Lithium battery production early warning and optimization method based on X-ray detection and neural network

Abstract

A lithium battery production early warning and optimization method based on X-ray detection and a neural network comprises the following steps: (1) obtaining operation data and an X-ray detection image of a lithium battery cell production process, and constructing lithium battery production big data; (2) constructing a BP neural network model reflecting a lithium battery process-defect mapping relation; (3) updating the lithium battery production big data, and performing prediction and alarm on lithium battery production abnormity; (4) optimizing production early warning and process of the lithium battery to form closed-loop control of'prediction-early warning-intervention-optimization '; and (5) optimal process parameters are calculated through the NSGA-BP model, and optimal production is carried out. According to the method, the X-ray detection data and the process parameters are fused, an intelligently-driven closed-loop optimization system is constructed, the bottleneck of process-defect nonlinear modeling is broken through, and dynamic compensation and real-time early warning are achieved. The defect identification accuracy is greater than 99%; the process parameter optimization error is less than 1%; the process adjustment response time is less than or equal to 60 seconds; and the yield is improved by 10%.

Description

Technical Field

[0001] The present invention relates to a lithium battery production early warning and optimization method based on X-ray detection and neural network, belonging to the technical field of lithium battery production. Background Art

[0002] By utilizing the varying attenuation levels of X-rays when penetrating different materials, the internal structure of lithium-ion batteries can be imaged. X-ray image analysis can effectively identify various defects within lithium-ion batteries, such as poor electrode alignment, incorrect electrode count, electrode wrinkles, and poor cell housing dimensions. Therefore, X-ray inspection has become an indispensable quality control step in the lithium-ion battery production process.

[0003] Lithium-ion battery production involves numerous and complex processes. Improper process parameters or environmental changes during production can lead to defects in lithium-ion batteries, resulting in an increase in defective products. Currently, X-ray inspection technology in the lithium battery production field can perform comprehensive online X-ray inspections on different types of lithium batteries and different production processes. This effectively identifies potential defects in lithium battery products and intelligently determines whether individual lithium-ion batteries are qualified, thus ensuring product safety and reliability.

[0004] The "Lithium-ion Battery X-Ray Online Inspection Device" (CN20120289578.4) integrates X-ray imaging with an automated conveying platform, using a control host to analyze image data in real time and sort qualified / unqualified batteries. The "Lithium-ion Battery Self-Discharge Detection System" (CN201510854690.1) combines X-ray tomography with electrochemical parameters (voltage, current, and impedance) to achieve real-time, non-destructive testing of the self-discharge rate. The "Neural Network-Based Lithium Battery Charging Detection System" (CN201811404518.6) uses a neural network to train charging data to predict the remaining charging time and optimize the charging process in real time using temperature and current sensors. The "X-ray Radiography-Based Fault Detection and Prediction for Battery Cells" (CN202310520029.1) combines X-ray imaging with an intelligent classification algorithm to achieve non-destructive battery testing and predict battery defects. Publication number CN 119494839A, “A method and system for detecting battery pole pieces based on image recognition,” uses X-ray transmission images to analyze the angular alignment and main body wrinkles of battery cell pole pieces, optimizing the detection process and reducing costs. Publication number CN222514611 U, “A test and detection device for lithium battery charging management chips,” collects data through multiple sensors (temperature, gas, smoke), optimizes safety warnings through multimodal data fusion and AI algorithms, and combines microprocessor analysis thresholds to achieve early warning. Publication number CN 118941864 A, “A method, device, equipment, and medium for detecting battery positive and negative poles based on X-ray images,” uses a battery boundary point detection model to identify and correct X-ray battery images, obtain the battery’s positive and negative pole boundaries, and improve the accuracy and efficiency of detecting the battery’s positive and negative poles.

[0005] X-ray inspection is widely used in lithium battery defect detection (e.g., electrode alignment, wrinkles, etc.). However, existing patents (e.g., CN 20120289578.4 and CN 201510854690.1) primarily focus on defect detection and classification, without addressing the correlation analysis between defects and process parameters. Current process adjustments rely on manual experience, resulting in response lags (>30 minutes) and difficulty in resolving parameter coupling effects.

[0006] The aforementioned patent document embodies the collaborative application of "X-ray detection + intelligent modeling to identify lithium battery defects and sort out qualified / unqualified products." However, there is no method for deducing the cause of product failure based on X-ray detection. The lithium battery production process is complex, and many factors can cause product failure. Currently, the main reliance is on engineers' experience to analyze the causes of defects detected by X-ray detection and adjust the process. This has problems such as slow response speed, low accuracy, and low efficiency. In addition, the lithium battery production process is complex, involving multiple steps and process parameters. Any slight deviation can result in a failed product. Therefore, there is an urgent need for an intelligent and automated solution to achieve real-time monitoring and optimization of the production process.

[0007] X-ray inspection is widely used in lithium battery defect detection (e.g., electrode alignment, wrinkles, etc.). However, existing patents (e.g., CN 20120289578.4 and CN 201510854690.1) primarily focus on defect detection and classification, without addressing the correlation analysis between defects and process parameters. Current process adjustments rely on manual experience, resulting in response lags (>30 minutes) and difficulty in resolving parameter coupling effects.

[0008] The current lithium battery intelligent manufacturing industry faces three major technical bottlenecks: (1) Process-defect correlation modeling is difficult, relying on manual experience and slow response speed; traditional methods have difficulty capturing nonlinear relationships; (2) Process adjustment lags, with the average time from X-ray inspection results to parameter adjustment exceeding 30 minutes, model update lags, and cannot adapt to dynamic production lines; (3) Multi-parameter coupling effects, with single parameter optimization easily leading to secondary defects. In addition, existing technologies also have obvious deficiencies in lithium battery defect prediction, such as low accuracy of traditional statistical methods or simple models, inability to handle complex nonlinear relationships, and unstable reliance on expert experience. Summary of the Invention

[0009] The purpose of this invention is to solve the above three major technical defects of existing lithium battery intelligent manufacturing; to provide a lithium battery production early warning and optimization method based on X-ray detection and neural network, so as to realize production abnormality risk prediction and process parameter optimization, and improve production yield and production efficiency.

[0010] The technical solution of the present invention is as follows: a lithium battery production early warning and optimization method based on X-ray detection and neural network, the steps are as follows: (1) By extracting internal defect features from lithium battery X-ray images in real time, we can obtain X-ray imaging big data of lithium battery defects; obtain the operating data and X-ray detection images of the lithium battery cell production process, construct lithium battery production big data, extract features from the lithium battery defective product production big data, and determine the main equipment and process parameters that cause cell defects; (2) Extracting big data of key production process parameters of lithium batteries, constructing a back propagation neural network (BP) that reflects the process-defect mapping relationship of lithium batteries, constructing an image-defect convolutional neural network, using the YOLO algorithm to identify X-ray images, extracting six types of feature classifications and grades: pole piece alignment, pole piece wrinkle area, pole piece number, battery cell shell size, negative electrode redundancy, and welding defects; (3) Update the big data of lithium battery production; use the genetic algorithm (GA) to optimize the BP neural network process of the lithium battery process-defect mapping relationship, optimize the model parameters, train and learn; generate a production process early warning model to predict and alarm lithium battery production anomalies; (4) Set the lithium battery production objective function; use the NSGA-II algorithm on the trained BP neural network to generate a lithium battery production optimization model, optimize the lithium battery production warning and process; output monitoring warning and production process optimal parameters; automatically or with the assistance of the operator through the control system to make dynamic adjustments, forming a closed-loop control of "prediction-warning-intervention-optimization"; (5) Improve the non-dominated sorting multi-objective genetic algorithm (NSGA-II) and the lithium battery process-defect BP neural coupling lithium battery production optimization process, optimize the multi-objective parameters, and generate a lithium battery production optimization model.

