A digital twin driven photovoltaic laminator operation monitoring and failure prediction system

Abstract

The application discloses a digital twin driven photovoltaic laminator operation monitoring and fault prediction system, and relates to the technical field of photovoltaic equipment.The application realizes real-time monitoring by deploying various types of sensors on the photovoltaic laminator, extracts temperature field features, pressure and vibration signal features, PLC working condition parameter features and twin residual features from historical fault data of the photovoltaic laminator, divides feature vectors of the historical fault data into the same fault feature data set, calculates a causal sensitivity matrix S and a mutual information matrix M, determines the first discriminant features corresponding to each fault according to the comprehensive contribution degree, calculates the deviation of each dimension of the actual feature vector from the specified threshold, finds out the feature corresponding to the maximum deviation, takes the feature as the contrast feature, matches the contrast feature with the first discriminant features of all faults, thereby predicting the fault, and adjusts the real-time parameters of the predicted different levels of faults by using a PPO algorithm.

Description

Technical Field

[0001] This invention relates to the field of photovoltaic equipment technology, specifically to a digital twin-driven photovoltaic laminator operation monitoring and fault prediction system. Background Technology

[0002] Photovoltaic laminators are the core equipment for photovoltaic module encapsulation. Their operational stability directly affects module quality and production efficiency. However, existing operation monitoring and fault prediction technologies have two major shortcomings that make it difficult to meet the needs of high-precision industrial-grade operation and maintenance.

[0003] First, the ability to detect weak anomalies is insufficient, making early fault identification difficult. Traditional monitoring methods rely on single-point temperature measurement and low-frequency pressure sensing technologies, which are difficult to capture weak signals during equipment operation. For example, features such as localized hot spots on heating plates, micro-leakage in hydraulic systems, and specific frequency vibrations caused by early bearing wear are easily overlooked because existing sensors cannot achieve high-frequency dynamic sampling, resulting in missed early fault detection.

[0004] Secondly, the model has poor adaptability under dynamic changes in process parameters, requiring frequent manual intervention. Existing fault prediction models are mostly trained based on fixed operating condition data and lack the ability to dynamically map the physical mechanism of the laminator to changes in process parameters. When process parameters such as heating rate, lamination time, and vacuum degree are adjusted during production, or when material characteristics and environmental conditions change, the model is difficult to adapt, resulting in a decrease in prediction accuracy.

[0005] Therefore, there is an urgent need for a new system that combines high-resolution sensing with digital twin mechanism modeling to improve monitoring accuracy and achieve precise assessment of equipment operating status and fault prediction. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides a digital twin-driven photovoltaic laminator operation monitoring and fault prediction system, which solves the problems of insufficient detection rate of weak features and failure of the process parameter change model.

[0007] To achieve the above objectives, the present invention provides the following technical solution: a digital twin-driven photovoltaic laminator operation monitoring and fault prediction system, comprising:

[0008] The sensor monitoring module deploys an infrared array temperature sensor, a high-frequency piezoelectric pressure sensor, a MEMS triaxial accelerometer, and Modbus TCP for integrating PLC data on the photovoltaic laminator.

[0009] The feature extraction module is used to extract temperature field features, pressure and vibration signal features, PLC operating condition parameter features, and twin residual features from historical fault data of photovoltaic laminators.

[0010] The fault prediction module divides the feature vectors of historical fault data into the same fault feature data set, calculates the causal sensitivity matrix S and mutual information matrix M, determines the first discriminant feature corresponding to each fault based on the comprehensive contribution, calculates the deviation of each dimension of the actual feature vector from the specified threshold, finds the feature corresponding to the largest deviation, uses this feature as the comparison feature, and matches the comparison feature with the first discriminant feature of all faults to predict the fault.

[0011] The parameter adjustment module uses the PPO algorithm to adjust the parameters in real time for different levels of predicted faults.

