Artificial intelligence-based coater mini end-side device failure prediction method

Abstract

The present application relates to the cross technical field of artificial intelligence and industrial equipment fault prediction diagnosis, and in particular to a coating machine micro end side equipment fault prediction method based on artificial intelligence, the method being: coating machine multi-dimensional time series data acquisition and early slight marking, data division and preprocessing, lightweight LightGBM model construction and reconstruction training, end side anomaly diagnosis reasoning, and automatic labeling and early warning effectiveness performance evaluation; the present application can realize real-time and accurate prediction of the end side of the coating machine fault, greatly reduce the data labeling cost, and adapt to the micro end side equipment and the actual production scene of the coating machine.

Description

Technical Field

[0001] This invention relates to the interdisciplinary field of artificial intelligence and industrial equipment fault prediction and diagnosis, specifically to a fault prediction method for micro end-side equipment of a coating machine based on artificial intelligence. Background Technology

[0002] As a core production equipment in the manufacturing of lithium battery cells and metal foil substrates, the coating machine's operating status directly determines key quality indicators such as coating thickness uniformity and surface smoothness. During the coating machine's production process, abnormal conditions such as roller misalignment, coating viscosity fluctuations, temperature runaway, and coating speed mismatch can easily lead to product scrap. Therefore, real-time and accurate fault prediction and diagnosis of the coating machine is crucial to ensuring production stability and reducing production losses.

[0003] Currently, anomaly detection methods related to coating machine fault diagnosis are mainly divided into four categories: statistical, clustering, classification, and proximity-based. With the development of IoT, industrial big data, and artificial intelligence technologies, traditional fault detection methods have gradually revealed many shortcomings: On the one hand, cloud-based centralized computing requires uploading all multi-dimensional operating data such as coating machine roller speed, coating pressure, and oven temperature to a remote server, which not only increases network transmission pressure but also easily leads to delayed fault warnings due to transmission delays, failing to meet the real-time requirements of production; on the other hand, local single-sensor acquisition and analysis methods have limited data dimensions, making it difficult to fully reflect the complex operating status of the coating machine, resulting in low accuracy of fault detection.

[0004] Artificial intelligence technologies, represented by deep learning, have provided a new research direction for fault diagnosis of coating machines. Existing anomaly detection solutions based on deep learning mostly adopt PCA dimensionality reduction modeling combined with the LightGBM algorithm. However, such solutions still have significant shortcomings: First, the models mostly rely on cloud or high-performance edge node deployment, which is difficult to adapt to the actual needs of the compact space and limited computing resources in the coating machine production site, and cannot achieve lightweight deployment on micro-edge devices. Second, the LightGBM algorithm model has extremely high requirements for the quality of normal training data. However, in the actual production environment of coating machines, sensors are easily affected by coating dust and vibration interference, resulting in superimposed data noise. In addition, a small amount of hidden abnormal data is inevitably mixed in, causing the constructed normal operating condition model to have deviations and failing to effectively identify minor fault precursors. Third, the traditional anomaly detection result evaluation system relies on accurate manual annotation of abnormal data. However, in coating machine production, only serious anomalies that cause obvious product defects or shutdowns can be annotated. Hidden and minor anomalies are difficult to annotate accurately. Full manual annotation will significantly increase the operation and maintenance costs and cannot objectively evaluate the early warning effect of the edge diagnostic system. Summary of the Invention

[0005] To address the limitations of existing coating machine fault prediction methods, such as the inability to achieve lightweight deployment at the micro-end, insufficient fault identification accuracy, and high data annotation costs, an AI-based fault prediction method for micro-end equipment in coating machines is proposed. This method aims to achieve real-time and accurate end-side prediction of coating machine faults, significantly reduce data annotation costs, and adapt to micro-end equipment and actual production scenarios of coating machines.