[0011] NSGA-BP is a hybrid algorithm that combines the multi-objective optimization capabilities of NSGA with the predictive power of BP neural networks. It is commonly used for parameter optimization or prediction tasks (such as process parameter tuning) in complex industrial scenarios. This paper uses a multi-objective genetic algorithm-back propagation neural network (NSGA-BP) model to calculate optimal process parameters. These parameters are used as process setpoints for lithium battery production equipment, allowing each equipment to independently adjust its process for optimal production.

[0012] The core goal of the BP neural network model of the lithium battery process-defect mapping relationship is to use back propagation and weight updating to predict or diagnose defects that may occur in the lithium battery production process; by learning historical production data, including taking process parameters as input and corresponding defect detection results as output, a complex nonlinear mapping relationship is established.

[0013] The BP neural network model of the lithium battery process-defect mapping relationship adopts a typical single hidden layer BP neural network, which is divided into an input layer, a hidden layer, and an output layer. Their mathematical expressions are as follows: Input layer: Receives input vector as: ; Hidden layer: The net input to the hidden layer neurons is: ; , arrive ; in, The input layer neurons to the hidden layer The connection weights of neurons; The hidden layer The bias of the neuron The output of the hidden layer neurons , ,in is the activation function; For the receives the input vector, i.e. the first The value of a process parameter; There are 8 process parameters, including coating speed, roller pressure, roller speed, welding current, oven temperature, slurry viscosity / solid content, winding / stacking tension, ambient temperature and humidity, and injection volume / injection rate; Output layer: The net input to the output layer neurons ; , arrive , is the number of neurons in the hidden layer; in, Hidden layer neurons to the output layer The connection weights of neurons; The output layer The bias of each neuron; For the The final output of the output layer neurons; ,in is the activation function of the output layer; The back propagation and weight update: The core is to calculate the gradient of the loss function L for all weights W and bias b ( ), use the chain rule to calculate backward from the output layer to the input layer, and then use gradient descent or its variants to update the parameters: The loss function L is: measure the network output and the true value the gap; The weight W and bias b are: ; Among them, η is the learning rate, which controls the update step size; The new weight calculated this time; is the old weight from last time; The new bias calculated for this time; The old offset from last time.

[0014] The BP neural network model of the lithium battery process-defect mapping relationship includes input, output, model training and model application.

[0015] The input, key process parameters, include coating speed, roller pressure, roller speed, oven temperature, slurry viscosity / solid content, winding / stacking tension, ambient temperature and humidity, and injection volume / injection rate.

[0016] The output includes the defect detection results of the batch of batteries, defect probability; main defect types, including short circuit, micro short circuit, lithium plating, uneven thickness or leakage; defect classification; and the severity of the defect, including minor, moderate or severe.

[0017] The model training inputs the process parameters into the input layer of the BP network to enter the model training.

[0018] The model is applied by inputting the production process parameters of a new, unknown batch of lithium batteries into the trained model; the model performs forward calculations; and the output layer provides defect prediction results for the batch, including probability, classification, or grading.

[0019] The optimization process of the lithium battery process-defect BP neural network using the genetic algorithm GA is as follows: (1) Data collection and preprocessing Data is collected from battery cell X-ray image detection equipment, production equipment sensors, and process parameter recording systems through lithium battery X-ray sources / imagers. The data is cleaned, normalized, and feature extracted to construct a large data set of production process parameters and defect characteristics.

[0020] (2) BP model construction and training BP neural network and decision tree machine learning algorithms are used to build a process parameter-defect optimization model; the model is trained through historical data to optimize the network structure and parameters, thereby improving the model's prediction accuracy and generalization ability.

[0021] (3) Process parameter optimization and control Based on the output of the BP neural network and process parameter optimization model, real-time warnings and dynamic adjustment of the set values ​​of production equipment process parameters are provided. A real-time feedback mechanism ensures the stability and consistency of the production process. The BP model is not only used for offline analysis or post-batch testing, but is also deeply integrated into the production line manufacturing execution system / supervisory control and data acquisition system (MES / SCADA) to achieve real-time defect probability prediction at the millisecond level. MES (Manufacturing Execution System) is positioned as a factory-level management system, connecting the enterprise planning layer and the workshop equipment layer, and is used for production scheduling and real-time monitoring, material tracking and quality management, equipment maintenance collaboration and performance analysis. SCADA (Supervisory Control and The Data Acquisition (DPA) system is positioned at the monitoring layer, directly connecting to field equipment such as PLCs and sensors for real-time data acquisition and equipment status monitoring, process alarm management and remote control, and industrial process visualization. When the proposed model predicts that the defect probability exceeds the standard, it generates specific, actionable process parameter adjustment recommendations, such as "increase the temperature in the third zone of the coating oven by 2°C and reduce the roller pressure by 5%." Dynamic adjustments are made automatically or with the assistance of the operator through the control system, forming a closed-loop control system of "prediction-warning-intervention-optimization." For predictive intelligent sorting / disposal, the model output (probability, classification, and grading) directly drives automated sorting equipment, such as automatically sorting high-short-circuit risk cells to special inspection lines or scrap channels, and diverting slightly defective cells to the repair process. This seamless, intelligent linkage from prediction results to action execution is a key feature.

[0022] The image-defect convolutional neural network is constructed, and the YOLO algorithm is used to identify X-ray images to extract the feature classification and grade of the electrode alignment, electrode wrinkle area, number of electrodes, battery cell shell size, negative electrode redundancy, and welding defects; a classifier is added on the existing basis, and the new category samples are re-trained and tested to further optimize the classifier and improve the classification performance of the classifier.

[0023] Construct defect-related feature engineering and data sets, introduce new defect characterization data, and introduce monitoring data sources that can more directly / deeply reflect defects as model input in addition to conventional process parameters, including online / offline detection data fusion. The X-ray imaging features of electrodes and battery cells are integrated into the BP model through feature extraction or fusion methods, that is, CNN processing images + BP processing process parameters, to construct a multimodal defect prediction model; based on process mechanism and data analysis, construct derivative features with physical significance or strong predictive power, including coating drying rate gradient, roller compaction density change rate, and interaction terms between adjacent process parameters, as model input, which makes up for the shortcomings of existing process parameters.

