A method and system for optimizing data transmission throughout the entire lithium battery separator production process.

Abstract

This invention discloses a method and system for optimizing data transmission throughout the entire lithium battery separator production process, specifically relating to the field of data transmission technology. It involves collecting sensor data from the entire lithium battery separator production process using a sensor array, extracting actual process parameters from the sensor data, comparing these actual process parameters with preset process parameters, and outputting abnormal data. Based on the spatiotemporal topology of the sensors, it calculates the spatial-temporal-functional coupling weights between abnormal data, constructs a structural coupling matrix, identifies error cluster candidates, and extracts their spatiotemporal distribution characteristics and abnormal combination patterns. A preset rule engine is used to structurally classify the error cluster candidates and output a deviation report. Based on the deviation report, it outputs sensor calibration instructions and process fine-tuning instructions, solving the problems of difficult-to-accurate identification of deviation sources, untimely transmission of key abnormal data, and high false alarm rates.

Description

Technical Field

[0001] This invention relates to the field of data transmission technology, and more specifically, to a method and system for optimizing data transmission throughout the entire production process of lithium battery separators. Background Technology

[0002] In the manufacturing process of lithium battery separators, to ensure the thickness uniformity and process stability of the final product, continuous monitoring and precise control of key parameters in multiple stages such as drying, stretching, and rolling are required. To this end, production systems typically deploy numerous sensors of various types, including temperature, pressure, tension, and thickness sensors, to collect real-time data from each stage. With the advancement of intelligent manufacturing, more and more edge computing nodes are being used to process local data, supporting local early warning, adaptive process adjustments, and quality traceability. However, against the backdrop of a surge in sensor data volume, accurately identifying the sources of deviations in the production process has become a critical factor affecting product quality and production line efficiency.

[0003] Traditional methods commonly employ residual analysis and single-point trend monitoring to detect deviations in process parameters. These methods are effective in handling single anomalies under stable operating conditions. However, when faced with the spatial coupling, temporal drift, and functional linkage characteristics of multi-source data in complex operating conditions, traditional methods often fail to identify systematic deviations, especially in effectively distinguishing whether quality anomalies are caused by sensor drift, process disturbances, or recipe deviations. Furthermore, existing technologies lack priority judgment and dynamic adaptation strategies for the bandwidth scheduling mechanism of anomaly data transmission, which may lead to delays or loss of critical data under network-limited conditions, further reducing the timeliness and accuracy of anomaly response. Summary of the Invention

[0004] To overcome the aforementioned deficiencies in the prior art, this invention provides a method and system for optimizing data transmission throughout the entire lithium battery separator production process. By constructing a spatiotemporal-functional coupling model of sensor data, introducing an anomaly scoring-driven bandwidth scheduling mechanism, and an error clustering identification and classification strategy, this invention addresses the problems mentioned in the background art, such as the difficulty in accurately identifying deviation sources, untimely transmission of key abnormal data, and high false alarm rates.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a method for optimizing data transmission throughout the entire production process of lithium battery separators, comprising the following steps:

[0006] Step 1: Collect sensor data for the entire lithium battery separator process through a sensor array and transmit it to the local edge database through edge nodes; the production data includes at least the sensor data corresponding to the drying process and the stretching process. By setting temperature sensors and pressure sensors on the edge gateways of each thermal control zone and stretching zone respectively, output the temperature-time curve of the thermal control zone and the pressure-time curve of the stretching zone with confidence level labels.

[0007] Step 2: Extract sensor data to obtain actual process parameters, compare the actual process parameters with preset process parameters, and output abnormal data;

[0008] Step 3: Deviation Source Identification: Receive abnormal data, calculate the spatial-temporal-functional coupling weights between abnormal data based on the spatiotemporal topology of the sensor, and construct a structural coupling matrix to reflect the correlation between sensor anomalies; based on the structural coupling matrix, identify error cluster candidates, extract their spatiotemporal distribution characteristics and anomaly combination patterns; use a preset rule engine to perform structural classification of error cluster candidates and output a deviation report.

[0009] Step 4: Based on the deviation report, output sensor calibration instructions and process fine-tuning instructions, execute the sensor calibration instructions and process fine-tuning instructions, and monitor the calibration results and process adjustment results.

[0010] Preferably, when the cumulative amount or frequency of abnormal data exceeds the limit, an alarm message is sent out. Based on the abnormal score of the sensor data and the current network bandwidth, the network bandwidth is matched for the sensor data to ensure the transmission priority of abnormal sensor data. Sensor data with higher abnormal score values ​​are transmitted first.

[0011] Preferably, the quantification method of the anomaly score is as follows: the anomaly score is calculated by comprehensively considering the degree of deviation between the actual process parameters and the preset process parameters based on sensor data, the duration of the deviation, and the degree of impact of the deviation on production quality.

[0012] Preferably, the sensor deviation source identification process includes the following steps:

[0013] Step S101: Data preprocessing and feature extraction: After receiving abnormal data from each sensor data channel, input a graph neural network model or clustering algorithm based on the spatiotemporal topology of the sensor. Calculate the spatial-temporal-functional coupling weights between abnormal data, and perform calculations based on time synchronization, spatial proximity, and functional category similarity to form a structural coupling matrix. Based on the structural coupling matrix, select error cluster candidates.