[0006] The technical solution of the present invention is as follows: The method for predicting faults in micro-end-side equipment of coating machines based on artificial intelligence includes the following steps: Step 1: Relying on the Internet of Things sensor network, collect the core operating parameters of multiple nodes in the coating machine production process in real time to form multi-dimensional time series data covering normal and abnormal operating states. Only a small number of severely abnormal data are manually marked in the early stage. Step 2: Divide the collected historical multi-dimensional time series data into training set and validation set, and remove the severely anomalous data that has been labeled in the training set; use the isolated forest algorithm to perform preliminary anomaly detection on the training set to locate latent anomalous data points, and repair the detected latent anomalous data points through linear interpolation. Step 3: Construct a lightweight LightGBM model adapted for deployment on the micro end of the coating machine. Input the repaired training set data containing only normal operation data into the LightGBM model for reconstruction training, so that the model can learn the characteristic rules of the normal operation state of the coating machine. Step 4: Input the validation set data containing the unrepaired normal operation data into the trained LightGBM model, calculate the reconstruction error score of each data sample; set a reconstruction error threshold. If the reconstruction error score of a sample exceeds the reconstruction error threshold, it is determined to be abnormal data, thereby realizing real-time end-side diagnosis of abnormal operating conditions of the coating machine. Step 5: Based on the well-trained LightGBM model and the real-time collected coating machine operation data, perform automatic annotation on latent and minor anomalies, and combine it with the previous manual minor annotation to form complete annotation data; calculate the early warning effectiveness evaluation index based on the complete annotation data to evaluate the model performance.

[0007] Furthermore, in step 1, the multi-node core operating parameters include at least one of the following: temperature, vibration, speed, tension, roller speed, coating viscosity, motor current, and coating pressure of the coating machine head and oven; the acquisition interval of the multi-dimensional time series data is 10 seconds, with the coating machine head generating 5-dimensional time series data and the coating machine oven generating 15-dimensional time series data.

[0008] Furthermore, in step 2, data preprocessing also includes: aggregating historical multi-dimensional time series data of the coating machine head and oven according to the collection interval; processing multiple data points within the same time interval using mean smoothing; and filling missing data using forward interpolation. The processed data is then standardized using Z-score to eliminate differences in the dimensions and numerical ranges of different features. The Z-score standardization formula is: ; Where x is the original data value of the feature. The characteristic mean, The standard deviation is the feature value; the ratio of the training set to the validation set is 8:2.

[0009] Furthermore, in step 2, when using linear interpolation to repair latent anomalies, the outlier is filled based on the mean of the 10 normal data points before and after the anomaly. The calculation formula is as follows: ; Where, x repaired The value after anomaly repair, x t-i x is the value of the i-th normal data point before the outlier. t+i This represents the value of the i-th normal data point after the outlier.

[0010] Furthermore, in step 3, when constructing the lightweight LightGBM model, feature engineering is performed on the feature differences between the 5-dimensional operating data of the coating machine head and the 15-dimensional operating data of the oven. Feature engineering includes data standardization and feature selection. Then, the processed machine head data and oven data are fused as model input. Model training uses gradient boosting decision trees for supervised training, and the training loss function is cross-entropy, with the formula: ; Where K is the number of categories, y i,k For the one-hot encoding of the true label of the i-th sample, p i,k Let be the predicted probability that the i-th sample belongs to the k-th class, and n be the number of samples.

[0011] Furthermore, during the training of the LightGBM model, each gradient boosting decision tree has 20 leaf nodes and a learning rate of 0.005, with the model hyperparameters tuned to minimize the loss function.

[0012] Furthermore, in step 4, the formula for calculating the mean μ of the validation set reconstruction error is: ; ; Among them, e j Let m be the reconstruction error of the j-th sample, and m be the number of samples in the validation set. The reconstruction error threshold is set to μ+1σ to μ+3σ. If the reconstruction error score of a sample is within the threshold range of the reconstruction error threshold, it is judged as slightly abnormal data. If the error score exceeds μ+3σ, it is judged as severely abnormal data, where σ represents the standard deviation.

[0013] Furthermore, in step 5, the evaluation indicators for the effectiveness of the early warning include the 24-hour effective early warning rate, the 24-hour false early warning rate, and the 24-hour missed early warning rate. 24-hour effective early warning rate = (Number of effective early warnings / Total number of early warnings) × 100%. An effective early warning is defined as an abnormal event occurring on the coating machine within 24 hours after the occurrence of an abnormal early warning point. 24-hour false alarm rate = (Number of invalid alarms / Total number of alarms) × 100% 24-hour missed warning rate = (Total number of abnormal events - Number of successful warnings before the occurrence of an anomaly) / Total number of abnormal events × 100%.