[0012] As a further aspect of the present invention, the temperature field features include the 10-dimensional temperature field PCA principal components extracted by PCA, the coordinates and area values ​​of the largest anomaly region, the heating slope and overshoot, and the standard deviation of temperature fluctuation during the heat preservation stage; the pressure signal features include kurtosis, margin, rising edge slope, and the energy proportion of the eight frequency bands obtained by three-layer wavelet decomposition; the vibration signal features include the STFT energy centroid frequency, the kurtosis of the first three IMFs after EMD decomposition, the triaxial effective value, and the triaxial peak factor; the PLC operating condition parameter features include the mean speed, mean vacuum level, speed slope, vacuum level slope, number of temperature exceedances, and number of pressure exceedances; the twin residual features are the absolute values ​​of the differences between the simulation features and the real features of the digital twin model.

[0013] As a further aspect of the present invention, based on the physical equations of the digital twin model, the causal dependency relationship between features and faults is defined, and a causal sensitivity matrix S is generated, where Si,j represents the physical sensitivity of feature fi changes to fault cj.

[0014] For the feature vector sequence in the fault set, the mutual information I(fi;cj) between feature fi and fault label cj is calculated according to the time window, and a dynamic weight factor wt is added to assign different weights to the data at different time points. The specific formula is wt=a^t, where a∈(0,1) and t is the time from the current time point.

[0015] Record the mutual information value of each feature to the fault in different time windows, and generate a mutual information matrix M;

[0016] The features in the same fault set are sorted according to the comprehensive contribution formula CMIC(fi)=ɑ*mean(Si,j)+(1-ɑ)*mean(Mi,j), and the feature with the highest contribution is selected as the first discriminant feature, where ɑ is the mechanism-data balance coefficient.

[0017] As a further aspect of the present invention, if at least two faults have the same first discriminant feature, the first discriminant feature is marked as a shared first feature. For the fault set marked as a shared first feature, the obtained twin residual is used to train a Local Sensitive Hash (LSH) model to obtain a second discriminant model.

[0018] As a further aspect of the present invention, if the comparison feature is a first discriminant feature but not a shared first feature, then the fault corresponding to the first discriminant feature is directly used as the predicted fault; if the comparison feature is a shared first feature, then the actually obtained twin residual vector is input into the second discriminant model corresponding to the fault of the shared first feature to predict the corresponding fault.

[0019] As a further aspect of the present invention, after predicting the corresponding fault, the actual twin residual vector is input into the trained GCN network, and the probability of predicting different levels of fault is output, wherein the fault levels can be divided into light, medium and heavy.

[0020] As a further aspect of the present invention, the specific steps for real-time parameter adjustment of predicted faults of different levels using the PPO algorithm are as follows:

[0021] Obtain the input multidimensional state vector S, which specifically includes the fault characteristic physical quantity Qpred, the current temperature rise rate vT, the pressure setpoint P, the lamination time tlam, the probability of different levels of fault, the real-time yield rate Ygood, the actual production cycle Tcycle, the ambient temperature, and the ambient humidity.

[0022] Design a continuous adjustable parameter space A, specifically including the following process parameter adjustment amounts: △vT∈[vTth1,vTth2], △P∈[Pth1,Pth2], △tlam∈[tlamth1,tlamth2];

[0023] The fault risk reward is calculated according to the formula Rrisk=-((Qpred-Qsafe) / (Qmax_risk-Qsafe))^2, where Qsafe is the corresponding fault physical quantity threshold and Qmax_risk is the severe fault threshold.

[0024] The production efficiency bonus is calculated using the formula Refficiency=λ1*Ygood-λ2*((Tcycle-Tnominal) / (Tmax_cycle-Tnominal))^2, where λ1 and λ2 are weighting coefficients, Tnominal is the production cycle threshold, and Tmax_cycle is the maximum production cycle threshold.

[0025] A reward function balancing failure risk and production efficiency is constructed based on the formula R=c*Rrisk+(1-c)*Refficiency, where c is an adjustment factor;

[0026] The edge device collects state St, action At, and reward Rt in real time to form a trajectory {St,At,Rt,St+1}, and triggers a policy update every 200 steps.

[0027] The advantage function At = Rt + γ * V(St+1) - V(St) is calculated using generalized advantage estimation, where γ is the discount factor.