[0024] The BP neural network model of the lithium battery process-defect mapping relationship is divided into an input layer, a hidden layer and an output layer; the input layer includes coating speed, roller pressure, roller speed, oven temperature, slurry viscosity / solid content, winding / stacking tension, ambient temperature and humidity, injection volume / injection rate process parameters; the hidden layer includes a three-layer network of convolutional layer, pooling layer and fully connected layer, and the activation function is ReLU; the output layer includes the total probability of output defects, the probability distribution of output defect types and the probability distribution of output defect levels.

[0025] The neural network model takes the main production factors that lead to internal defects of lithium batteries as input and the internal defect characteristics of lithium batteries as output, and establishes a forward multi-layer BP neural network model of lithium battery production process-defects; based on the BP neural network algorithm and samples, multiple learnings are performed to calculate the weights and thresholds of each layer, thereby solving the calculation model between input and output; the simulation error of the model is less than 1%, that is, it has good memory ability. Using the established model, untrained production conditions are calculated to calculate the most likely production factors that cause battery defects, and the error is required to be within 1%.

[0026] The defect characteristics include at least one of poor electrode alignment, missing electrode, wrinkled electrode, poor cell shell size, negative electrode redundancy, welding defects, and other internal defects; the defect characteristics include at least one of poor electrode alignment, missing electrode, wrinkled electrode, poor cell shell size, negative electrode redundancy (missing, excessive), welding defects (cold solder joints, over-soldering, etc.), and other internal defects (impurities, short circuits, fractures, etc.).

[0027] The main production equipment, process parameters and control measures of the lithium battery include at least one of the following: The slurry quality, coating uniformity and rolling thickness in the electrode preparation section are key control points; the equipment conditions of the mixer, coater and roller press directly affect the defects of the lithium battery. The electrode defects are adjusted by stirring time, drying temperature, coating speed, nozzle condition and rolling condition.

[0028] The electrode cutting accuracy, electrode winding / stacking quality, battery cell assembly quality, and welding quality in the battery cell assembly section are key control points; the equipment conditions of the mold laser knife, winding machine, stacking machine, and welding machine directly affect the defects of lithium batteries.

[0029] The process adjustment measures of the winding machine and the stacking machine include: unwinding correction, stroke correction, sheet feeding correction, infrared positioning, and tension control; the stacking machine adjusts the tightness of the stacked battery cells through the temperature and pressure of the hot pressing process; the welding power and welding current adjust the welding quality.

[0030] The lithium battery production optimization model, through mathematical modeling, algorithm design and data analysis, maximizes production efficiency, reduces costs, improves yield, or optimizes other key indicators while meeting quality, safety, and environmental constraints. Its construction method and mathematical expression are as follows: (1) Objective function, optimization goal: Minimize costs, including raw materials, energy consumption, equipment depreciation, and labor; Maximize production efficiency, including output per unit time and equipment utilization; Optimal quality, improved yield, and consistency control; Minimize energy consumption and carbon emissions to meet green manufacturing requirements; Timely order delivery and reduced delay penalties.

[0031] (2) Decision variables: production scheduling (order sequence, batch size), equipment parameter setting (coating speed, drying temperature, injection volume), resource allocation (manpower, equipment switching strategy), inventory control (raw materials, semi-finished product buffer).

[0032] (3) Constraints, including physical constraints and process constraints; Physical constraints, including equipment capacity limits and process parameter ranges (e.g., temperature / humidity limits); quality constraints, including key parameter tolerances (coating thickness ±2μm) and electrochemical performance requirements; Process constraints, process sequence (coating → rolling → slitting → winding), cleaning time; Resource constraints, material supply, equipment maintenance time, and personnel scheduling; Environmental safety, solvent volatilization limits, and wastewater treatment capabilities.

[0033] (4) Optimization focus of key links, including electrode manufacturing, coating uniformity control, and drying energy consumption optimization; assembly process, winding / stacking accuracy, and vacuum injection efficiency; formation / capacity separation, and charge and discharge strategy optimization (time / energy consumption balance); supply chain collaboration, and matching of raw material procurement with production rhythm.

[0034] Taking defect risk minimization as the objective function, the process parameter combination is iteratively optimized; each batch of new data triggers the update of model parameters and the learning rate is dynamically adjusted.

[0035] The process parameter-defect optimization model introduces new defect characterization data: in addition to conventional process parameters, a new monitoring data source that can more directly / deeply reflect defects is introduced as model input; including online / offline detection data fusion, the X-ray imaging features of the pole piece and battery cell are integrated into the BP model through innovative feature extraction or fusion methods (such as CNN processing image + BP processing process parameters) to build a multimodal defect prediction model; advanced feature construction, based on process mechanism and data analysis, constructs derivative features with physical significance or strong predictive power (such as coating drying rate gradient, roller compaction density change rate, interaction terms between adjacent process parameters) as model input; high-quality annotated data set construction method: efficient, accurate and low-cost defect sample annotation method (such as semi-supervised learning based on the use of a large amount of unlabeled data, innovative online detection and offline disassembly verification linkage mechanism).

[0036] After training, the lithium battery process-defect BP model is output to generate a production process early warning model; the lithium battery production optimization process is coupled with the improved non-dominated sorting genetic algorithm NSGA-II and the lithium battery process-defect BP neural coupling to optimize multi-objective parameters and generate a lithium battery production optimization model.

[0037] The beneficial effect of the present invention is that the present invention uses lithium battery X-ray detection technology to obtain various defects inside lithium batteries, and forms lithium battery defect production big data with production processes and key production process parameters, conducts AI learning, and performs intelligent analysis to realize intelligent risk judgment of production processes, process parameters, etc. of lithium battery-related manufacturing processes, and provides production risk warnings, thereby providing a higher precision, efficiency, and reliability lithium-ion battery intelligent manufacturing solution, improving production yield and production efficiency.

[0038] This invention integrates X-ray detection data with process parameters to construct an intelligently driven closed-loop optimization system, breaking through the bottleneck of process-defect nonlinear modeling and achieving dynamic compensation and real-time early warning. Defect recognition accuracy is >99%; process parameter optimization error is <1%; process adjustment response time is ≤60 seconds; and the yield rate is increased by 10%. After intelligent transformation, a production line with an annual output of 100 million 26700-5000mAh lithium batteries has shortened its process adjustment response time to within 60 seconds, increasing annual output by 10 million units and generating an additional 50 million yuan in profits. The defect rate has also been reduced from 0.2% to 0.01%, increasing the number of good products by 2 million, generating direct economic benefits of 20 million yuan and total economic benefits of 70 million yuan. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 A flowchart for lithium-ion battery production early warning and process optimization based on X-ray detection and neural networks; Figure 2This is the workflow diagram for the lithium battery intelligent manufacturing production line; Figure 3 This is the optimization flow chart of GA for lithium battery process-defect BP neural network; Figure 4 Optimized flow chart for lithium battery production by coupling NSGA-II with BP; Figure 5 This is the overall architecture diagram of the control system; Figure 6 Flowchart for lithium-ion battery defect model detection; Figures 7 (1) to 7 (9) are characteristic diagrams of defects in lithium battery X-ray imaging; Among them: Figure 7 (1) is the original X-ray imaging image of the lithium battery; Figure 7 (2) shows the poor size of the lithium battery in the shell; Figure 7 (3) shows the poor alignment of the lithium battery pole piece; Figure 7 (4) shows the missing lithium battery pole piece; Figure 7 (5) shows the wrinkled lithium battery pole piece; Figure 7 (6) shows the normal redundancy of the lithium battery, the missing redundancy, the excessive redundancy, and the bent pole piece end; Figure 7 (7) shows the wrinkled lithium battery pole piece; Figure 7 (8) shows the poor alignment of the lithium battery pole piece; Figure 7 (9) shows the broken lithium battery pole piece. DETAILED DESCRIPTION

[0040] With reference to the accompanying drawings, the specific implementation of the lithium battery production early warning and optimization method based on X-ray detection and neural network of the present invention is as follows.