[0014] Furthermore, it also includes a statistical alarm step for primary anomaly detection: after data preprocessing, statistical methods are used to perform first-level anomaly detection on the training set. If the real-time data exceeds the specified threshold or the offset per unit time is greater than the specified threshold, an alarm is directly triggered, and the anomaly points in the alarm are simultaneously repaired by linear interpolation, forming a multi-level anomaly detection mechanism of "statistical primary detection + isolated forest fine detection" to further improve the quality of training data.

[0015] Furthermore, step 6 is also included: adopting a computing architecture that allows the edge and local server to work together, with the local server uniformly completing the training and hyperparameter tuning of the LightGBM model, and deploying the LightGBM model after pruning and optimization to the micro edge device of the coating machine, thereby separating data acquisition from data computation; multi-dimensional data collected by the Internet of Things is transmitted to the edge in a lightweight manner to complete diagnostic inference; the micro edge device includes at least one of the following: coating head, tension controller, feed pump, and edge detection sensor.

[0016] Compared with the prior art, the present invention has the following beneficial effects: Dual improvement in data quality and model accuracy: A multi-level anomaly detection collaborative mechanism of "basic statistical anomaly detection + refined isolation forest detection" is constructed. By repairing explicit and implicit anomalies in the training data through linear interpolation, the interference of "dirty data" on the model is avoided, and the dependence of the LightGBM model on "clean data" is greatly reduced. This enables the model to adapt to the mixed complexity of data affected by dust and vibration in the coating machine production scenario, and significantly improves the model's accuracy in identifying minor faults and implicit anomalies.

[0017] Significantly reduce labor and maintenance costs: A semi-supervised labeling mechanism of "early-stage minor labeling + later-stage automatic labeling" is proposed. In the early stage, only a small amount of seriously abnormal data needs to be manually labeled. In the later stage, implicit and minor anomaly labels are automatically completed by relying on a well-trained model. No full manual participation is required, which not only reduces the labor cost of data labeling, but also ensures the comprehensiveness of labeled data and provides sufficient data support for continuous model optimization.

[0018] Real-time fault early warning at the edge: The LightGBM model is pruned and optimized to adapt to the low computing power and small storage hardware resources of the micro edge device of the coating machine. At the same time, the data collected by the Internet of Things is transmitted to the edge in a lightweight manner to complete the diagnostic inference, which significantly reduces the data transmission bandwidth occupation and diagnostic latency. Experimental verification shows that, under the premise of the same missing early warning rate, the effective early warning rate in 24 hours is 23% higher than that of the traditional method. It can accurately match the fault early warning needs of the high real-time production of the coating machine, and provide strong support for timely handling of anomalies and reducing product scrap.

[0019] The system boasts strong operational stability and adaptability: It adopts a computing architecture that integrates edge and local server operations, separating data acquisition from data computation, thus significantly improving the overall system's operational stability. The core detection framework supports flexible replacement of different anomaly detection algorithms, which can be adapted and adjusted according to the characteristics of different production processes of the coating machine. It is also compatible with the integration and expansion of multiple encoder structures, meeting the personalized diagnostic needs of different coating machine models and different production scenarios.

[0020] The evaluation system aligns with actual production needs: It establishes an evaluation system for early warning effectiveness based on the 24-hour effective early warning rate, false early warning rate, and missed early warning rate. It eliminates the need for full manual annotation of abnormal data, solving the problem of traditional evaluation indicators relying on precise annotation. It can objectively and efficiently evaluate the early warning effect of the end-side diagnostic system, providing a scientific basis for model iteration and optimization. Attached Figure Description

[0021] Figure 1 This is a flowchart of the multi-level prediction algorithm in an embodiment of the present invention; Figure 2 This is a flowchart of the micro end-side diagnostic method for coating machines in an embodiment of the present invention. Detailed Implementation