[0014] Step S102: Analyze error clustering candidates, extract the spatiotemporal distribution features and abnormal combination patterns of error clustering candidates, perform structural classification of error clustering candidates based on rule engine and machine learning classification model, and output error labels;

[0015] Step S103: Output a deviation report containing the deviation type and coverage. The deviation type includes process-induced error, formulation-induced error, and sensor-induced error. The coverage refers to the time range in which the deviation occurred, the process section and sensor index involved, and the degree of impact of the thickness deviation.

[0016] Preferably, if the deviation label indicates that the error is caused by the process, the process adjustment is delayed, sensor data is continuously collected, and steps S101 to S103 are repeated for confirmation to avoid erroneous feedback from disturbing the process.

[0017] If the deviation label indicates that the error is caused by the sensor, a calibration command is sent to the edge node where the corresponding sensor is located to adjust the corresponding sensor gain or bias calibration coefficient.

[0018] If the process causes errors, a process fine-tuning command is sent to the edge node where the corresponding sensor is located.

[0019] Preferably, the sensor calibration command refers to adjusting the gain / bias calibration coefficient of the sensor within the abnormal section; the process fine-tuning command refers to adjusting the process parameters within the abnormal section. The method for executing the sensor calibration command is as follows:

[0020] If the historical drift amplitude of the sensor has been consistently exceeding the threshold for a long period of time or the residual fluctuation before calibration is abnormally drastic, it is judged that the sensor performance has deteriorated or failed, and the sensor component is prompted to be replaced. If the replacement conditions are not met, the calibration process continues.

[0021] Send a sensor calibration command to the edge node where the corresponding sensor is located, and execute the sensor calibration command, which includes the calibration target value and calibration coefficient adjustment strategy.

[0022] Depending on the sensor type, either gain calibration or bias calibration is performed. Gain calibration refers to correcting the slope coefficient of the sensor measurement output through linear regression, while bias calibration refers to compensating for the reference offset of the measured value.

[0023] The sensor output within a certain time window before and after calibration is compared with the preset standard parameters, and the calibration error residual is calculated.

[0024] When the calibration error is below the set threshold, the calibration is considered successful, and the calibration parameters and their effects for this round are recorded.

[0025] Preferably, the method for executing process fine-tuning instructions is as follows:

[0026] Step S201: Based on the identified error cluster data, a temperature-pressure-thickness coupled model with an adaptive weight adjustment mechanism is constructed using time-aligned temperature, pressure and thickness measurement data. An ensemble learning algorithm is used to fuse ridge regression and decision tree regression to build corresponding sub-models for different production formula scenarios. The model is dynamically switched under different production formula scenarios to improve the prediction accuracy under nonlinear working conditions.

[0027] Step S202: Introduce an online residual monitoring mechanism based on a sliding window to evaluate the residual between the predicted thickness and the actual thickness data in real time. When the residual continuously exceeds the threshold range or shows a deviation trend, trigger anomaly localization reasoning to automatically locate the variable causing the residual and its corresponding work section.

[0028] Step S203: Generate process fine-tuning instructions based on the importance ranking and deviation direction of the variables, according to the principle of minimum disturbance.

[0029] Step S204: After the correction command is issued, the execution segment data is collected in real time and fed back to the temperature-pressure-thickness coupled model to update the model parameters, realize the dynamic convergence of adaptive working condition evolution, and fine-tune the process parameters based on the predicted thickness.

[0030] Step S205: Update the logs and the results of each round of process fine-tuning instructions, thickness improvement range, and residual fluctuation changes, and archive them in the edge database for subsequent quality backtracking and process stability assessment to improve the system's steady-state operation capability.

[0031] Preferably, the method further includes: Step 5: Collect historical production data covering temperature in the temperature control zone, pressure in the stretching zone, and thickness data, covering different production formulas and working conditions; analyze the historical production data; use the ridge regression equation and combine cross-validation to optimize the regression coefficients and establish an initial thickness prediction model; select a drift-free validation dataset to test the prediction performance; trigger the thickness prediction model update based on the abnormal data accumulation threshold; test the mean square error of the updated model under actual working conditions; if the target is met, the new model is activated; otherwise, it reverts to the historical version and triggers an alarm.

[0032] Preferably, the thickness prediction model is constructed as follows:

[0033] Step S301: Collect historical production data, including temperature data of the thermal control zone, pressure data of the stretching zone, and actual thickness data of the downstream thickness gauge;

[0034] Step S302: Correct the timing offset between the sensors in the thermal control zone and the tensile zone to ensure synchronization of temperature, pressure and thickness data; remove random noise by moving average or low-pass filtering; correct sensor drift based on historical calibration records; use the 3σ criterion or median absolute deviation to remove some abnormal data and reduce data deviation caused by drift or failure; normalize the temperature, pressure and thickness data to the 0 to 1 range to eliminate dimensional differences.

[0035] Step S303: Train the ridge regression equation using ridge regression and optimize the regression coefficients using cross-validation; train until the loss function meets the requirements. The loss function is the mean square error between the actual thickness data and the predicted thickness.

[0036] Step S304: After training is completed, select a set of known drift-free validation datasets to test the prediction performance of the initial thickness prediction model; deploy the application after successful validation.

