, the diagnostic process will not be triggered. Instead, it will directly follow its inherent scheduling strategy and periodically execute proactive optimization simulations. The intervention hypotheses generated in this case are not targeted at specific faults, but rather seek performance improvements based on existing stable operation. For example, to optimize the core objective of "cobalt recovery rate," the following hypotheses might be generated: Intervention Hypothesis D (Optimized Sorting): Fine-tuning the wind speed of the air separator ,from tentatively increased to m / s (meters per second) m / s.

[0065] Intervention hypothesis E (optimized pretreatment): Adjust the pre-discharge current I, increasing it by 10%.

[0066] Regardless of the path along which it is generated, the subsequent estimation will be based on the second state. As the latest and most reliable starting point for all state variables, for each intervention hypothesis, parameter changes are simulated in the CCG, and then all physical mechanism equations are driven in a forward deduction to obtain the intervention prediction trajectory for the future time window.

[0067] Subsequently, the decision-making module evaluates the final state results of all intervention prediction trajectories. The evaluation criteria include hard constraints and optimization objectives. Hard constraints must be met, such as: a safe upper limit for tool wear, bearing temperature, and a lower limit for black powder purity. The optimization objective is to maximize cobalt recovery. First, any assumptions that violate constraints during the prediction period are eliminated. Then, from the remaining feasible assumptions, the scheme that maximizes the predicted final state cobalt recovery rate is selected as the optimal intervention assumption.

[0068] Ultimately, the optimal intervention hypothesis is transformed into executable instructions (i.e., decision instructions). After the instructions are issued, monitoring continues, and new sensor data initiates the next round of state reconstruction, thus forming an autonomous closed loop of continuous perception, diagnosis, optimization, and execution. In this example, the diagnostic conclusion of "accelerated tool wear" directly shapes the direction of intervention hypothesis generation, ensuring the targetedness and effectiveness of subsequent decision-making actions.

[0069] Second embodiment, see Figure 7 This embodiment aims to disclose an intelligent dismantling system for lithium battery recycling, including a processor and a memory. The processor executes a program stored in the memory to implement a control method for lithium battery recycling. It also includes: The first sensing data acquisition module is used to acquire the first sensing data; the first sensing data refers to the sensing data collected by the sensor at the current moment. The first state estimation generation module is used to receive the first sensing data, and reconstruct the state based on the first sensing data to generate the first state estimate; The baseline predicted trajectory generation module is used to receive the first state estimate, use the first state estimate as the initial condition, drive the forward inference of the computable causal graph (CCG), and generate the baseline predicted trajectory. The second sensing data acquisition module is used to acquire second sensing data; the second sensing data refers to the sensing data collected by the sensor at the next moment. The second state estimation generation module is used to receive the second sensing data, reconstruct the state based on the second sensing data, and generate the second state estimate.

[0070] The comparison and diagnosis module is used to receive the second state estimate and compare the second state estimate with the baseline prediction value generated at the same time as the baseline prediction trajectory; If the difference exceeds a preset threshold, a diagnosis is triggered and a diagnosis result is output. Multiple intervention hypotheses are generated based on the diagnosis result. If the difference does not exceed the preset threshold, multiple intervention hypotheses seeking proactive optimization are generated directly. The intervention prediction trajectory generation module is used to sequentially deduce intervention hypotheses and generate a set of intervention prediction trajectories. The decision determination module is used to select the intervention hypothesis corresponding to the best intervention prediction trajectory in the set of intervention prediction trajectories and generate the final decision instruction; The disassembly decision command output module transmits the final decision command to the execution end used for lithium battery disassembly. Deeply integrated into the equipment's underlying control system, this module is not a simple command relay but a control node with highly reliable interlocking logic. When the decision-making module selects the optimal intervention assumption (such as emergency speed reduction), this module can send the highest-priority hard command to the crusher drive inverter via hard-wired connection or a secure industrial bus to achieve instantaneous speed adjustment. Simultaneously, it sends a pause signal to the upstream feeding PLC to interrupt the process and sends an early warning to the fire and environmental monitoring system to initiate a local emergency response. For planned maintenance commands, the system can also dynamically adjust the production cycle through the upper-level production scheduling interface to ensure minimal overall capacity loss during maintenance. The entire system forms a rigid closed loop with safety as its absolute core, capable of immediate and autonomous response to physical risks.