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

[0023] See Figure 1-2This invention provides an artificial intelligence-based method for predicting faults in micro-end-side equipment of a coating machine. The LightGBM model of this method, after pruning and optimization, is deployed on the micro-end-side equipment of the coating machine (coating head, tension controller, feed pump, end-side detection sensors, etc.). A computational architecture is adopted that allows the end-side equipment and the local server to work collaboratively. The local server is responsible for model training and hyperparameter tuning, while the end-side equipment is responsible for data acquisition and real-time diagnostic inference. The specific steps of the method are described in detail below: Step 1: Multi-dimensional time series data collection and preliminary light labeling A multi-node data acquisition system for the coating machine is built based on the Internet of Things sensor network. It collects core operating parameters of multiple nodes of the core components such as the coating head and the oven in real time during the production process of the coating machine. The core operating parameters include at least one of temperature, vibration, speed, tension, roller speed, coating viscosity, motor current, and coating pressure. The data acquisition interval is set to 10 seconds. The parameters of the coating machine head form 5-dimensional time series data, and the parameters of the coating machine oven form 15-dimensional time series data. Finally, a multi-dimensional time series data covering both normal and abnormal operating states of the coating machine is formed.

[0024] The collected multi-dimensional time series data is lightly manually labeled in the early stage. Only a small number of serious abnormal data that cause obvious product defects or equipment downtime are labeled as anomalies. There is no need to label hidden or minor abnormal data, which greatly reduces the initial manual labeling cost.

[0025] Seriously abnormal data refers to abnormal node data that affects the normal operation of the coating machine, including abnormal data of the coating head, drying oven, etc. that affect the operation of the coating machine.

[0026] Step 2: Data partitioning and preprocessing 2.1 Dataset Partitioning The collected historical multi-dimensional time series data were divided into training and validation sets in an 8:2 ratio. The training set was used for the reconstruction training of the LightGBM model, and the validation set was used for the performance verification of the model and the setting of anomaly detection thresholds. After the division, the severely abnormal data that had been labeled in the training set were removed, and the data that was suspected to be operating normally were retained.

[0027] 2.2 Multi-stage data preprocessing The historical data of the coating machine head and oven in both the training and validation sets are preprocessed in a unified manner, specifically including: Data aggregation and smoothing: Data is aggregated at 10-second intervals, and multiple data points collected within the same time interval are processed using mean smoothing to reduce the interference of sensor noise on the data. Missing value imputation: Missing values ​​in the data are filled using forward interpolation to ensure the continuity of the time series data; Z-score standardization: To eliminate differences in the units and numerical ranges of different features and improve model training performance, the processed data is standardized using Z-score, with the formula as follows: ; Where x is the original data value of the feature. The characteristic mean, The characteristic standard deviation; 2.3 Multi-level anomaly detection and data repair This invention establishes a two-level anomaly detection mechanism to locate and repair anomalies in suspected normal operating data in the training set, specifically as follows: Level 1: Basic Statistical Anomaly Detection and Alarm: Statistical methods are used to detect anomalies in the preprocessed training set. If the real-time data exceeds a preset threshold or the offset per unit time is greater than a preset threshold, an alarm is directly triggered and the data is marked as an explicit anomaly. Level 2: Isolation Forest Fine Anomaly Detection: The Isolation Forest algorithm is used to perform preliminary anomaly detection on the training set after statistical testing, accurately locating hidden anomalous data points in the data; Linear interpolation repair: For both explicit and implicit outliers detected above, linear interpolation is used for repair. Outliers are filled with the mean of the 10 normal data points before and after the outlier. The calculation formula is as follows: ; Where, x repaired The value after anomaly repair, x t-i x is the value of the i-th normal data point before the outlier. t+i This represents the value of the i-th normal data point after the outlier.

[0028] After the repair is completed, the training set contains only clean, normally functioning data, providing high-quality data support for model training.

[0029] Step 3: Lightweight LightGBM Model Construction and Reconstruction Training 3.1 Model Structure Design A lightweight LightGBM model adapted for deployment on the micro end side of a coating machine is constructed. To address the feature differences between the 5-dimensional operating data of the coating machine head and the 15-dimensional operating data of the oven, feature engineering (including data standardization and feature filtering) is performed on the two types of data respectively. Invalid features are removed and core features related to faults are retained. The processed head data and oven data are then fused as model input to improve the model's feature learning efficiency and fault identification accuracy.

[0030] 3.2 Model Training Parameter Settings The repaired training set containing only normally functioning data is input into the lightweight LightGBM model, and supervised training is performed using Gradient Boosting Decision Tree (GBDT). The training loss function is cross-entropy (LogLoss), and the formula is as follows: ; Where K is the number of categories, y i,k For the one-hot encoding of the true label of the i-th sample, p i,k Let be the predicted probability that the i-th sample belongs to the k-th class, and n be the number of samples.