[0041] 1. Analysis of lithium battery production process and defects and technical solutions to solve the defects 1. Determination of the main equipment and process parameters causing battery cell defects Lithium-ion battery production involves numerous steps and complex processes. Any error can reduce the battery's reliability. Lithium-ion battery production can generally be divided into three stages: electrode preparation, cell assembly, and post-filling.

[0042] (1) Electrode preparation section. In the electrode preparation section, the three processes of material drying, sol and slurry stirring directly determine the quality of the slurry. Material drying temperature is a key process parameter. It controls the stirring speed, dispersion speed, vacuum degree, stirring time and temperature of the mixer, sets the process parameters that match the formula, and generally uses a viscometer to monitor the slurry viscosity. The extrusion coating machine consists of unwinding, nozzle (equipped with a feeding system), drying tunnel, winding and other drive systems. The manufacturing accuracy of the nozzle, nozzle width, running speed, dynamic tension control, stability, drying method, air pressure, air supply position, air supply direction and temperature curve setting are all influencing factors. It is mainly necessary to control the width of the nozzle opening, temperature gradient distribution, air volume and belt speed to ensure that the electrode thickness is consistent, fully dried, and without cracks, curling and other phenomena. The electrode entering the rolling process must be properly dried, otherwise it is easy to lose powder and fall off during the rolling process. The uniformity and surface density of the coated electrode are necessary conditions for the quality of the rolled electrode to meet the standards. Factors such as the no-load roll gap of the roller press, the thickness of the rolled piece at the entrance, the rolling mill stiffness, the bearing oil film thickness, the deformation resistance of the rolled piece and the eccentricity of the roller are all influencing factors.

[0043] (2) Cell assembly section. Assembling the positive and negative electrodes and separators into a battery core is the core process of battery manufacturing and plays a decisive role in the electrical performance and safety of the battery. Using a mold laser cutter will cause rapid wear, short service life and poor compatibility. Therefore, attention should be paid to the dimensional accuracy and burr condition of the electrode cutting length and width. The core of cell assembly quality control is that the negative electrode should be excessively large compared to the positive electrode. Otherwise, lithium dendrites will easily accumulate during charging, causing internal short circuits in the battery and posing a safety hazard. During stacking and winding operations, the positive and negative electrodes should be completely separated by the separator, and the length and width of the separator should be greater than the negative electrode, and the length and width of the negative electrode should be greater than the positive electrode. During winding, the negative electrode wraps around the positive electrode. During stacking, the negative electrode has one more layer than the positive electrode. The unwinding correction, stroke correction, sheet feeding correction and infrared positioning functions of the winding machine and stacking machine should be ensured to ensure that the electrode is centered and matched, and to obtain qualified wound and stacked cells. The tightness of the contact between the electrode and the diaphragm affects the internal resistance of the battery cell, and the tension control of the winding machine is also a key indicator. The laminated structure is adjusted by controlling the temperature and pressure of the hot pressing process to adjust the tightness of the battery cell. The battery cell manufacturing process involves many welding-related steps, including lamination (tab welding), tab transfer, spot bottom welding, and cap welding. Strong welds are required, so welding power and welding current are important considerations, and the welding area and / or number of welds need to be precisely designed.

[0044] (3) Post-liquid injection process. The core processes after liquid injection are activation, formation, aging and secondary sealing, and their quality control has nothing to do with X-ray detection. Improper storage and drying temperature of raw materials will affect the physical and chemical properties, such as metal corrosion, membrane shrinkage and deformation, etc. The high or low ambient temperature directly affects the rate of solvent evaporation during the coating process, which in turn affects the coating effect of the electrode. In the production process of lithium-ion batteries, in addition to strictly controlling the quality control points of each process and optimizing the production process, the temperature, humidity and cleanliness of the production environment must also be strictly controlled. The battery production process is closely linked, and there are some important factors that require special attention, such as strict control of moisture, dust and metal foreign matter throughout the manufacturing process. Dust and metal foreign matter can cause the diaphragm aperture to be blocked or even pierce the diaphragm. In mild cases, it will increase self-discharge; in severe cases, it will cause short circuits, leakage and even explosions in the battery, causing safety accidents. Dust and metal foreign matter can be detected by X-rays.

[0045] Through analysis, the main equipment and process parameters that cause defects are identified: (1) The slurry quality, coating uniformity, and roller thickness in the electrode preparation section are key control points. The process parameters of the mixer, coater, and roller press will directly affect the defects of the lithium battery. The electrode defects can be adjusted by mixing time, drying temperature, coating speed, nozzle working condition, and roller pressing working condition. (2) The electrode cutting accuracy, electrode winding / stacking quality, battery assembly quality, and welding quality in the battery cell assembly section are key control points. The process parameters of the mold laser knife, winder and stacker, and welder will directly affect the defects of the lithium battery. The process adjustment measures for the winder and stacker include: unwinding correction, stroke correction, sheet feeding correction, infrared positioning, and tension control of the winder. The stacker adjusts the tightness of the stacked battery cell through the temperature and pressure of the hot pressing process. The welding power and welding current adjust the welding quality.

[0046] 2. Determination of lithium battery defect characteristics Construct the internal defect characteristics of lithium batteries identified by X-rays, perform fuzzy numerical conversion on the collected big data, and build a sample database.

[0047] This embodiment uses X-ray testing to perform 100% quality inspection on the battery cells. Unqualified battery cells have a variety of defects, and the defect characteristics are normalized. The defect characteristics of the battery cells, such as the length direction of the wound battery cell must ensure that the negative electrode has a margin, and the diaphragm wraps the negative electrode. The relative position of the pole piece is very important. The specific parameters include the diaphragm length m, the relative dimensions j and k of the negative electrode, etc. The relative position L between the positive and negative electrodes, etc. These dimensions are affected by the pole piece coating specifications, winding technology, etc. In the width direction of the pole piece, the negative electrode also needs to exceed the positive electrode by a certain margin V, and the diaphragm exceeds the negative electrode by a certain margin W. The alignment in the width direction directly depends on the accuracy of the winding process. When the accuracy is high, the margin can be smaller and the energy density of the battery can be higher, but the battery safety factor is reduced, and the process accuracy must be guaranteed, otherwise it will bring great safety hazards to the battery.