[0028] The objective function is optimized by pruning importance sampling: L=Et[min(rt(A)*At,clip(rt(A),1-ε,1+ε)*At)], where rt(A) is the probability ratio of the new and old policies, and the Actor-Critic network parameters are updated synchronously, where ε is the pruning parameter;

[0029] When a fault occurs, the PPO strategy is activated, and the optimal action is output every 10 seconds and sent to the PLC for execution via the industrial protocol.

[0030] As a further aspect of the present invention, the impact of parameter changes on lamination quality is simulated in a digital twin model before parameter adjustment. If the expected results are not achieved after three consecutive adjustments, the most recent valid parameter is automatically restored and an early warning is triggered.

[0031] As a further aspect of the present invention, failure to meet expectations specifically includes Qpred not decreasing or yield being less than 98%.

[0032] This invention provides a digital twin-driven photovoltaic laminator operation monitoring and fault prediction system, which has the following advantages compared with the prior art:

[0033] (1) This invention achieves comprehensive monitoring of the operating status of photovoltaic laminators by combining multiple sensors, which can promptly detect potential faults and improve the reliability of equipment operation;

[0034] (2) The fault prediction method based on digital twin technology and multi-source data feature extraction of the present invention makes full use of the advantages of historical fault data and digital twin model, improves the accuracy of fault prediction and reduces production costs.

[0035] (3) The present invention uses the PPO algorithm to dynamically adjust parameters, which can optimize process parameters in real time according to equipment operating status and production indicators, thereby enhancing the system's adaptability. Attached Figure Description

[0036] Figure 1 This is a block diagram illustrating the system principle of the present invention. Detailed Implementation

[0037] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0038] like Figure 1 This invention provides a digital twin-driven photovoltaic laminator operation monitoring and fault prediction system, comprising:

[0039] The sensor monitoring module deploys an infrared array temperature sensor, a high-frequency piezoelectric pressure sensor, a MEMS triaxial accelerometer, and Modbus TCP for integrating PLC data on the photovoltaic laminator.

[0040] The infrared array temperature sensor can cover 1024 temperature measurement points with an accuracy of ±0.5℃, replacing the traditional single-point temperature measurement and realizing millimeter-level resolution scanning of the temperature field of the heating plate to capture local hot spots;

[0041] The high-frequency piezoelectric pressure sensor, with a sampling rate that can be set to 10kHz, can monitor dynamic pressure fluctuations in hydraulic systems, such as micro-leakage at the 0.1MPa level.

[0042] MEMS triaxial accelerometers can monitor the time-frequency domain characteristics of mechanical vibrations, such as the 100Hz resonance peak caused by bearing failure.

[0043] Modbus TCP integrates PLC data, is compatible with older equipment, and enables real-time synchronization of operating parameters such as oil pump speed and vacuum level.

[0044] The feature extraction module is used to extract temperature field features, pressure and vibration signal features, PLC operating condition parameter features, and twin residual features from historical fault data of photovoltaic laminators.

[0045] Temperature field feature extraction specifically includes PCA principal component analysis of the temperature field. The 1024 temperature data points collected by the infrared array sensor contain a large amount of redundant information. The top 10 principal components with a cumulative variance contribution rate greater than 95% are extracted by PCA dimensionality reduction, which can compress the amount of data and retain the main change patterns of the temperature field.

[0046] The largest abnormal region is identified, including its coordinates and area value. 3D features are extracted from this region. The temperature difference between each temperature measurement point and its 8 neighbors is calculated. Regions with gradients greater than 2℃ / mm are marked to form multiple discrete abnormal regions. The area of ​​each region is calculated, and only the region with the largest area is retained to ensure that only the most significant abnormal region is processed each time.

[0047] The heating slope and overshoot are used to extract two-dimensional features and identify abnormalities in the temperature control system, such as PID algorithm failure leading to slow heating or overshoot. This helps to prevent EVA film from local over-melting or insufficient curing due to uneven temperature, thus ensuring the quality of component encapsulation.

[0048] The standard deviation of temperature fluctuation during the heat preservation stage is used to extract one-dimensional features and evaluate the long-term stability of the heating system. When the standard deviation exceeds the specified threshold, it indicates poor contact of the heating wire or failure of the temperature control module, thus realizing early warning of gradual heating failure.

[0049] The pressure signal features specifically include kurtosis, one-dimensional features are extracted, and it is sensitive to pressure spikes to detect transient impact faults.