[0031] The hyperparameters for model training are set as follows: the number of leaf nodes in each gradient boosting decision tree is 20, the learning rate is 0.005, the training objective is to minimize the loss function, and the hyperparameters such as tree depth and number of iterations are tuned to ensure the accuracy of model training while controlling the complexity and computing power requirements of the model and adapting it to deployment on micro edge devices.

[0032] 3.3 Model Lightweight Optimization The trained LightGBM model is pruned and optimized to remove redundant branches, simplify the model structure, reduce the model's storage footprint and inference computing power requirements, and ensure that the model can run stably and efficiently on the coating machine's micro-end device.

[0033] Step 4: Diagnostic reasoning for end-sided abnormalities The validation set data, which includes data from unrepaired normal operation, is transmitted in a lightweight manner and then input into the trained LightGBM model deployed on a micro-device. The reconstruction error score of each data sample in the validation set is calculated through model inference. The reconstruction error score reflects the degree of deviation of the sample data from the characteristics of the normal operation state of the coating machine.

[0034] Based on the reconstruction error scores from the validation set, the mean μ and standard deviation σ of the reconstruction error are calculated using the following formulas: ; ; Where e j Let m be the reconstruction error of the j-th sample, and m be the number of samples in the validation set. The reconstruction error threshold is set to μ+1σ to μ+3σ. If the reconstruction error score of a sample is within the threshold range of the reconstruction error threshold, it is judged as slightly abnormal data. If the error score exceeds μ+3σ, it is judged as seriously abnormal data, thus realizing real-time end-side diagnosis of abnormal working conditions of the coating machine.

[0035] This invention classifies parameters such as temperature, vibration, speed, tension, and roller speed of the coating machine head and oven into states using 3 times standard deviation and 1 time standard deviation: data within 1 time standard deviation is considered normal, data between 1 and 3 times standard deviation is considered slightly abnormal, and data exceeding 3 times standard deviation is considered seriously abnormal. This conforms to the steady-state monitoring specifications for industrial equipment and the classification is reasonable and reliable. Step 5: Performance Evaluation of Automatic Labeling and Early Warning Functions 5.1 Semi-supervised automatic annotation Based on the well-trained LightGBM model and real-time data collected from the coating machine via IoT, the model continuously infers from the real-time data, automatically identifies latent and minor anomalies, and automatically labels them. The automatically labeled anomaly data is then merged with the minor anomaly data manually labeled in step 1 to form complete labeled data of coating machine faults, providing a sufficient and comprehensive dataset for iterative optimization of the model.

[0036] 5.2 Evaluation of the effectiveness of early warning Based on the above complete labeled data, the effectiveness evaluation index of the early warning system is calculated to objectively assess the fault prediction performance of the model. The evaluation index includes the 24-hour effective early warning rate, the 24-hour false early warning rate, and the 24-hour missed early warning rate. The specific calculation method is as follows: 24-hour effective early warning rate = (Number of effective early warnings / Total number of early warnings) × 100%, where an effective early warning is defined as: an actual abnormal event occurring on the coating machine within 24 hours after the occurrence of an abnormal early warning point; 24-hour false alarm rate = (Number of invalid alarms / Total number of alarms) × 100%, where invalid alarms are defined as: no abnormal events occur on the coating machine within 24 hours after the occurrence of an abnormal alarm point; 24-hour missed warning rate = (Total number of abnormal events - Number of successful warnings before the occurrence of an anomaly) / Total number of abnormal events × 100%.

[0037] The above three indicators can be used to comprehensively and objectively evaluate the model's early warning accuracy, false positive rate, and false negative rate. If the indicators do not meet the preset requirements, the complete labeled data will be re-input into the model for iterative training until the model performance meets the fault prediction requirements of the coating machine production.

[0038] Step 6: The client-side and the local server work together. This invention adopts a computing architecture that integrates edge devices and local servers. The local server uniformly completes the training, hyperparameter tuning, and iterative optimization of the LightGBM model. The trained and pruned lightweight model is then distributed to each micro edge device of the coating machine. The edge devices collect operational data in real time through an IoT sensor network. After lightweight transmission, fault diagnosis inference and anomaly warning are performed locally. This achieves the separation of data acquisition and data computation, making full use of the high-performance computing power of the local server while leveraging the low latency advantage of the edge devices, thus significantly improving the overall system's operational stability and real-time performance.