[0071] Third embodiment, see Figure 8 This embodiment aims to disclose an intelligent sorting system for lithium battery recycling, including a processor and a memory. The processor executes a program stored in the memory to implement a control method for lithium battery recycling. It also includes: The first sensing data acquisition module is used to acquire the first sensing data; the first sensing data refers to the sensing data collected by the sensor at the current moment. The first state estimation generation module is used to receive the first sensing data, and reconstruct the state based on the first sensing data to generate the first state estimate; The baseline predicted trajectory generation module is used to receive the first state estimate, use the first state estimate as the initial condition, drive the forward inference of the computable causal graph (CCG), and generate the baseline predicted trajectory. The second sensing data acquisition module is used to acquire second sensing data; the second sensing data refers to the sensing data collected by the sensor at the next moment. The second state estimation generation module is used to receive the second sensing data, reconstruct the state based on the second sensing data, and generate the second state estimate.

[0072] The comparison and diagnosis module is used to receive the second state estimate and compare the second state estimate with the baseline prediction value generated at the same time as the baseline prediction trajectory; If the difference exceeds a preset threshold, a diagnosis is triggered and a diagnosis result is output. Multiple intervention hypotheses are generated based on the diagnosis result. If the difference does not exceed the preset threshold, multiple intervention hypotheses seeking proactive optimization are generated directly. The intervention prediction trajectory generation module is used to sequentially deduce intervention hypotheses and generate a set of intervention prediction trajectories. The decision determination module is used to select the intervention hypothesis corresponding to the best intervention prediction trajectory in the set of intervention prediction trajectories and generate the final decision instruction; The sorting decision command output module transmits the final decision command to the execution end used for lithium battery sorting. This module is an intelligent gateway supporting multiple industrial communication protocols. It converts the optimized parameter values ​​output by the decision module into standard control commands recognizable by the underlying actuators (e.g., writing new frequency setpoints to the fan inverter via the PROFINET network, or adjusting the high-voltage power supply voltage via an analog output card). Furthermore, its execution concept extends beyond stand-alone control to the coordination of the entire production system. For example, if a diagnostic check reveals that a batch of materials generally has high moisture content affecting sorting efficiency, the system can, while adjusting process parameters, suggest or trigger coordinated adjustments to the parameters of the upstream drying process through an interface with the Manufacturing Execution System (MES). It can also automatically update production formulas or generate quality inspection work orders based on optimized decisions, achieving closed-loop quality management.

[0073] While dismantling and sorting systems serve different stages of lithium battery recycling, they both embed the same CCG-driven autonomous decision-making loop: continuous state perception and reconstruction, model-based prediction, causal-driven diagnosis, and simulation-verified decision generation. Their fundamental difference stems from scenario constraints: dismantling systems directly face the risks of mechanical damage and thermal runaway, thus their system design emphasizes safety rigidity, immediate response, and equipment health management. The execution end is deeply integrated with the underlying control of heavy equipment, emphasizing direct and mandatory intervention capabilities. Sorting systems, on the other hand, address the complex variations in material characteristics. Their system design pursues quality flexibility, parameter precision, and process coordination, with the execution end focusing on flexible integration with process control and production management systems to achieve precise and optimized adjustments. The two complement each other: the intelligent dismantling system provides stable and safe material input for subsequent processes, while the intelligent sorting system maximizes the value extraction of input materials, jointly constructing a complete intelligent production line system with autonomous perception, analysis, decision-making, and execution capabilities that runs through the key stages of lithium battery recycling.

[0074] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A control method for lithium battery recycling, characterized by, The method comprises the following steps: obtaining a first state estimation; driving a digital twin forward to obtain a baseline prediction trajectory based on the first state estimation as an initial condition; obtaining a second state estimation; comparing the second state estimation with a baseline prediction value generated at the same time point as the baseline prediction trajectory, and calculating a difference; if the difference exceeds a preset threshold, triggering diagnosis and outputting a diagnosis result, and generating an intervention hypothesis based on the diagnosis result; if the difference does not exceed the preset threshold, directly generating an intervention hypothesis; taking the second state estimation as a starting point, sequentially deducing the intervention hypothesis in the digital twin to obtain a set of intervention prediction trajectories; selecting an intervention prediction trajectory corresponding to the intervention hypothesis in the set of intervention prediction trajectories to generate a final decision instruction, and outputting the final decision instruction.