[0050] The margin, which extracts one-dimensional features, is more sensitive to low-amplitude periodic impacts, such as the initial micro-leakage of the sealing ring, which is greater than 1.5, and is superior to the traditional effective value.

[0051] Rising edge slope is used to extract 1D features and identify transient faults such as hydraulic valve jamming.

[0052] The energy proportions of the eight frequency bands obtained by the three-level wavelet decomposition are used to extract eight-dimensional features. Gradual faults are monitored by the frequency band energy change trend, which improves the false negative rate and makes up for the shortcomings of FFT in analyzing non-stationary signals.

[0053] The vibration signal characteristics specifically include the STFT energy center of gravity frequency. One-dimensional features are extracted, and when the bearing fails, the energy center of gravity shifts to higher frequencies.

[0054] After EMD decomposition, the first three IMF kurtosis values ​​are used to extract 3D features. After empirical mode decomposition, the first three IMF kurtosis values ​​of different faults show differences. For example, gear cracks will cause IMF2 kurtosis to be greater than 5, which is different from the IMF1 kurtosis anomaly of bearing faults.

[0055] Triaxial RMS values ​​are used to extract 1D features that reflect vibration intensity.

[0056] The three-axis peak factor is used to extract one-dimensional features. A peak factor greater than 3 indicates impact failure, such as bearing spalling.

[0057] The specific characteristics of PLC operating parameters include average speed, average vacuum, speed slope, and vacuum slope. Four-dimensional features are extracted, and the 5-minute sliding window mean and trend slope are calculated for low-frequency data such as oil pump speed and vacuum to identify progressive anomalies.

[0058] The number of temperature and pressure out-of-range times is extracted into 2D features. The number of times parameters such as lamination time and temperature setpoint exceed the process range within a specified time is counted. High-frequency out-of-range times indicate control system failure or sensor drift.

[0059] Twin residual characteristics: For any of the above characteristics f measured, the twin residual is defined as: r = |f measured - f twin simulation|, where f twin simulation is the theoretical value calculated by the digital twin model based on the current operating parameters and physical mechanism. The larger the absolute value of the residual, the further the actual state deviates from the physical law.

[0060] Since a total of 39 dimensions of measured features were extracted, the twin residual features are also 39 dimensions. Combining the two forms a feature vector with a dimension of 78.

[0061] The fault prediction module divides historical fault data into sets based on fault type using feature vectors. For each set, it calculates the causal sensitivity matrix S and mutual information matrix M. Based on the comprehensive contribution, it determines the first discriminant feature corresponding to each fault. If at least two faults have the same first discriminant feature, it marks the first discriminant feature as a shared first feature. It trains a Local Sensitive Hash (LSH) model using the twin residuals of the fault to obtain a second discriminant model. It calculates the deviation of each dimension of the actual feature vector from a specified threshold, finds the feature with the largest deviation, and identifies it as a first discriminant feature but not a shared first feature. If the feature is a shared first feature, it directly predicts the fault corresponding to the first discriminant feature. If the feature is a shared first feature, it inputs the actual twin residual vector into the second discriminant model to predict the corresponding fault. It also inputs the actual twin residual vector into the trained GCN network to output the probability of different levels of the same fault. The fault levels can be divided into light, medium, and severe.

[0062] The specific steps for determining the first discriminant feature based on the overall contribution are as follows:

[0063] Based on the physical equations of the digital twin model, the causal dependency between features and faults is defined as directed edges, and a causal sensitivity matrix S is generated, where Si,j represents the physical sensitivity of feature fi changes to fault cj.

[0064] For the feature vector sequence in the fault set, the mutual information I(fi;cj) between feature fi and fault label cj is calculated according to the time window, and a dynamic weight factor wt is added to assign different weights to the data at different time points. The specific formula is wt=a^t, which is used to capture the time-varying characteristics of feature contribution.

[0065] Record the mutual information value of each feature to the fault in different time windows, and generate a mutual information matrix M;

[0066] The features in the same fault set are sorted according to the comprehensive contribution formula CMIC(fi)=ɑ*mean(Si,j)+(1-ɑ)*mean(Mi,j), and the feature with the highest contribution is selected as the first discriminant feature. Here, ɑ is the mechanism-data balance coefficient, which can be taken as 0.6.