[0039] The pruning optimization process is as follows: First, based on the trained base learner structure, calculate the leaf node split gain, node sample number and feature contribution of each decision tree, and identify and remove redundant branches with split gain below the set threshold, insufficient sample number and no significant impact on the prediction results. Secondly, based on the importance of features, redundant features and invalid split nodes with contribution rates below the threshold are removed to reduce the feature dimension and computational cost of the model. Finally, under the constraint of ensuring that the model's prediction accuracy loss is less than 1%, global pruning is performed on the number and depth of decision trees to further simplify the model structure and reduce computational complexity.

[0040] The lightweight LightGBM model, optimized through multi-level pruning, is distributed and deployed to the micro-end device of the coating machine through a secure channel, realizing the physical separation and collaborative operation of data acquisition and model inference calculation, and meeting the industrial application requirements of low computing power, low latency, and high real-time performance of the end device. Application Examples

[0041] To further verify the practical application effect of the present invention, the implementation effect of the present invention will be described in detail below in two practical application scenarios: lithium battery cell production line and photovoltaic module glass coating production line.

[0042] Example 1: Fault Prediction of Coating Machine in Lithium Battery Cell Production Line Application scenarios

[0043] The coating machine is the core production equipment of a lithium battery cell production line. It is necessary to predict in real time the faults such as roller misalignment, coating viscosity fluctuation and oven temperature runaway. The production line collects more than 20 dimensions of operating data of the coating machine, with a data collection interval of 10 seconds. The production site space is compact and the computing power and storage resources of the end-side equipment are limited.

[0044] The fault prediction method for micro end-side equipment of coating machine based on artificial intelligence of the present invention is applied to the coating machine. Data acquisition, preprocessing, model building and training, end-side deployment and collaborative operation are completed according to the above steps 1-6. The model is deployed on micro end-side equipment such as coating head and tension controller of coating machine, and adopts end-side and local server collaborative computing architecture. Implementation Results

[0045] Compared with traditional single-algorithm fault detection methods, the method of this invention improves the 24-hour effective warning rate from 59% to 82% under the same leak warning rate, achieving accurate identification of minor faults and hidden anomalies in the coating machine. At the same time, only 5% of serious abnormal data needs to be manually labeled, while the remaining abnormal data is automatically labeled by the model, reducing the cost of manual data labeling by more than 90%. The inference latency of the model on the edge device is less than 50ms, and the data transmission bandwidth usage is reduced by 70%, fully meeting the real-time and low computing power requirements of lithium battery cell production lines. The product scrap rate caused by coating machine failures is reduced by 65%, and production stability is greatly improved.

[0046] Example 2: Fault Prediction of Coating Machine in Photovoltaic Module Glass Coating Production Line The fault prediction method of this invention can be flexibly extended and adapted to the field of photovoltaic module glass coating. Targeting the coating requirements of large-size, high-flatness, and high-adhesion photovoltaic glass, the coating machine is specifically adapted and modified. Then, the core fault prediction method of this invention is applied to achieve real-time end-side fault prediction of the photovoltaic module glass coating machine. Specific adaptation, modification, and implementation effects are as follows: 1. Core adaptation and modification

[0047] The core steps of the fault prediction method of this invention remain unchanged. However, considering the characteristics of photovoltaic module glass coating, the coating machine is adapted as follows. The adapted coating machine still uses the lightweight LightGBM model of this invention for end-side fault prediction: Equipment size adaptation: The coating width of the coating machine is extended to 1600~2200mm to meet the large size requirements of conventional photovoltaic glass (1660mm×1000mm, 2100mm×1600mm); the traction mechanism adopts the "multi-roller synchronous traction" design, with 4~6 sets of traction rollers evenly distributed on the upper and lower surfaces of the photovoltaic glass to avoid glass deviation and bending during operation.