[0037] Preferably, the thickness prediction model is updated based on the accumulation of abnormal data, and the updated model is output; the update method of the thickness prediction model is as follows:

[0038] When abnormal data accumulates to a threshold, the thickness prediction model is updated.

[0039] Abnormal data is aggregated by production batch and preprocessed, including at least timestamp alignment, noise reduction and drift correction, and normalization.

[0040] Receive normalized outlier data, perform ridge regression, output updated thickness prediction model coefficients, reduce the influence of historical data, and use weighted regression to assign higher weights to recent outlier data based on the exponential decay function.

[0041] The residuals of the updated model are compared with those of the historical model. If the mean square error of the updated model decreases, the historical model is replaced. The validated model is then stored in the edge database according to the timestamp, production formula, and sensor calibration status.

[0042] To achieve the above objectives, the present invention provides the following technical solution: a data transmission optimization system for the entire production process of lithium battery separators, comprising:

[0043] The sensor data acquisition module collects sensor data throughout the entire process of lithium battery separator production via a sensor array and transmits it to the local edge database through edge nodes.

[0044] The abnormal data filtering module extracts sensor data to obtain actual process parameters, compares the actual process parameters with preset process parameters, and outputs abnormal data.

[0045] The deviation source identification module receives abnormal data, calculates the spatial-temporal-functional coupling weights between abnormal data based on the spatiotemporal topology of the sensor, constructs a structural coupling matrix, identifies error cluster candidates based on the structural coupling matrix, extracts their spatiotemporal distribution features and abnormal combination patterns, and performs structural classification of error cluster candidates using a preset rule engine, outputting a deviation report.

[0046] The instruction execution module outputs sensor calibration instructions and process fine-tuning instructions based on deviation reports, executes these instructions, and monitors the calibration and adjustment results. The process fine-tuning instructions are executed by: constructing a coupled model integrating temperature, pressure, and thickness data; using an ensemble learning algorithm to achieve dynamic switching prediction under multiple production formulation scenarios; introducing an online residual monitoring mechanism based on a sliding window to achieve real-time diagnosis and anomaly localization of predicted deviations; and generating minimum disturbance process fine-tuning instructions based on the importance of key variables and the direction of deviation.

[0047] The technical effects and advantages of this invention are as follows:

[0048] (1) The lithium battery separator full-process production data transmission optimization method provided by the present invention introduces the spatiotemporal topology and structural coupling matrix of the sensor, and combines graph neural network, rule engine and machine learning model to realize the cluster identification and fine classification of sensor anomalies, and automatically generate response instructions according to the deviation type; it not only improves the accuracy and interpretation of anomaly source identification, but also effectively distinguishes different types of errors such as sensor drift, process deviation and formula anomaly, and reduces the risk of misjudgment and misoperation; it supports scoring of abnormal data and dynamically adjusting its transmission priority to ensure rapid response and reporting of high-risk anomalies.

[0049] (2) The lithium battery separator full-process production data transmission optimization method provided by this invention combines a thickness prediction model, a process correction algorithm, and a sensor adaptive calibration mechanism to establish a feedback closed-loop control system based on the temperature-pressure-thickness coupling relationship. By dynamically constructing sub-models under the formulation scenario through ensemble learning methods and integrating online residual monitoring and variable importance analysis, real-time fine-tuning of process parameters is achieved with minimal disturbance while ensuring product quality, effectively avoiding the impact of drastic process fluctuations on separator quality. Attached Figure Description

[0050] Figure 1 This is a flowchart showing the optimized data transmission process for the entire production of lithium battery separators according to the present invention.

[0051] Figure 2 The flowchart for constructing the thickness prediction model of this invention is shown below.

[0052] Figure 3 This is a block diagram of the data transmission optimization system for the entire production process of lithium battery separators according to the present invention. Detailed Implementation

[0053] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0054] At the same time, it should be understood that, for ease of description, the dimensions of the various parts shown in the accompanying drawings are not drawn according to actual scale.

[0055] The following description of at least one exemplary embodiment is merely illustrative and is in no way intended to limit the scope of this application and its application or use.

[0056] Techniques, methods, and equipment known to those skilled in the art may not be discussed in detail, but where appropriate, such techniques, methods, and equipment should be considered part of the specification.

[0057] Example 1, see Figure 1 This invention provides an optimized data transmission flow chart for the entire production process of lithium battery separators. Figure 1 The method for optimizing data transmission throughout the entire production process of lithium battery separators, as shown, includes the following steps:

[0058] Step 1: Collect sensor data for the entire lithium battery separator process using a sensor array and transmit it to the local edge database via edge nodes. The production data includes sensor data corresponding to the drying and stretching processes. Temperature and pressure sensors are set on the edge gateways of each thermal control zone and stretching zone, respectively, to output temperature-time curves for the thermal control zone and pressure-time curves for the stretching zone with confidence level labels. The preprocessing method in Step 1 is as follows: timestamp the sensor data; and use a dynamic time warping algorithm to correct the time sequence offset between sensors in real time.