[0067] The specific steps for training a GCN network using twin residuals are as follows:

[0068] For the same type of fault, a 39-dimensional twin residual vector is extracted from historical monitoring data to form a dataset. Equipment experts label each data sample as light, medium, or severe according to the degree of impact of the fault on production. For example, a slight leak that does not affect production is light, a moderate leak that requires a warning is medium, and a severe leak that causes shutdown is severe. Finally, a training set containing residual vectors and corresponding level labels is obtained.

[0069] Each residual vector is considered as a node in the graph, and the feature of the node is the residual vector. The similarity between any two nodes is calculated, and the difference can be judged by the angle between the vectors. When the similarity exceeds a set threshold, an edge is connected between the two nodes to form a graph structure that represents the association of fault states. This structure can capture the spatial distribution pattern of data of different levels of the same type of fault. For example, the residual vectors of severe faults often cluster in specific areas of the graph.

[0070] A GCN network suitable for processing graph data is constructed: The input layer directly receives the 39-dimensional residual vector of each node as the initial feature; the hidden layer, through multi-layer graph convolution operations, allows each node to not only learn its own residual features, but also fuse the correlation features of neighboring nodes, enhancing the discrimination of the same level; the output layer outputs three values ​​through a fully connected layer, corresponding to the probabilities of "light", "medium", and "heavy", which are converted into probability values ​​between 0 and 1 by the Softmax function, and the sum is 1.

[0071] The loss function uses cross-entropy loss to measure the difference between the model's predicted fault level probability and the manually labeled fault level, making the model output closer to the true level. The Adam optimizer is used to adjust the network parameters. The dataset is divided into training and validation sets in a 7:3 ratio to avoid overfitting. During training, the model performance is evaluated by the classification accuracy of the validation set. If the performance does not improve after several consecutive rounds, training will automatically stop.

[0072] Once the system predicts the specific fault type, it inputs the real-time collected twin residual vectors into the trained GCN network. The model outputs three levels of probability. For example, if the output probability is [0.2, 0.7, 0.1], the current fault level is determined to be the middle level with the highest probability. Combined with the corresponding processing strategy for this level, it provides accurate decision support for operation and maintenance personnel.

[0073] The parameter adjustment module uses the PPO algorithm to adjust parameters in real time for different predicted fault levels. The specific operation steps are as follows:

[0074] The input to the PPO algorithm is a multi-dimensional state vector S, specifically the physical quantity of the fault feature Qpred, including but not limited to the leakage rate of the seal, the local temperature residual of the heating plate, the main frequency offset of the bearing vibration, the current temperature rise rate vT, the pressure setpoint P, and the lamination time tlam. The risk probability index is the probability of different levels of faults, obtained from the GCN network. The production index parameters are the real-time yield rate Ygood and the actual production cycle Tcycle. The environmental variable parameters are the ambient temperature and ambient humidity.

[0075] The design of the continuous adjustable parameter space A includes the following process parameter adjustments: △vT∈[vTth1,vTth2], △P∈[Pth1,Pth2], △tlam∈[tlamth1,tlamth2];

[0076] The reward function that balances failure risk and production efficiency is constructed based on the formula R=c*Rrisk+d*Refficiency. The failure risk reward is Rrisk=-((Qpred-Qsafe) / (Qmax_risk-Qsafe))^2, where Qsafe is the corresponding failure physical quantity threshold, Qmax_risk is the severe failure threshold, and the reward increases as the failure rate decreases. The production efficiency reward is calculated based on the formula Refficiency=λ1*Ygood-λ2*((Tcycle-Tnominal) / (Tmax_cycle-Tnominal))^2, which includes a positive reward for yield rate and a penalty for cycle deviation. λ1 and λ2 are weighting coefficients, Tnominal is the production cycle threshold, and Tmax_cycle is the maximum production cycle threshold.

[0077] The edge device collects state St, action At, and reward Rt in real time to form a trajectory {St,At,Rt,St+1}, and triggers a policy update every 200 steps.

[0078] The advantage function At = Rt + γ * V(St+1) - V(St) is calculated using the generalized advantage estimation (GAE), where γ is the discount factor.