[0048] Coating process adaptation: The coating method is optimized to "scraper + spray composite coating". The anti-reflective coating slurry is sprayed first and then smoothed with a scraper to avoid scratches on the glass surface caused by single scraper coating; The coating gap adjustment mechanism is equipped with a "multi-point gap detection" function. One gap detection sensor is set at each end and the middle of the coating head to detect and automatically adjust the coating gap in real time with an adjustment accuracy of ±0.5μm to compensate for the slight flatness deviation of the photovoltaic glass surface.

[0049] Coating supply system adaptation: Add an "online viscosity detection unit" to detect the viscosity of the anti-reflective coating slurry in real time, with a control range of 500~1500mPa・s, and automatically adjust the amount of thinner added when the viscosity fluctuates; Add a vacuum degassing unit, with the vacuum degree controlled at -0.08~-0.09MPa, to remove air bubbles in the coating and avoid pinholes and bubble defects after coating.

[0050] Added field-specific functions: Added "Glass Surface Pretreatment Unit", which integrates dust removal, oil removal and plasma treatment modules, with a plasma treatment power of 500~1000W, to improve the surface tension of the glass and make the coating adhesion ≥4B (compliant with photovoltaic module industry standards); Added "Coating Curing Pretreatment Unit", which adopts an infrared heating device with a temperature control range of 80~120℃ to pre-cur the coating after coating, shortening the subsequent curing process time. 2. Implementation Results

[0051] The fault prediction method of this invention is applied to the adapted photovoltaic module glass coating machine, realizing real-time end-side prediction of faults such as viscosity fluctuation, gap deviation, and temperature runaway. The effective early warning rate is over 85% in 24 hours, and the missed early warning rate is less than 5%. At the same time, the coating machine achieves large-size, high-precision coating of photovoltaic glass, with coating uniformity ≤ ±0.8μm. The coating adhesion and light transmittance both meet industry standards, and the light transmittance is 3%~5% higher than that of traditional coating machines. The coating weather resistance (high and low temperature, ultraviolet aging) meets the long-term outdoor use requirements of photovoltaic modules and is adapted to the production needs of large-scale photovoltaic module production lines.

[0052] The above description is merely an embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural or procedural transformations made based on the content of the present invention's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. A method for predicting faults in micro-end-side equipment of a coating machine based on artificial intelligence, characterized in that, Includes the following steps: Step 1: Relying on the Internet of Things sensor network, collect the core operating parameters of multiple nodes in the coating machine production process in real time to form multi-dimensional time series data covering normal and abnormal operating states. Only a small number of severely abnormal data are manually marked in the early stage. Step 2: Divide the collected historical multi-dimensional time series data into training set and validation set, and remove the severely anomalous data that has been labeled in the training set; use the isolated forest algorithm to perform preliminary anomaly detection on the training set to locate latent anomalous data points, and repair the detected latent anomalous data points through linear interpolation. Step 3: Construct a lightweight LightGBM model adapted for deployment on the micro end of the coating machine. Input the repaired training set data containing only normal operation data into the LightGBM model for reconstruction training, so that the model can learn the characteristic rules of the normal operation state of the coating machine. Step 4: Input the validation set data containing the unrepaired normal operation data into the trained LightGBM model, calculate the reconstruction error score of each data sample; set a reconstruction error threshold. If the reconstruction error score of a sample exceeds the reconstruction error threshold, it is determined to be abnormal data, thereby realizing real-time end-side diagnosis of abnormal operating conditions of the coating machine. Step 5: Based on the well-trained LightGBM model and the real-time collected coating machine operation data, perform automatic annotation on latent and minor anomalies in the later stage, and combine it with the previous manual minor annotation to form complete annotation data; Based on the complete labeled data, calculate the early warning effectiveness evaluation index and evaluate the model performance.

2. The method for predicting faults in a micro-end-side device of a coating machine based on artificial intelligence according to claim 1, characterized in that, In step 1, the multi-node core operating parameters include at least one of the following: temperature, vibration, speed, tension, roller speed, coating viscosity, motor current, and coating pressure of the coating machine head and oven; the acquisition interval of the multi-dimensional time series data is 10 seconds, the coating machine head generates 5-dimensional time series data, and the coating machine oven generates 15-dimensional time series data.