[0059] Step 2: Extract sensor data to obtain actual process parameters, compare the actual process parameters with preset process parameters, and output abnormal data;

[0060] Step 3: Deviation Source Identification: Receive anomalous data; based on the spatiotemporal topology of the sensor, calculate the spatial-temporal-functional coupling weights between anomalous data. These weights are calculated based on temporal synchronization, spatial proximity, and functional category similarity. Construct a structural coupling matrix to reflect the correlation between sensor anomalies. Based on this structural coupling matrix, identify error cluster candidates and extract their spatiotemporal distribution characteristics and anomaly combination patterns. Use a preset rule engine to perform structural classification on the error cluster candidates and output a deviation report.

[0061] Step 4: Based on the deviation report, output sensor calibration instructions and process fine-tuning instructions, execute the sensor calibration instructions and process fine-tuning instructions, and monitor the calibration results and process adjustment results.

[0062] In one possible embodiment, when the cumulative amount or frequency of abnormal data exceeds the limit, an alarm message is issued. Based on the anomaly score of the sensor data and the current network bandwidth, network bandwidth is matched to the sensor data to ensure the transmission priority of abnormal sensor data. Sensor data with higher anomaly scores are transmitted first. Specific implementation methods include:

[0063] To avoid bandwidth constraints during critical periods and potential delays or loss of important data, temperature data at edge nodes is compressed at three levels: "abnormal data priority, drift data secondary, and normal data lowest compression" with different compression ratios (e.g., lossless compression for abnormal data, slight loss for drift data, and high compression for normal data), significantly reducing bandwidth consumption. Only high-priority data within a ±30s window is transmitted in real-time at the second level, while non-abnormal data is accumulated at specified time intervals (e.g., every 5 minutes) before being uploaded in batches, minimizing bandwidth peak conflicts. The available bandwidth of the link is periodically measured (e.g., every 10 seconds), and the results are sent to edge nodes. Based on this, the nodes automatically adjust the number of parallel transmission threads and the fragment size to ensure that critical data can still be transmitted with priority even when bandwidth is limited.

[0064] To avoid communication reliability issues caused by single path dependence, two or more links are connected to the edge node simultaneously, and a health detection probe is configured for each link. When the latency or packet loss rate of the main link exceeds a preset threshold, the node automatically switches high-priority data to an available backup link and reroutes it. During this process, the switching event is logged and alarms are generated. Sensor data with the highest anomaly score can be sent in parallel on both the main and backup channels. The results of the two channels are compared by CRC check. If they match, the reception is confirmed. If one channel fails, the other channel can still ensure data integrity.

[0065] In one possible embodiment, the anomaly score is quantified by comprehensively calculating the anomaly score based on the degree of deviation between the actual process parameters and preset process parameters from the sensor data, the duration of the deviation, and the degree of impact of the deviation on production quality. Specifically, for each sensor data point, the residual between its actual value and preset value is calculated. The residual is defined as the ratio of the difference between the actual value and the preset value to the preset value, to normalize the magnitude of the deviation. A sliding window (e.g., a window size of 5 minutes) is used to statistically analyze the cumulative mean and standard deviation of the residual to reflect the persistence and volatility of the deviation. In addition, combined with process knowledge, different sensor data are assigned weights to their impact on the uniformity of lithium battery separator thickness. For example, the weight of the temperature sensor in the thermal control zone is higher than that of the pressure sensor in the stretching zone. The weight values ​​are determined based on historical production data and regression analysis of thickness deviations. The anomaly score is obtained by weighted summation. The weight coefficients can be dynamically adjusted according to the production formula and operating conditions. The higher the anomaly score, the greater the impact of the anomaly of the sensor data on production quality, and the higher the priority.

[0066] Background: In lithium battery separator production, identifying the source of deviation is a key step in ensuring thickness uniformity and process stability. Traditional deviation identification methods (such as residual analysis) can locate sensor drift or process deviations, but they may misjudge when faced with transient disturbances, leading to false alarms or production interruptions.

[0067] In this embodiment of the invention, it needs to be further explained that the sensor deviation source identification process includes the following steps:

[0068] Step S101: Data preprocessing and feature extraction: After receiving abnormal data from each sensor data channel, input a graph neural network model or clustering algorithm based on the spatiotemporal topology of the sensor. Calculate the spatial-temporal-functional coupling weights between abnormal data, and perform calculations based on time synchronization, spatial proximity, and functional category similarity to form a structural coupling matrix. Based on the structural coupling matrix, select error cluster candidates.

[0069] The spatiotemporal topology of a sensor refers to the network structure formed based on the physical location of the sensor in the lithium battery separator production process (spatial topology, such as the distribution of the thermal control zone and stretching zone) and the time series of data acquisition (temporal topology). By using graph neural networks or clustering algorithms, combined with time synchronization, spatial proximity and functional category similarity, the coupling relationship between abnormal data is calculated to identify the source of deviation.

[0070] The explanation explains that the spatial-temporal-functional coupling weight is a weighted index used to measure the similarity or correlation of multiple sensor anomalous data in the dimensions of spatial location, temporal behavior, and functional type. It reflects whether the anomalous data is correlated, and is thus used to discover whether clustered anomalous behaviors originate from systemic process problems rather than isolated sensor anomalousness. The structural coupling matrix is ​​a matrix constructed with sensors as nodes. Each element in the matrix represents the coupling strength between any two sensors (i.e., the coupling weight mentioned above), similar to the weighted form of the node adjacency matrix in a graph neural network. It is used to identify which sensor data anomalousness has group consistency (i.e., clustering).