[0079] The objective function is optimized by pruning importance sampling: L=Et[min(rt(A)*At,clip(rt(A),1-ε,1+ε)*At)], where rt(A) is the probability ratio of the new and old strategies, and ε is the pruning parameter, which can be 0.2. The Actor-Critic network parameters are updated synchronously.

[0080] When a fault occurs, the PPO strategy is activated, and the optimal action is output every 10 seconds and sent to the PLC for execution via the industrial protocol.

[0081] Before adjustment, simulate the impact of parameter changes on lamination quality in the digital twin model, such as whether a decrease in temperature leads to insufficient curing of the adhesive film. Execute the adjustment after verification.

[0082] If three consecutive adjustments fail to meet expectations (i.e., Qpred does not decrease or the yield rate is less than 98%), the most recent valid parameter will be automatically restored and an alert will be triggered.

[0083] Before using the PPO algorithm, the hyperparameters of the Actor and Critic layers need to be set, with the specific requirements as follows:

[0084] The hyperparameters of the Actor layer specifically include an input dimension of S, 2 hidden layers, 256 and 128 hidden layer neurons, ReLU activation function, 3 output dimensions, a learning rate of 3e-4, and a batch size of 128.

[0085] The hyperparameters of the Critic layer specifically include the input dimension being S, the number of hidden layers being 2, the number of hidden layer neurons being 256 and 128, the activation function being ReLU, the output being a scalar function V(s), and the learning rate being 3e-4.

[0086] Some of the data in the above formulas are numerical calculations with dimensions removed, and the contents not described in detail in this specification are all prior art known to those skilled in the art.

[0087] The above embodiments are only used to illustrate the technical methods of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical methods of the present invention without departing from the spirit and scope of the technical methods of the present invention.

Claims

1. A digital twin-driven photovoltaic laminator operation monitoring and fault prediction system, characterized in that, include: The sensor monitoring module deploys an infrared array temperature sensor, a high-frequency piezoelectric pressure sensor, a MEMS triaxial accelerometer, and Modbus TCP for integrating PLC data on the photovoltaic laminator. The feature extraction module is used to extract temperature field features, pressure and vibration signal features, PLC operating condition parameter features, and twin residual features from historical fault data of photovoltaic laminators. The fault prediction module divides the feature vectors of historical fault data into a set of feature data for the same fault, calculates the causal sensitivity matrix S and the mutual information matrix M, determines the first discriminant feature corresponding to each fault based on the comprehensive contribution, calculates the deviation of each dimension of the actual feature vector from a specified threshold, finds the feature corresponding to the largest deviation, uses this feature as a comparison feature, and matches the comparison feature with the first discriminant features of all faults to predict the fault. The specific steps for determining the first discriminant feature based on the comprehensive contribution are as follows: based on the physical equations of the digital twin model, the causal dependency relationship between the feature and the fault is defined, and the causal sensitivity matrix S is generated, where Si,j represents the physical sensitivity of the change in feature fi to the fault cj. For the feature vector sequence in the fault set, calculate the mutual information I(fi;cj) between feature fi and fault label cj according to the time window, and add a dynamic weight factor wt to assign different weights to the data at different time points. The specific formula is wt=a^t, where a∈(0,1), and t is the time from the current time point; record the mutual information value of each feature to the fault in different time windows to generate a mutual information matrix M; sort the features in the same fault set according to the comprehensive contribution formula CMIC(fi)=ɑ*mean(Si,j)+(1-ɑ)*mean(Mi,j), and select the one with the highest contribution as the first discriminant feature, where ɑ is the mechanism-data balance coefficient; The parameter adjustment module uses the PPO algorithm to adjust the parameters in real time for different levels of predicted faults.