[0048] Through X-ray detection equipment, real-time image data of internal defects of lithium batteries are obtained, and the following key defect features are extracted: (1) electrode alignment (relative position of positive and negative electrodes); (2) number of electrodes (whether the number of layers meets the process requirements); (3) electrode wrinkles (areas with uneven grayscale values); (4) poor size of battery cell into the shell (the distance between the battery cell and the shell); (5) redundancy of negative electrode (missing, excessive); (6) welding defects (cold solder joints, over-soldering, etc.); (7) other internal defects (impurities, short circuits, fractures, etc.).

[0049] 3. Constructing a BP neural network model for lithium battery process-defect mapping The main production factors that lead to internal defects in lithium batteries are used as input, and the characteristics of lithium battery internal defects are used as output. A forward-looking, multi-layer BP neural network architecture is established to analyze lithium battery production processes and defects. Based on the BP neural network algorithm and samples, multiple learning cycles are performed to calculate the weights and thresholds of each layer, thereby solving the computational model between input and output. The model has a simulation error of less than 1%, indicating good memory. Using the established model, calculations are performed on untrained production conditions to calculate the most likely production factors that lead to battery defects, with an error requirement of less than 1%. The established model is then used to relearn new sample data, allowing the model to be constructed under new operating conditions, indicating the model's adaptive capabilities.

[0050] This example builds an image-defect convolutional neural network (CNN) and uses the YOLO algorithm to identify X-ray images, extracting and categorizing six features, such as electrode alignment and wrinkle area. Based on the existing model, a classifier is added, and training and testing are performed on samples from the new categories to further optimize the classifier and improve its classification performance. X-ray online inspection equipment inspects 100% of lithium batteries. If any product does not meet the requirements, the equipment automatically issues an alarm and rejects the defective product.

[0051] This example uses a genetic algorithm (GA) to optimize and calculate the weights and thresholds of each layer in the BP network, outputting a lithium battery process-defect BP model and generating a lithium battery production early warning model. An improved non-dominated sorting genetic algorithm (NSGA-II) is used to optimize multi-objective parameters and generate a lithium battery production optimization model. An online incremental learning algorithm (such as OS-ELM) is used to update model weights hourly to adapt to process fluctuations.

[0052] The lithium battery production early warning and process optimization process is as follows: (1) Data collection and preprocessing Collect data from X-ray inspection equipment, production equipment sensors, and process parameter recording systems. Clean, normalize, and extract features from the data to construct a large dataset of production process parameters and defect characteristics.

[0053] (2) BP model construction and training Using machine learning algorithms such as BP neural networks and decision trees, we build a defect optimization model based on process parameters. We train the model using historical data, optimize the network structure and parameters, and improve the model's prediction accuracy and generalization capabilities.

[0054] (3) Process parameter optimization and control Based on the output of the BP neural network and process parameter optimization model, the set values ​​of production equipment process parameters are dynamically adjusted. Through the real-time feedback mechanism, the stability and consistency of the production process are ensured.

[0055] 4. Control system composition of lithium battery intelligent manufacturing Hardware configuration of the control system for intelligent manufacturing of lithium batteries: (1) X-ray detection equipment: used to obtain real-time images of internal defects of lithium batteries.

[0056] (2) Industrial Internet of Things (IIoT) devices: used to collect the operating status and process parameters of production equipment.

[0057] (3) Edge computing devices: used for real-time data processing and model reasoning.

[0058] Control system software configuration for lithium battery intelligent manufacturing: (1) Data management platform: statistics, classification and quantification of defect data and process parameters.

[0059] (2) BP algorithm platform: realizes defect classification, process parameter optimization and production early warning.

[0060] (3) Visual interface: Real-time display of production status, defect distribution and process adjustment suggestions.

[0061] Further upgrades to the BP algorithm and model enhance the intelligent platform's adaptability and emergency response speed under extreme conditions, bringing customers a more extreme experience in production efficiency and product quality improvement. The software counts, classifies, and quantifies defect data, and classifies and quantifies production processes and process parameters to generate production-defect big data. After segmenting and adjusting the battery area, it is necessary to extract the defect feature parameters of the battery area and correctly classify them to facilitate the rapid detection of internal battery faults. The extracted assembly defect features must consider the impact of classifier selection and training methods on classification accuracy. 2. Specific embodiments Example 1: Workflow of a lithium battery intelligent manufacturing production line The workflow diagram of the lithium battery intelligent manufacturing production line in this embodiment is as follows: Figure 2 As shown: (1) Data acquisition: X-ray images and process parameters are collected simultaneously to build a multimodal data set;

[0063] (2) Defect classification: CNN identifies defects in lithium battery cell images and outputs quantitative features (such as alignment deviation ±0.02mm);

[0064] (3) Risk prediction: BP network calculates the defect probability under the current parameters. If the risk threshold is greater than 5%, an early warning is triggered.

[0065] (4) Process parameter optimization: Genetic algorithm generates the optimal parameter combination and sends it to coating machines, welding machines and other equipment in real time;

[0066] (5) Closed-loop control: the adjusted parameters are fed back to the model to form a production-testing-optimization closed loop.

[0067] A lithium battery process-defect mapping model was constructed. The input layer includes eight process parameters, such as coating speed, roller pressure, and welding current; the hidden layer includes a three-layer BP network with ReLU as the activation function; and the output layer includes defect probability (alignment deviation, wrinkle area, etc.).

[0068] Genetic algorithm optimization: Minimizing defect risk is the objective function, and process parameter combinations are iteratively optimized. Incremental learning mechanism: Each batch of new data triggers model parameter updates, and the learning rate is dynamically adjusted. This algorithm acquires operational data from lithium battery cell production processes and X-ray inspection data, extracts features from the production data of defective lithium battery products, and outputs monitoring warnings and optimal production process parameters.

[0069] Genetic algorithm GA optimizes the process of lithium battery process-defect BP neural network, such as Figure 3 As shown, after the training is completed, the lithium battery process-defect BP model and the production process early warning model are output.

[0070] Improved non-dominated sorting genetic algorithm NSGA-II and lithium battery process-defect BP neural coupling lithium battery production optimization process, such as Figure 4 As shown in the figure, multi-objective parameters are optimized to generate a lithium battery production optimization model.

[0071] like Figure 5 As shown in the figure, the control system of lithium battery intelligent manufacturing based on X-ray detection and neural network is as follows: the monitoring agency PC is loaded with lithium battery production monitoring software, and the BP neural network model of lithium battery production process-defects is the core monitoring module, which accepts production conditions (production process parameters) and production quality (defects), outputs monitoring warnings and optimal production process parameters, and sends execution instructions to the execution agency through the control agency PLC, so that each lithium battery production equipment operates according to the optimal parameters.

[0072] Example 2: Lithium Battery Cell Defect Detection Process The lithium battery cell defect detection process of this embodiment is as follows: Figure 6 As shown, defect data is preprocessed and a convolutional neural network model is built to intelligently identify defects in lithium battery cell images and output quantitative features. Convolutional neural networks (CNNs) are the most representative deep learning algorithms in the field of image recognition. The You Only Look Once (YOLO) family of algorithms is an optimal class of object detection algorithms within convolutional neural networks. This example develops a lithium battery defect extraction model based on the YOLOv8 deep learning algorithm.