[0071] Step S102: Analyze error cluster candidates, extract the spatiotemporal distribution characteristics and abnormal combination patterns of error cluster candidates. Based on the rule engine and machine learning classification model, perform structural classification of error cluster candidates and output error labels. The rule engine matches the following conditions based on process knowledge and sensor configuration: whether the error cluster candidate is located in the tensile zone or thermal control zone, whether it involves multiple sensor types (such as temperature, pressure, thickness, strain), and whether it is accompanied by tension feedback delay or temperature control lag. If the conditions are met, it is marked as a process-induced error.

[0072] If only a single sensor is malfunctioning and has a high malfunction score, it is marked as a sensor causing the error.

[0073] If the abnormal score is related to the deviation of the formula parameters and has no structural coupling, it is marked as an error caused by the formula.

[0074] If the abnormal score is related to the deviation of process parameters in a specific section, it is marked as a process-induced error.

[0075] To explain, the rule engine makes conditional judgments based on process knowledge. For example, large temperature fluctuations in the thermal control zone accompanied by abnormal tension indicate process-related errors. It is used to determine whether the error has structural features. The machine learning classification model is a pre-trained classification algorithm model (such as decision tree, SVM, random forest, or graph neural network) used to learn abnormal feature patterns from clustered data. When the rule engine cannot completely cover all abnormal patterns, the machine learning classification model can make data-driven judgments. Combining spatiotemporal features, deviation trends, anomaly scores, and other inputs, it outputs error labels.

[0076] Step S103: Output a deviation report containing the deviation type and coverage. The deviation type includes process-induced error, formulation-induced error, and sensor-induced error. The coverage refers to the time range in which the deviation occurred, the process section and sensor index involved, and the degree of impact of the thickness deviation.

[0077] In one possible embodiment, if the deviation label indicates that the error is caused by the process, the process adjustment is delayed (e.g., an alarm is set after 5 minutes), sensor data is continuously collected, and steps S101 to S103 are repeated for confirmation to avoid erroneous feedback from disturbing the process.

[0078] If the deviation label indicates that the error is caused by the sensor, a calibration command is sent to the edge node where the corresponding sensor is located to adjust the corresponding sensor gain or bias calibration coefficient.

[0079] If the process causes errors, a process fine-tuning command is sent to the edge node where the corresponding sensor is located. For example, the target heating rate of the thermal control zone or stretching zone is corrected, or the equipment status is checked.

[0080] In this embodiment of the invention, it needs to be further explained that the sensor calibration command refers to adjusting the gain / bias calibration coefficient of the sensor in the abnormal section (such as the temperature sensor and the pressure sensor); the process fine-tuning command refers to adjusting the process parameters in the abnormal section, correcting the target heating rate of each section during the next start-up preheating or formula change, correcting the pressure curve during diaphragm rolling or winding, minimizing the layer thickness fluctuation, and making the thickness prediction model continuously converge with the on-site working conditions; the method of executing the sensor calibration command is as follows:

[0081] If the historical drift amplitude of the sensor has been consistently exceeding the threshold for a long period of time or the residual fluctuation before calibration is abnormally drastic, it is judged that the sensor performance has deteriorated or failed, and the sensor component is prompted to be replaced. If the replacement conditions are not met, the calibration process continues.

[0082] Send a sensor calibration command to the edge node where the corresponding sensor is located, and execute the sensor calibration command, which includes the calibration target value and calibration coefficient adjustment strategy.

[0083] Depending on the sensor type, either gain calibration or bias calibration is performed. Gain calibration refers to correcting the slope coefficient of the sensor measurement output through linear regression, while bias calibration refers to compensating for the reference offset of the measured value.

[0084] The sensor output within a certain time window before and after calibration is compared with the preset standard parameters, and the calibration error residual is calculated.

[0085] If the calibration error is below the set threshold, the calibration is considered successful, and the calibration parameters and their effects for this round are recorded; otherwise, the calibration process is repeated or a manual maintenance suggestion is triggered.

[0086] In this embodiment of the invention, it needs to be further explained that the method of executing the process fine-tuning instruction is as follows:

[0087] Step S201: Based on the identified error clustering data, a temperature-pressure-thickness coupled model with an adaptive weight adjustment mechanism is constructed using time-aligned temperature, pressure, and thickness measurement data. An ensemble learning algorithm is used to fuse ridge regression and decision tree regression to build corresponding sub-models for different production formulation scenarios. Dynamic switching is performed under different production formulation scenarios to improve prediction accuracy under nonlinear operating conditions. A thickness prediction model is built based on key factors. The temperature-time curve, pressure-time curve, and lithium battery separator thickness data from downstream laser / β-ray thickness gauges are subjected to multivariate regression analysis to output the thickness prediction model.

[0088] Step S202: Introduce an online residual monitoring mechanism based on a sliding window to evaluate the residual between the predicted thickness and the actual thickness data in real time. When the residual continues to exceed the threshold range or shows a deviation trend, trigger anomaly localization reasoning to automatically locate the variable causing the residual (for example, through principal component analysis, the key factors causing thickness anomalies include temperature and pressure) and its corresponding process section.

[0089] Step S203: Based on the importance ranking and deviation direction of the variables, generate process fine-tuning instructions under the principle of minimum disturbance, including but not limited to fine-tuning of the tension change rate in the stretching section, adjusting the heating slope in the thermal control zone, and optimizing the temperature stabilization time in the preheating zone.