3. The method for predicting faults in a micro-end-side device of a coating machine based on artificial intelligence according to claim 1, characterized in that, Step 2, data preprocessing also includes: aggregating historical multi-dimensional time series data of the coating machine head and oven according to the collection interval; applying mean smoothing to multiple data points within the same time interval; and filling missing data using forward interpolation; and applying Z-score standardization to the processed data to eliminate differences in the units and numerical ranges of different features. The Z-score standardization formula is: ； Where x is the original data value of the feature. The characteristic mean, The standard deviation is the feature value; the ratio of the training set to the validation set is 8:

2.

4. The method for predicting faults in a micro-end-side device of a coating machine based on artificial intelligence according to claim 1, characterized in that, In step 2, when using linear interpolation to repair latent anomalies, the outlier is filled with the mean of the 10 normal data points before and after the outlier. The calculation formula is as follows: ； Where, x repaired The value after anomaly repair, x t-i x is the value of the i-th normal data point before the outlier. t+i This represents the value of the i-th normal data point after the outlier.

5. The method for predicting faults in a micro-end-side device of a coating machine based on artificial intelligence according to claim 1, characterized in that, In step 3, when constructing the lightweight LightGBM model, feature engineering is performed on the feature differences between the 5-dimensional operating data of the coating machine head and the 15-dimensional operating data of the oven. Feature engineering includes data standardization and feature selection. Then, the processed machine head data and oven data are fused as model input. Model training uses gradient boosting decision trees for supervised training, and the training loss function is cross-entropy, with the formula: ； Where K is the number of categories, y i,k For the one-hot encoding of the true label of the i-th sample, p i,k Let be the predicted probability that the i-th sample belongs to the k-th class, and n be the number of samples.

6. The method for predicting faults in a micro-end-side device of a coating machine based on artificial intelligence according to claim 5, characterized in that, During the training of the LightGBM model, each gradient boosting decision tree has 20 leaf nodes and a learning rate of 0.

005. The model hyperparameters are tuned with the goal of minimizing the loss function.

7. The method for predicting faults in a micro-end-side device of a coating machine based on artificial intelligence according to claim 1, characterized in that, In step 4, the formula for calculating the mean μ of the validation set reconstruction error is: ； ； Among them, e j Let m be the reconstruction error of the j-th sample, and m be the number of samples in the validation set. The reconstruction error threshold is set to μ+1σ to μ+3σ. If the reconstruction error score of a sample is within the threshold range of the reconstruction error threshold, it is judged as slightly abnormal data. If the error score exceeds μ+3σ, it is judged as severely abnormal data, where σ represents the standard deviation.

8. The method for predicting faults in a micro-end-side device of a coating machine based on artificial intelligence according to claim 1, characterized in that, In step 5, the evaluation indicators for the effectiveness of the early warning include the 24-hour effective early warning rate, the 24-hour false early warning rate, and the 24-hour missed early warning rate; 24-hour effective early warning rate = (Number of effective early warnings / Total number of early warnings) × 100%. An effective early warning is defined as an abnormal event occurring on the coating machine within 24 hours after the occurrence of an abnormal early warning point. 24-hour false alarm rate = (Number of invalid alarms / Total number of alarms) × 100% 24-hour missed warning rate = (Total number of abnormal events - Number of successful warnings before the occurrence of an anomaly) / Total number of abnormal events × 100%.

9. The method for predicting faults in a micro-end-side device of a coating machine based on artificial intelligence according to claim 1, characterized in that, It also includes a statistical alarm step for primary anomaly detection: after data preprocessing, statistical methods are used to perform first-level anomaly detection on the training set. If the real-time data exceeds the specified threshold or the offset per unit time is greater than the specified threshold, an alarm is directly triggered, and the anomaly points in the alarm are simultaneously repaired by linear interpolation.

10. The method for predicting faults in a micro-end-side device of a coating machine based on artificial intelligence according to claim 1, characterized in that, It also includes step 6: adopting a computing architecture that allows the edge and local server to work together, with the local server uniformly completing the training and hyperparameter tuning of the LightGBM model, and deploying the LightGBM model after pruning and optimization to the micro edge device of the coating machine, thereby achieving the separation of data acquisition and data calculation; multi-dimensional data collected by the Internet of Things is transmitted to the edge in a lightweight manner to complete diagnostic inference; the micro edge device includes at least one of the following: coating head, tension controller, feed pump, and edge detection sensor.

Images