[0073] The specific implementation is as follows: (1) Dataset annotation and image preprocessing. The locations of targets and the extent of defects in X-ray images of lithium battery cells are annotated, and defects are classified and assessed. The lithium battery cell images are randomly divided into training, validation, and test sets in proportion. Data augmentation is performed on the images in the training set, which effectively improves the generalization ability and robustness of the model.

[0074] (2) Construction of the YOLOv8 diagnostic model. The data-enhanced lithium battery cell X-ray images are used as input information for the YOLOv8 deep learning model training. The Backbone network in the YOLOv8 deep learning model is used to extract the feature representation of the lithium battery cell X-ray images layer by layer, and an attention mechanism is added. After the Neck part of the feature pyramid, feature maps of different scales are generated. The features after Neck processing are fed into the YOLOv8 deep learning model. The validation data set is re-entered into the training model to adjust the relevant parameters and obtain the optimal parameters. Finally, the test data set is input into the verified training model, and the defect classification and defect grading results are finally output.

[0075] (3) Evaluation indicators of model performance. A script file was written in Python to calculate the accuracy, sensitivity (recall rate), F1 value, and area under the precision-recall curve (PRAUC) to evaluate the detection and classification performance of the model.

[0076] X-ray detection of lithium battery cell defect characteristics images, as shown in Figures 7 (1) to 7 (9), shows the following: (1) The electrode alignment is poor, as shown in Figure 7 (3). The distance between the negative electrode and the positive electrode of the battery cell is not within the process value requirement. The image shows that the distance between the vertex of the negative electrode and the vertex of the positive electrode is greater or less than the process value;

[0077] (2) Missing pole pieces, as shown in Figure 7 (4). The number of core pole pieces does not meet the process value requirements. The image shows that the number of pole pieces is more or less than the process value;

[0078] (3) Pole wrinkles, as shown in Figure 7 (5). The poles of the battery cell were wrinkled during production. The image shows a black shadow with uneven grayscale value in the main body of the battery cell;

[0079] (4) The size of the battery cell in the shell is not good, as shown in Figure 7 (2). When the battery cell is installed in the battery shell, the structural distance between the battery cell and the shell does not meet the process value requirements. The image shows that the structural distance between the battery cell and the shell is not within the process value range;

[0080] (5) Negative electrode redundancy (missing, excessive), as shown in Figure 7 (6);

[0081] (6) Welding defects (false welding, over-welding, etc.);

[0082] (7) Other internal defects (impurities, short circuit, fracture, etc.). Relatively rare internal defects of lithium batteries, such as internal impurities, distortion of the positive and negative electrodes, short circuit of the positive and negative electrodes, fracture of the positive and negative electrodes, etc.

[0083] Example 3: Constructing a BP neural network model for lithium battery process-defect mapping The core goal of the BP neural network model in this embodiment is to use a backpropagation neural network to predict or diagnose defects that may occur during the lithium battery production process. It establishes a complex nonlinear mapping relationship by learning from historical production data (process parameters as input and corresponding defect detection results as output). This BP neural network model, which maps the lithium battery process-defect relationship, predicts defects by learning the relationship between process parameters and defect outcomes. The model consists of an input layer (which receives eight process parameters: coating speed, roller pressure, roller speed, oven temperature, slurry viscosity / solids content, winding / lamination tension, ambient temperature and humidity, and injection volume / injection rate); a hidden layer (which performs nonlinear feature extraction using activation functions such as ReLU); and an output layer (which generates prediction results). The eight input layer parameters are typically core parameters for key processes such as coating, roller pressing, and winding / lamination. The output layer outputs not only defect probabilities but also defect classification and grading information.

[0084] The core idea and process of the model (a) Input (X), key process parameters (e.g. coating speed, roller pressure, etc.);

[0085] (b) Output (Y): defect detection results for the battery batch, defect probability (Prob), main defect types (e.g., short circuit, micro-short circuit, lithium deposition, uneven thickness, leakage, etc.), defect grade (Grade), and defect severity (e.g., minor, moderate, severe);

[0086] (c) Model training, inputting process parameters into the input layer of the BP network;

[0087] (d) Model Application: The trained model is fed with the production process parameters for a new, unknown batch of lithium batteries. The model performs forward computations. The output layer provides defect predictions (probability, classification, and grading) for the batch.

[0088] Mathematical model expression: Input layer: Receives input vector as: ; Hidden layer: The net input to the hidden layer neurons is: ; , arrive ; in, The input layer neurons to the hidden layer The connection weights of neurons; The hidden layer The bias of the neuron The output of the hidden layer neurons , ,in is the activation function; For the receives the input vector, i.e. the first the values ​​of process parameters; There are 8 process parameters, including coating speed, roller pressure, roller speed, welding current, oven temperature, slurry viscosity / solid content, winding / stacking tension, ambient temperature and humidity, and injection volume / injection rate; Output layer: The net input to the output layer neurons ; , arrive , is the number of neurons in the hidden layer; in, The hidden layer neurons to the output layer The connection weights of neurons; The output layer The bias of each neuron; For the The final output of the output layer neurons; ,in is the activation function of the output layer; The back propagation and weight update: The core is to calculate the gradient of the loss function L with respect to all weights W and bias b ( ); use the chain rule to calculate backward from the output layer to the input layer, and then use gradient descent or its variant to update the parameters: The loss function L is: measure the network output and the true value the gap; The weight W and bias b are: ; Among them, η is the learning rate, which controls the update step size; The new weight calculated for this time; is the old weight from last time; The new bias calculated for this time; The old offset from last time.

[0089] There are many steps in the production of lithium batteries. Common key process parameters that have a significant impact on defects may include: (a) Coating process. Coating speed affects the uniformity and thickness of the wet coating. Coating gap / thickness directly determines the thickness of the active material coating. Oven temperature (multiple temperature zones) affects the solvent evaporation rate and determines the coating pore structure and binder distribution (sometimes the average or key temperature zone value is used as input). Slurry viscosity / solids content affects coating rheology and coating quality.

[0090] (b) Rolling process. Rolling pressure determines the final thickness, density, and porosity of the electrode sheet. Rolling speed, in combination with pressure, influences the compaction effect. The gap between the rollers directly controls the final thickness.

[0091] (c) Other processes. Winding / stacking tension affects the flatness and alignment of the electrode and separator, and is associated with short circuits and deformation. Injection volume / injection rate affects the degree of electrolyte infiltration. Formation / aging conditions (such as temperature and time) affect the quality of SEI film formation. (Sometimes included). Ambient temperature and humidity (critical process): Affect process stability and material properties.