[0090] Step S204: After the correction command is issued, the execution segment data is collected in real time and fed back to the temperature-pressure-thickness coupled model to update the model parameters, realize the dynamic convergence of adaptive working condition evolution, and fine-tune the process parameters based on the predicted thickness.

[0091] Step S205: Update the logs and the results of each round of process fine-tuning instructions, thickness improvement range, and residual fluctuation changes, and archive them in the edge database for subsequent quality backtracking and process stability assessment to improve the system's steady-state operation capability.

[0092] The production process of lithium battery separators involves extrusion, drying, stretching, thickness measurement, and winding. The thermal control zone controls the crystallinity of the membrane or removes solvents, the stretching zone forms a microporous structure through tension, and the thickness measurement stage ensures uniform membrane thickness. The uniformity and quality stability of the separator directly affect the safety and performance of the battery. Production involves multiple thermal control and stretching zones, and the temperature and pressure parameters in each zone need to be precisely controlled to ensure the uniformity of the separator thickness. However, in actual production, long-term operation of sensors may lead to drift, and changes in production formulas (such as the process requirements of different separator models) can cause changes in operating conditions, resulting in the lithium battery separator thickness prediction gradually deviating from the actual operating conditions, leading to thickness prediction deviations and quality fluctuations. To address the problem of the thickness prediction model being uncalibrated over a long period and unable to reflect changes in on-site operating conditions, Example 2 is provided.

[0093] Example 2, see Figure 2 The flowchart for constructing a thickness prediction model is provided. The method further includes: Step 5: Collect historical production data covering temperature in the temperature control zone, pressure in the stretching zone, and thickness data, covering different production formulas and working conditions. Analyze the historical production data, use the ridge regression equation, and optimize the regression coefficients by combining cross-validation to establish an initial thickness prediction model. Select a drift-free validation dataset to test the prediction performance. Trigger the thickness prediction model update based on the abnormal data accumulation threshold. Test the mean square error of the updated model under actual working conditions. If the target is met, the new model is activated; otherwise, it reverts to the historical version and triggers an alarm.

[0094] In this embodiment of the invention, it needs to be further explained that the thickness prediction model is constructed in the following way:

[0095] Step S301: Collect historical production data, including temperature data in the thermal control zone, pressure data in the stretching zone, and actual thickness data from the downstream thickness gauge; the collected data covers different production formulas (such as different thickness specifications) and operating conditions (such as normal operation and formula switching).

[0096] Step S302: Correct the timing offset between the sensors in the thermal control zone and the tensile zone to ensure synchronization of temperature, pressure, and thickness data; remove random noise by moving average or low-pass filtering; correct sensor drift based on historical calibration records (e.g., adjust bias or gain); use the 3σ criterion or median absolute deviation to remove some abnormal data, reducing data deviation caused by drift or fault; normalize the temperature, pressure, and thickness data to the 0 to 1 range to eliminate dimensional differences.

[0097] Step S303: Train the ridge regression equation using ridge regression (if multicollinearity exists between features, ridge regression is preferred), and optimize the regression coefficient βi using cross-validation (such as 5-fold cross-validation); train until the loss function meets the requirements. The loss function is the mean square error between the actual thickness data and the predicted thickness.

[0098] Step S304: After training is completed, select a set of known drift-free validation datasets (sensor data calibrated in the laboratory or recently maintained) to test the prediction performance of the initial thickness prediction model; deploy the application after successful validation.

[0099] To address the multicollinearity of temperature in the thermal control zone, pressure in the stretching zone, and thickness data downstream, the regularization parameter λ is dynamically adjusted, and the optimal λ value is determined through k-fold cross-validation to balance model bias and generalization ability. To adapt to changes in production formula, recent data is given higher weight based on an exponential decay function, and the thickness prediction model is updated to reduce the interference of historical data on nonlinear operating conditions.

[0100] Background: The initial thickness prediction model was trained based on historical data. However, it has not been updated for a long time and cannot adapt to equipment wear, environmental changes, or formula switching, resulting in increased thickness prediction errors. The thickness prediction model is updated based on the accumulation of abnormal data, and the updated model is output.

[0101] In this embodiment of the invention, it needs to be further explained that the update method of the thickness prediction model is as follows:

[0102] When abnormal data accumulates to a threshold, the thickness prediction model is updated.

[0103] Abnormal data is aggregated by production batch and preprocessed, including at least timestamp alignment, noise reduction and drift correction, noise removal by moving average, drift correction based on gain / bias calibration coefficient, and normalization.

[0104] Receive normalized outlier data, perform ridge regression, output updated thickness prediction model coefficients, reduce the influence of historical data, and use weighted regression to assign higher weights to recent outlier data based on the exponential decay function.

[0105] The updated model is compared with the historical model in terms of residuals (mean squared error). If the mean squared error of the updated model decreases (e.g., by 10%), the historical model is replaced. The validated models (including the initial thickness prediction model and each updated model) are stored in the edge database according to timestamps, production formulas, and sensor calibration status. This supports rollback to the historical model, ensuring convergence with the actual working conditions and reducing thickness prediction errors caused by drift or formula switching.