2. The control method for lithium battery recycling according to claim 1, characterized by, The method comprises the following steps: obtaining perception data at a current time point; reconstructing the state of the perception data at the current time point to obtain a virtual state of the lithium battery at the current time point; generating the first state estimation based on the virtual state of the lithium battery at the current time point.

3. The control method for lithium battery recycling according to claim 1, characterized by, The method comprises the following steps: obtaining perception data at a next time point; reconstructing the state of the perception data at the next time point to obtain a virtual state of the lithium battery at the next time point; generating the second state estimation based on the virtual state of the lithium battery at the next time point.

4. The control method for lithium battery recycling according to claim 2 or 3, characterized by, The method comprises the following steps: preprocessing the perception data; inputting the perception data after preprocessing into a state reconstruction algorithm; calculating a set of state variable values that best fit all observation values and physical constraints in probability based on the state reconstruction algorithm; generating the virtual state of the lithium battery based on the set of state variable values.

5. The control method for lithium battery recycling according to claim 4, characterized by, The state reconstruction algorithm comprises the following steps: an internal computable causal diagram is built, in which the nodes define the state variables and the edges define the physical or chemical constraint equations between the state variables.

6. The control method for lithium battery recycling according to claim 1, characterized by, The method comprises the following steps: loading the current set of process parameters; driving all mechanism equations in the digital twin to forward deduce a fixed time window; generating a baseline prediction trajectory, which is used to describe the baseline prediction values of the state variables at future time points.

7. The control method for lithium battery recycling according to claim 1, characterized by, The method comprises the following steps: locking the state variable with the deviation; locating the root cause node with the largest prediction and actual deviation by reversely tracking the causal dependence path in the digital twin; outputting the diagnosis result based on the root cause node.

8. The control method for lithium battery recycling according to claim 1, characterized by, The method comprises the following steps: evaluating the final results of all intervention prediction trajectories in the set of intervention prediction trajectories; the evaluation criteria include the improvement of the core target indicators and whether the safety and quality constraints are violated; selecting all feasible intervention prediction trajectories that meet the constraint conditions, and selecting the intervention prediction trajectory with the best optimization effect on the core target from the feasible intervention prediction trajectories. 9.The intelligent disassembling system for lithium battery recycling comprising a processor and a memory, wherein the processor is configured to execute a program stored in the memory to implement the control method for lithium battery recycling according to claim 1. The method further comprises the following steps: a baseline prediction trajectory generation module is configured to receive the first state estimation, drive the digital twin forward based on the first state estimation as an initial condition, and generate a baseline prediction trajectory; a comparison and diagnosis module is configured to receive the second state estimation, compare the second state estimation with a baseline prediction value generated at the same time point as the baseline prediction trajectory, and calculate a difference; if the difference exceeds a preset threshold, triggering diagnosis and outputting a diagnosis result, and generating an intervention hypothesis based on the diagnosis result; If the difference does not exceed the preset threshold, an intervention hypothesis is directly generated; an intervention prediction trajectory generation module, configured to sequentially deduce the intervention hypothesis to generate a set of intervention prediction trajectories; a decision determination module, configured to select an intervention hypothesis corresponding to one of the intervention prediction trajectories to generate a final decision instruction; and, a disassembly decision instruction output module, configured to transmit the final decision instruction to an execution end for disassembly. 10.An intelligent sorting system for lithium battery recycling, comprising a processor and a memory, the processor being configured to execute a program stored in the memory to implement the control method for lithium battery recycling according to claim 1, characterized in that, Further comprising: a baseline prediction trajectory generation module, configured to receive the first state estimation, take the first state estimation as an initial condition, drive a digital twin forward to deduce, and generate a baseline prediction trajectory; a comparative diagnosis module, configured to receive the second state estimation, compare the second state estimation with a baseline prediction value generated at the same time by the baseline prediction trajectory, and calculate a difference; if the difference exceeds the preset threshold, triggering diagnosis and outputting a diagnosis result, and generating an intervention hypothesis based on the diagnosis result; if the difference does not exceed the preset threshold, an intervention hypothesis is directly generated; an intervention prediction trajectory generation module, configured to sequentially deduce the intervention hypothesis to generate a set of intervention prediction trajectories; a decision determination module, configured to select an intervention hypothesis corresponding to one of the intervention prediction trajectories to generate a final decision instruction; and, a sorting decision instruction output module, configured to transmit the final decision instruction to an execution end for sorting.

Images