[0092] Example 4: Comparison of implementation effects Through Example 1 and Example 2, a lithium-ion battery intelligent manufacturing system was created, and a dedicated data environment was built. The implementation results are compared as follows: (1) Optimizing the economic effects of production Based on X-ray images and production condition big data, the lithium-ion battery intelligent manufacturing system can provide real-time production warnings and optimal parameters to the winding process, identify the winding machine number, and adjust the winding machine's correction mechanism. The proportion of production abnormalities has been reduced from the original 0.1-0.2% to 0.01%.

[0093] Actual operational data shows that after intelligent transformation, a production line with an annual output of 100 million 26,700-5,000mAh lithium batteries can increase annual output by 10 million units, generating an additional 50 million yuan in profits. The defect rate has also been reduced from 0.2% to 0.01%, resulting in an additional 2 million good products, direct economic benefits of 20 million yuan, and total economic benefits of 70 million yuan, significantly enhancing the company's core competitiveness and sustainable development capabilities. In terms of social benefits, the implementation of the intelligent production project for new energy vehicle power batteries has played a positive role in guiding and demonstrating the improvement of the intelligence level of industry enterprises, enhancing product quality and stability, and improving production efficiency and operating benefits, actively promoting the healthy and sustainable development of the domestic battery industry.

[0094] (2) Examples of process optimization effects An example of optimized winding process operation: X-ray inspection revealed poor coverage of the positive and negative electrodes in the battery cell, resulting in a defective cell. The BP model attributed the anomaly to fluctuations in the winding machine's tension (±5%). The winding machine's deviation-correcting mechanism was automatically adjusted, and the neural network weights were simultaneously updated. After 30 days of continuous production, the defect rate for this process dropped from 0.13% to 0.008%.

[0095] Welding process optimization example: X-ray inspection of tab solder joint defects (abnormal grayscale values); the BP model attributed the abnormality to insufficient welding current (10% below the set value); the welding power was adjusted to 1200W, and the number of points was increased by 2; the solder joint defect rate decreased by 85%.

[0096] (3) Comparison of production line transformation effects .

Claims

1. A lithium-ion battery production early warning and optimization method based on X-ray detection and neural network, characterized in that: The method steps are as follows: (1) By extracting internal defect features from lithium battery X-ray images in real time, we can obtain X-ray imaging big data of lithium battery defects; obtain the operating data and X-ray detection images of the lithium battery cell production process, construct lithium battery production big data, extract features from the lithium battery defective product production big data, and determine the main equipment and process parameters that cause cell defects; (2) Extracting big data of key production process parameters of lithium batteries, constructing a BP neural network model that reflects the process-defect mapping relationship of lithium batteries; constructing an image-defect convolutional neural network, using the YOLO algorithm to identify X-ray images, extracting six types of feature classifications and grades: electrode alignment, electrode wrinkle area, number of electrodes, battery cell shell size, negative electrode redundancy, and welding defects; (3) Update lithium battery production big data; use genetic algorithm (GA) to optimize the BP neural network process of lithium battery process-defect mapping relationship, and perform model parameter optimization, training, and learning; Generate a production process early warning model to predict and alarm abnormalities in lithium battery production; (4) Set the lithium battery production objective function; use the NSGA-II algorithm on the trained BP neural network to generate a lithium battery production optimization model, optimize the lithium battery production warning and process; output monitoring warning and production process optimal parameters; automatically or with the assistance of the operator through the control system to make dynamic adjustments, forming a "prediction-warning-intervention-optimization" closed-loop control; (5) Improve the lithium battery production optimization process by coupling the non-dominated sorting genetic algorithm NSGA-II with the lithium battery process-defect BP neural coupling, optimize the multi-objective parameters, generate a lithium battery production optimization model, and use the NSGA-BP model to calculate the optimal process parameters as the process setting values ​​of the lithium battery production equipment. Each device can adjust the process by itself to achieve optimal production.

2. The lithium-ion battery production early warning and optimization method based on X-ray detection and neural network according to claim 1, characterized in that: The core goal of the BP neural network model of the lithium battery process-defect mapping relationship is to use back propagation and weight update to predict or diagnose defects that may occur in the lithium battery production process; By learning historical production data, including process parameters as input and corresponding defect detection results as output, a complex nonlinear mapping relationship is established; The BP neural network model of the lithium battery process-defect mapping relationship adopts a typical single hidden layer BP neural network, which is divided into an input layer, a hidden layer, and an output layer. Their mathematical expressions are as follows: Input layer: Receives input vector as: ; Hidden layer: The net input to the hidden layer neurons is: ; , arrive ; in, The input layer neurons to the hidden layer The connection weights of neurons; The hidden layer The bias of the neuron The output of the hidden layer neurons , , in is the activation function; For the receives the input vector, i.e. the first The value of a process parameter; There are 8 process parameters, including coating speed, roller pressure, roller speed, welding current, oven temperature, slurry viscosity / solid content, winding / stacking tension, ambient temperature and humidity, and injection volume / injection rate; Output layer: The net input to the output layer neurons ; , arrive , is the number of neurons in the hidden layer; in, The hidden layer neurons to the output layer The connection weights of neurons; The output layer The bias of each neuron; For the The final output of the output layer neurons; ,in is the activation function of the output layer; The back propagation and weight update: The core is to calculate the gradient of the loss function L for all weights W and bias b ( ); Use the chain rule to calculate backward from the output layer to the input layer, and then use gradient descent or its variant to update the parameters: The loss function L is: measure the network output and the true value the gap; The weight W and bias b are: ； Where: η is the learning rate, which controls the update step size; The new weight calculated for this time; is the old weight from last time; The new bias calculated for this time; The old offset from last time.

3. The lithium-ion battery production early warning and optimization method based on X-ray detection and neural network according to claim 1, characterized in that: The optimization process of the lithium battery process-defect BP neural network using the genetic algorithm GA is as follows: (1) Data collection and preprocessing Collect data from battery cell X-ray image detection equipment, production equipment sensors, and process parameter recording systems through lithium battery X-ray sources / imagers; Clean, normalize, and extract features from the data to build a large dataset of production process parameters and defect characteristics; (2) BP model construction and training Using BP neural network and decision tree machine learning algorithms, we build a process parameter-defect optimization model. We train the model with historical data, optimize the network structure and parameters, and improve the model's prediction accuracy and generalization ability. (3) Process parameter optimization and control Based on the BP neural network and the output of the process parameter optimization model, real-time warning and dynamic adjustment of the set values ​​of the process parameters of production equipment; Ensure the stability and consistency of the production process through real-time feedback mechanism; The BP model is not only used for offline analysis or post-batch testing, but is also deeply integrated into the production line's MES / SCADA system to achieve real-time defect probability prediction at the millisecond level. When the defect probability is predicted to exceed the standard, the model can generate specific and executable process parameter adjustment suggestions and dynamically adjust them automatically or with the assistance of the operator through the control system, forming a closed-loop control of "prediction-warning-intervention-optimization".