[0106] Example 3, see Figure 3 The structural block diagram of the lithium battery separator end-to-end production data transmission optimization system differs from Embodiments 1 and 2 in that the present invention provides: a lithium battery separator end-to-end production data transmission optimization system, comprising:

[0107] The sensor data acquisition module collects sensor data throughout the entire process of lithium battery separator production via a sensor array and transmits it to the local edge database through edge nodes.

[0108] The abnormal data filtering module extracts sensor data to obtain actual process parameters, compares the actual process parameters with preset process parameters, and outputs abnormal data.

[0109] The deviation source identification module receives abnormal data, calculates the spatial-temporal-functional coupling weights between abnormal data based on the spatiotemporal topology of the sensor, constructs a structural coupling matrix, identifies error cluster candidates based on the structural coupling matrix, extracts their spatiotemporal distribution features and abnormal combination patterns, and performs structural classification of error cluster candidates using a preset rule engine, outputting a deviation report.

[0110] The instruction execution module outputs sensor calibration instructions and process fine-tuning instructions based on the deviation report, executes the sensor calibration instructions and process fine-tuning instructions, and monitors the calibration results and process adjustment results;

[0111] The process fine-tuning instructions are executed as follows: a coupled model integrating temperature, pressure, and thickness data is constructed, and dynamic switching prediction under multiple production formula scenarios is achieved through ensemble learning algorithms; an online residual monitoring mechanism based on sliding windows is introduced to achieve real-time diagnosis and anomaly location of prediction deviations; and minimum disturbance process fine-tuning instructions are generated based on the importance of key variables and the direction of deviation.

[0112] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for optimizing data transmission in a full-process production of lithium battery separators, characterized in that, The method comprises the following steps: Collecting sensor data of the whole process of the lithium battery separator through a sensor array; Extracting actual process parameters from the sensor data, comparing the actual process parameters with preset process parameters, and outputting abnormal data; Bias source identification: receiving abnormal data, calculating the space-time-function coupling weight between abnormal data based on the space-time topology structure of the sensor, constructing a structural coupling matrix to reflect the correlation between sensor abnormalities; Based on the structural coupling matrix, identify error cluster candidates, extract their spatio-temporal distribution characteristics and abnormal combination patterns; Using a preset rule engine to classify the error cluster candidates structurally and output a bias report; Output sensor calibration instructions and process fine-tuning instructions, and monitor the calibration results and process adjustment results; The process fine-tuning instruction is executed in the following manner: a coupling model is constructed by fusing temperature, pressure and thickness data, and a dynamic switching prediction is realized in multiple production formula scenarios through an ensemble learning algorithm; an online residual monitoring mechanism based on a sliding window is introduced to realize real-time diagnosis and abnormal positioning of prediction bias; and a minimum disturbance process fine-tuning instruction is generated according to the importance and bias direction of the key variables.

2. The method of claim 1, wherein the method is characterized by: When the cumulative amount or frequency of abnormal data jointly exceeds the limit, an alarm information is sent out, the abnormal score of the sensor data is matched with the current network bandwidth based on the sensor data, the transmission priority of the abnormal sensor data is ensured, and the sensor data with a high abnormal score is transmitted preferentially.

3. The method of claim 2, wherein the method further comprises: The quantification method of the abnormal score is: based on the deviation degree, deviation duration and influence degree of the abnormal score on the production quality, the abnormal score is calculated comprehensively.

4. The method of claim 1, wherein the method is characterized by: The sensor bias source identification process comprises the following steps: Step S101: data preprocessing and feature extraction: after receiving abnormal data from each sensor data channel, input the graph neural network model based on the space-time topology structure of the sensor, calculate the space-time-function coupling weight between the abnormal data based on time synchronization, space proximity and functional category similarity, and form a structural coupling matrix; Based on the structural coupling matrix, error cluster candidates are screened; Step S102: analyze the error cluster candidates and extract the spatio-temporal distribution characteristics and abnormal combination patterns of the error cluster candidates; based on the rule engine and machine learning classification model, the error cluster candidates are classified structurally, and an error label is output; Step S103: output a bias report containing bias types and coverage range; the bias types include process-induced errors, formula-induced errors and sensor-induced errors; the coverage range refers to the time range of the bias, the involved work section and sensor index, and the influence degree of the thickness bias.

5. The method of claim 1, wherein the method is characterized by: The sensor calibration instruction refers to adjusting the gain / bias calibration coefficient of the sensor in the abnormal section; the process fine-tuning instruction refers to adjusting the process parameters in the abnormal section; and the sensor calibration instruction is executed in the following manner: If the sensor history drift amplitude is stable for a long time and exceeds the threshold value or the residual fluctuation before calibration is abnormally severe, it is judged that the sensor performance is degraded or failed, prompting to replace the sensor component, if the replacement condition is not met, the calibration process is continued; The sensor calibration instruction is issued to the edge node where the corresponding sensor is located, and the sensor calibration instruction includes a calibration target value and a calibration coefficient adjustment strategy; According to the sensor type, gain calibration or bias calibration is performed respectively, the gain calibration refers to correcting the slope coefficient of the sensor measurement output by linear regression, and the bias calibration refers to performing reference offset compensation on the measurement value; The sensor output in a certain time window before and after calibration is compared with the preset standard parameter, and the calibration error residual is calculated; When the calibration error is lower than the set threshold, it is determined that the calibration is successful, and the calibration parameters and their effects of this round are recorded.