2. The digital twin-driven photovoltaic laminator operation monitoring and fault prediction system according to claim 1, characterized in that, The temperature field features include the 10-dimensional temperature field PCA principal components extracted by PCA, the coordinates and area of ​​the largest anomaly region, the heating slope and overshoot, and the standard deviation of temperature fluctuation during the heat preservation stage; the pressure signal features include kurtosis, margin, rising edge slope, and the energy proportion of the eight frequency bands obtained by three-layer wavelet decomposition; the vibration signal features include the STFT energy centroid frequency, the kurtosis of the first three IMFs after EMD decomposition, the triaxial RMS value, and the triaxial peak factor; the PLC operating condition parameter features include the mean speed, mean vacuum level, speed slope, vacuum level slope, number of temperature exceedances, and number of pressure exceedances; the twin residual features are the absolute values ​​of the differences between the simulation features and the real features of the digital twin model.

3. The digital twin-driven photovoltaic laminator operation monitoring and fault prediction system according to claim 1, characterized in that, If at least two faults have the same first discriminant feature, then the first discriminant feature is marked as a shared first feature. For the fault set marked as a shared first feature, the obtained twin residual is used to train the Local Sensitive Hash (LSH) model to obtain the second discriminant model.

4. The digital twin-driven photovoltaic laminator operation monitoring and fault prediction system according to claim 1, characterized in that, If the comparison feature is the first discriminant feature but not the shared first feature, then the fault corresponding to the first discriminant feature is directly used as the predicted fault. If the comparison feature is the shared first feature, then the actual obtained twin residual vector is input into the second discriminant model corresponding to the fault of the shared first feature to predict the corresponding fault.

5. The digital twin-driven photovoltaic laminator operation monitoring and fault prediction system according to claim 1, characterized in that, After predicting the corresponding fault, the actual twin residual vector is input into the trained GCN network, which outputs the probability of different fault levels, which can be divided into light, medium and heavy faults.

6. The digital twin-driven photovoltaic laminator operation monitoring and fault prediction system according to claim 1, characterized in that, The specific steps for real-time parameter adjustment of predicted faults of different levels using the PPO algorithm are as follows: Obtain the input multidimensional state vector S, which specifically includes the fault characteristic physical quantity Qpred, the current temperature rise rate vT, the pressure setpoint P, the lamination time tlam, the probability of different levels of fault, the real-time yield rate Ygood, the actual production cycle Tcycle, the ambient temperature, and the ambient humidity. Design a continuous adjustable parameter space A, specifically including the following process parameter adjustment amounts: △vT∈[vTth1,vTth2], △P∈[Pth1,Pth2], △tlam∈[tlamth1,tlamth2]; The fault risk reward is calculated according to the formula Rrisk=-((Qpred-Qsafe) / (Qmax_risk-Qsafe))^2, where Qsafe is the corresponding fault physical quantity threshold and Qmax_risk is the severe fault threshold. The production efficiency bonus is calculated using the formula Refficiency=λ1*Ygood-λ2*((Tcycle-Tnominal) / (Tmax_cycle-Tnominal))^2, where λ1 and λ2 are weighting coefficients, Tnominal is the production cycle threshold, and Tmax_cycle is the maximum production cycle threshold. A reward function balancing failure risk and production efficiency is constructed based on the formula R=c*Rrisk+(1-c)*Refficiency, where c is an adjustment factor; The edge device collects state St, action At, and reward Rt in real time to form a trajectory {St,At,Rt,St+1}, and triggers a policy update every 200 steps. The advantage function At = Rt + γ * V(St+1) - V(St) is calculated using generalized advantage estimation, where γ is the discount factor. The objective function is optimized by pruning importance sampling: L=Et[min(rt(A)*At,clip(rt(A),1-ε,1+ε)*At)], where rt(A) is the probability ratio of the new and old policies, and the Actor-Critic network parameters are updated synchronously, where ε is the pruning parameter; When a fault occurs, the PPO strategy is activated, and the optimal action is output every 10 seconds and sent to the PLC for execution via the industrial protocol.

7. The digital twin-driven photovoltaic laminator operation monitoring and fault prediction system according to claim 1, characterized in that, Before parameter adjustment, the impact of parameter changes on lamination quality is simulated in the digital twin model. If the expected results are not achieved after three consecutive adjustments, the most recent valid parameters are automatically restored and an early warning is triggered.

8. The digital twin-driven photovoltaic laminator operation monitoring and fault prediction system according to claim 7, characterized in that, Failure to meet expectations specifically includes Qpred not decreasing or a yield rate of less than 98%.

Images