4. The lithium-ion battery production early warning and optimization method based on X-ray detection and neural network according to claim 1, characterized in that: The image-defect convolutional neural network is constructed, and the YOLO algorithm is used to identify X-ray images to extract the feature classification and grade of the electrode alignment, electrode wrinkle area, number of electrodes, battery cell shell size, negative electrode redundancy, and welding defects; Add a classifier based on the existing one, and then retrain and test the new category samples to further optimize the classifier and improve the classification performance of the classifier; Construct defect-related feature engineering and data sets, introduce new defect characterization data, and introduce monitoring data sources that can more directly / deeply reflect defects as model input in addition to conventional process parameters, including online / offline detection data fusion. The X-ray imaging features of electrodes and battery cells are integrated into the BP model through feature extraction or fusion methods, that is, CNN processing images + BP processing process parameters, to construct a multimodal defect prediction model; based on process mechanism and data analysis, construct derivative features with physical significance or strong predictive power, including coating drying rate gradient, roller compaction density change rate, and interaction terms between adjacent process parameters, as model input, which makes up for the shortcomings of existing process parameters.

5. The lithium-ion battery production early warning and optimization method based on X-ray detection and neural network according to claim 1, characterized in that: The BP neural network model of the lithium battery process-defect mapping relationship is divided into an input layer, a hidden layer and an output layer; the input layer includes coating speed, roller pressure, roller speed, oven temperature, slurry viscosity / solid content, winding / lamination tension, ambient temperature and humidity, injection volume / injection rate process parameters; the hidden layer includes a three-layer network of convolutional layer, pooling layer and fully connected layer, and the activation function is ReLU; the output layer includes the total probability of output defects, the probability distribution of output defect types and the probability distribution of output defect levels; The neural network model takes the main production factors that lead to internal defects of lithium batteries as input and the internal defect characteristics of lithium batteries as output, and establishes a forward multi-layer BP neural network model of lithium battery production process-defects; based on the BP neural network algorithm and samples, multiple learnings are performed to calculate the weights and thresholds of each layer, thereby solving the calculation model between input and output; the simulation error of the model is less than 1%, that is, it has good memory ability. Using the established model, untrained production conditions are calculated to calculate the most likely production factors that cause battery defects, and the error is required to be within 1%.

6. The lithium-ion battery production early warning and optimization method based on X-ray detection and neural network according to claim 1, characterized in that: The defect characteristics include at least one of poor alignment of the pole piece, missing pole piece, wrinkled pole piece, poor size of the battery cell into the shell, redundancy of the negative electrode, welding defects, and other internal defects; The defect characteristics include at least one of poor alignment of the pole piece, missing pole piece, wrinkled pole piece, poor size of the battery cell into the shell, redundancy of the negative electrode, welding defects, and other internal defects, including impurities, short circuits, and fractures; The main production equipment, process parameters and control measures of the lithium battery include at least one of the following: The slurry quality, coating uniformity, and roller thickness in the electrode preparation section are key control points. The operating conditions of the mixer, coater, and roller press directly affect the defects of lithium batteries. The electrode defects can be adjusted by adjusting the mixing time, drying temperature, coating speed, nozzle working conditions, and roller pressing conditions. The key control points in the cell assembly process are the electrode cutting accuracy, electrode winding / stacking quality, cell assembly quality, and welding quality. The working conditions of the mold laser cutter, winding machine, stacking machine, and welding machine are the equipment that directly affects the defects of lithium batteries. The process adjustment measures of the winding machine and the stacking machine include: unwinding correction, stroke correction, sheet feeding correction, infrared positioning, and tension control; the stacking machine adjusts the tightness of the stacked battery cells through the temperature and pressure of the hot pressing process; the welding power and welding current adjust the welding quality.

7. The lithium-ion battery production early warning and optimization method based on X-ray detection and neural network according to claim 1, characterized in that: The lithium battery production optimization model, through mathematical modeling, algorithm design and data analysis, maximizes production efficiency, reduces costs, improves yield, or optimizes other key indicators while meeting quality, safety, and environmental constraints. Its construction method and mathematical expression are as follows: (1) Objective function, optimization goal: Minimize costs, including raw materials, energy consumption, equipment depreciation, and labor; Maximize production efficiency, including output per unit time and equipment utilization; Optimal quality, improved yield, and consistency control; Minimize energy consumption and carbon emissions to meet green manufacturing requirements; Timely order delivery and reduced delay penalties; (2) Decision variables, production scheduling, equipment parameter setting, resource allocation, and inventory control; (3) Constraints, including physical constraints and process constraints; Physical constraints, equipment capacity limits, and process parameter ranges; Quality constraints, key parameter tolerances, and electrochemical performance requirements; Process constraints, process sequence, cleaning time; Resource constraints, material supply, equipment maintenance time, and personnel scheduling; Environmental safety, solvent volatilization limits, and wastewater treatment capabilities; (4) Key optimization areas include electrode manufacturing, coating uniformity control, and drying energy optimization; assembly process, winding / stacking accuracy, and vacuum injection efficiency; formation / capacity separation, and charge / discharge strategy optimization (time / energy balance); supply chain collaboration, and matching raw material procurement with production rhythm. Taking defect risk minimization as the objective function, the process parameter combination is iteratively optimized; each batch of new data triggers the update of model parameters and the learning rate is dynamically adjusted.

8. The lithium-ion battery production early warning and optimization method based on X-ray detection and neural network according to claim 1, characterized in that: The process parameter-defect optimization model, Introducing new defect characterization data: In addition to conventional process parameters, new monitoring data sources that can more directly / deeper reflect defects are introduced as model inputs; This includes online / offline inspection data fusion, integrating the X-ray imaging features of pole pieces and battery cells into the BP model through innovative feature extraction or fusion methods to build a multi-modal defect prediction model; Advanced feature construction: Based on process mechanisms and data analysis, derived features with physical significance or strong predictive power are constructed, including coating drying rate gradients, roller compaction density change rates, and interaction terms between adjacent process parameters as model inputs. High-quality annotated dataset construction methods: Efficient, accurate, and low-cost defect sample annotating methods, including semi-supervised learning-based utilization of large amounts of unlabeled data and innovative online inspection and offline disassembly verification linkage mechanisms. After training, the lithium battery process-defect BP model is output to generate a production process early warning model; the lithium battery production optimization process is coupled with the improved non-dominated sorting genetic algorithm NSGA-II and the lithium battery process-defect BP neural coupling to optimize multi-objective parameters and generate a lithium battery production optimization model.

9. The lithium-ion battery production early warning and optimization method based on X-ray detection and neural network according to claim 2, characterized in that: The BP neural network model of the lithium battery process-defect mapping relationship includes input, output, model training and model application; The input, key process parameters, include coating speed, roller pressure, roller speed, oven temperature, slurry viscosity / solid content, winding / lamination tension, ambient temperature and humidity, injection volume / injection rate; The output includes the defect detection results of the battery batch, defect probability; main defect types, including short circuit, micro short circuit, lithium plating, uneven thickness or leakage; defect classification; and defect severity, including minor, moderate or severe. The model training includes inputting the process parameters into the input layer of the BP network to enter the model training; The model is applied by inputting the production process parameters of a new, unknown batch of lithium batteries into the trained model; the model performs forward calculations; and the output layer provides defect prediction results for the batch, including probability, classification, or grading.

Images