6. The method of claim 5, wherein the method is characterized by: The execution mode of the process fine-tuning instruction is: Step S201: Based on the identified error cluster data, a temperature-pressure-thickness coupling model with an adaptive weight adjustment mechanism is constructed using time-aligned temperature, pressure and thickness measurement data, an integrated learning algorithm is used to fuse ridge regression and decision tree regression, corresponding sub-models under different production formula scenarios are built, and dynamic switching is performed under different production formula scenarios to improve the prediction accuracy under nonlinear working conditions; Step S202: An online residual monitoring mechanism based on a sliding window is introduced, the residual between the predicted thickness and the actual thickness data is evaluated in real time, when the residual continuously exceeds the threshold range or shows a deviation trend, an abnormal positioning reasoning is triggered, and the variable causing the residual and the corresponding process section are automatically located; Step S203: According to the importance ranking and deviation direction of the variable, a process fine-tuning instruction under the principle of minimum disturbance is generated; Step S204: After the correction instruction is issued, the execution section data is collected in real time and fed back to the temperature-pressure-thickness coupling model, the model parameters are updated, the dynamic convergence of adaptive working condition evolution is realized, and the process parameters are fine-tuned based on the predicted thickness; Step S205: The update log and the results of each round of process fine-tuning instruction, the thickness improvement amplitude and the residual fluctuation change are all archived in the edge database for subsequent quality backtracking and process stability evaluation, and the system steady-state operation ability is improved.

7. The method of claim 1, wherein the method is characterized by: The method further comprises: Step five: Collect historical production data covering temperature in the temperature control area, pressure in the stretching area and thickness data, cover different production formulas and working conditions, analyze the historical production data, use the ridge regression equation, combine cross-validation to optimize the regression coefficient, and establish an initial thickness prediction model; select non-drift verification data set to test the prediction performance, trigger thickness prediction model update based on abnormal data accumulation threshold, test the mean square error of the updated model under actual working conditions, if it meets the standard, use the new model, otherwise, fall back to the historical version and alarm.

8. The method of claim 7, wherein the method is characterized by: The construction method of the thickness prediction model is: Step S301: Collect historical production data, including temperature data in the temperature control area, pressure data in the stretching area and actual thickness data of the downstream thickness gauge; Step S302: Correct the timing offset between the thermal control zone and the stretching zone sensors to ensure synchronization of temperature, pressure, and thickness data; remove random noise through moving average or low-pass filtering; correct sensor drift based on historical calibration records; remove part of the abnormal data using the 3σ criterion or median absolute deviation to reduce data deviation caused by drift or failure; normalize the temperature, pressure, and thickness data to the 0-1 interval to eliminate dimensional differences; Step S303: Train the ridge regression equation using ridge regression, and optimize the regression coefficients using cross-validation; train until the loss function meets the requirements. The loss function is the mean square error of the actual thickness data and the predicted thickness; Step S304: After training is completed, select a set of known non-drift validation data sets to test the initial thickness prediction model prediction performance; after verification, deploy and apply.

9. The method of claim 8, wherein the method further comprises: Based on the accumulation of abnormal data, the thickness prediction model is updated, and the updated model is output; the update method of the thickness prediction model is: When the abnormal data accumulates to the threshold, the thickness prediction model is updated; Abnormal data is summarized by production batch and preprocessed; At least including timestamp alignment, denoising and drift correction, normalization processing; Receive normalized abnormal data, perform ridge regression, output updated thickness prediction model coefficients, reduce the influence of historical data, use weighted regression, and give higher weight to recent abnormal data based on an exponential decay function; Compare the residual error of the updated model with the historical model. If the mean square error of the updated model is reduced, replace the historical model, and store the verified model to the edge database according to the timestamp, production formula, and sensor calibration state.

10. A system for optimizing data transmission in a full-process production of a lithium battery separator, for implementing the method for optimizing data transmission in a full-process production of a lithium battery separator according to any one of claims 1 or 7, characterized in that, It includes: Sensor data acquisition module, which acquires sensor data of lithium battery separator full process through sensor array and transmits to local edge database through edge node; Abnormal data screening module, which extracts actual process parameters from sensor data and compares them with preset process parameters to output abnormal data; Bias source identification module, which receives abnormal data, calculates the space-time-function coupling weight between abnormal data based on the space-time topology of sensors, constructs a structural coupling matrix, and identifies error cluster candidates based on the structural coupling matrix, extracts their space-time distribution characteristics and abnormal combination patterns; Using a preset rule engine to classify error cluster candidates structurally and output a bias report; Instruction execution module, which outputs sensor calibration instructions and process fine-tuning instructions based on the bias report, executes sensor calibration instructions and process fine-tuning instructions, and monitors calibration results and process adjustment results; The method for executing process fine-tuning instructions is to construct a coupling model that integrates temperature, pressure, and thickness data, and to realize dynamic switching prediction in multiple production formula scenarios through ensemble learning algorithm; introduce an online residual error monitoring mechanism based on sliding window to realize real-time diagnosis and abnormal positioning of prediction bias; generate minimum disturbance process fine-tuning instructions according to the importance and bias direction of key variables.

Images