ERP data collaborative analysis method for photovoltaic product defect detection

Abstract

The invention relates to the technical field of photovoltaic product detection, and discloses an ERP data collaborative analysis method for photovoltaic product defect detection, and the method comprises the steps: obtaining the production process data of an ERP system and the defect detection data of a defect detection device; the two types of data are arranged in a time and dimension structured mode, multi-level feature data are obtained through time coding and grouping feature extraction, and production-defect collaborative feature vectors are generated after collaborative attention fusion, rule correction and feature enhancement; and based on the vector, a detection analysis result containing defect types and position coordinates is generated through classification prediction and spatial mapping. Wherein the production process data comprises raw material batches, lamination temperature and the like, and the defect detection data comprises an electroluminescent image gray value, hot spot coordinates and the like. According to the method, through multi-source data cross-domain association and deep fusion, accurate detection and positioning of photovoltaic product defects are realized, the accuracy and timeliness of defect analysis are improved, and a scientific basis is provided for production process optimization.

Description

Technical Field

[0001] The present invention relates to the technical field of photovoltaic product detection, and in particular to an ERP data collaborative analysis method for photovoltaic product defect detection. Background Art

[0002] In the photovoltaic industry's manufacturing process, the quality of photovoltaic products directly impacts their power generation efficiency, service life, and market competitiveness. Defect detection is a critical step in ensuring product quality. Traditional photovoltaic product defect detection methods often suffer from isolated data and a single analysis dimension, making it difficult to conduct in-depth correlation analysis between production process data and defect detection data.

[0003] From a production process data perspective, photovoltaic production involves multiple key parameters, including raw material batches, lamination temperature, soldering process, and packaging pressure. Fluctuations in these parameters across time and batches can have a direct impact on product quality. However, traditional methods typically monitor only a single parameter or a specific production step, failing to build a comprehensive production process data model across time and parameter categories. This makes it difficult to accurately trace defects back to their source.

[0004] In terms of defect detection data, indicators such as electroluminescence image grayscale values, hot spot distribution coordinates, and hidden crack length can intuitively reflect the defect characteristics of photovoltaic modules. However, existing technologies often separate and analyze this detection data from production process data, lacking cross-domain correlation technology. For example, it is impossible to effectively establish a correlation between temperature anomalies during production and hot spot defects in subsequent inspections. This makes defect analysis superficial and makes it difficult to deeply explore the inherent connection between production processes and product defects.

[0005] Furthermore, with the large-scale development of the photovoltaic industry, production and inspection data has experienced explosive growth. Traditional analysis methods based on rules or simple statistics are unable to cope with the demands of processing large-scale, multi-dimensional data. They are unable to efficiently extract features and perform deep pattern recognition from data, resulting in defect detection accuracy and timeliness that are difficult to meet modern production requirements. The integration of production process data and defect detection data to build a cross-domain collaborative analysis model to accurately locate and analyze the causes of photovoltaic product defects has become a key issue that needs to be addressed in the field of photovoltaic inspection technology. Summary of the Invention

[0006] The purpose of the present invention is to provide an ERP data collaborative analysis method for photovoltaic product defect detection to solve the problems raised in the above background technology.

[0007] To achieve the above objectives, the present invention provides the following technical solution: an ERP data collaborative analysis method for photovoltaic product defect detection, the method comprising:

[0008] Obtaining photovoltaic production process data of multiple inspection batches within a predetermined period collected by the ERP system, and photovoltaic module defect detection data of the multiple inspection batches collected by the defect detection device;

[0009] performing feature extraction and cross-domain correlation on the production process data of the multiple inspection batches and the defect detection data of the multiple inspection batches to obtain a production-defect collaborative feature vector;

[0010] generating defect detection analysis results based on the production-defect collaborative feature vector;

[0011] The process of extracting features and performing cross-domain correlation on the production process data of the multiple inspection batches and the defect detection data of the multiple inspection batches to obtain a production-defect collaborative feature vector includes:

[0012] Arranging the production process data of the multiple inspection batches according to the time dimension and the parameter category dimension into production process structured input data, and arranging the defect detection data of the multiple inspection batches according to the time dimension and the detection index dimension into defect detection structured input data;

[0013] Performing time coding and group feature extraction on the production process structured input data and the defect detection structured input data respectively to obtain production process multi-level feature data and defect detection multi-level feature data;

[0014] The multi-level feature data of the production process and the multi-level feature data of the defect detection are fused across data sources to obtain the production-defect collaborative feature vector.

[0015] Preferably, the production process data includes raw material batch numbers, lamination temperature records, welding process parameters and packaging pressure values.

[0016] Preferably, the defect detection data includes electroluminescent image grayscale value, hot spot distribution area coordinates and hidden crack length parameters.

[0017] Preferably, performing time coding and group feature extraction on the production process structured input data and the defect detection structured input data respectively to obtain production process multi-level feature data and defect detection multi-level feature data, including:

[0018] Inputting the production process structured input data into a group feature extraction network based on a sliding time window to obtain multi-level feature data of the production process;

[0019] The defect detection structured input data is input into a feature extraction network based on temporal association coding to obtain the defect detection multi-level feature data.

[0020] Preferably, cross-data source fusion of the production process multi-level feature data and the defect detection multi-level feature data to obtain the production-defect collaborative feature vector includes:

[0021] Inputting the production process multi-level feature data and the defect detection multi-level feature data into a collaborative attention fusion network to obtain production-defect association weight data;

[0022] Inputting the production-defect association weight data into a rule-based constraint correction module to obtain optimized association weight data;

[0023] Inputting the production process multi-level feature data and the defect detection multi-level feature data into a feature enhancement network to obtain enhanced production process multi-level feature data and enhanced defect detection multi-level feature data respectively;

[0024] The enhanced production process multi-level feature data and the enhanced defect detection multi-level feature data are jointly encoded to obtain the production-defect collaborative feature vector.

[0025] Preferably, the multi-level feature data of the production process and the multi-level feature data of the defect detection are input into a collaborative attention fusion network to obtain production-defect association weight data, including:

[0026] The feature similarity vector of the multi-level feature data of the production process and the feature difference vector of the multi-level feature data of the defect detection are calculated, and the production-defect association weight data are generated through a vector superposition operation.

[0027] Preferably, the production process multi-level feature data and the defect detection multi-level feature data are respectively input into a feature enhancement network to obtain enhanced production process multi-level feature data and enhanced defect detection multi-level feature data, including:

[0028] Performing dimension compression on the multi-level feature data of the production process to obtain a production process feature compression vector;

[0029] Performing a dimensionality reduction operation after performing feature fusion on the optimized association weight data and the production process feature compression vector to obtain enhanced multi-level feature data of the production process;

[0030] After feature normalization is performed on the defect detection multi-level feature data, weighted modulation is performed on the data with the optimized associated weight data to obtain enhanced defect detection multi-level feature data.

[0031] Preferably, generating defect detection analysis results based on the production-defect collaborative feature vector includes:

[0032] Inputting the production-defect collaborative feature vector into a classification prediction model to obtain a defect type identification;

[0033] The defect detection analysis result including the defect location coordinates is generated according to the defect type identification.

[0034] Preferably, inputting the production-defect collaborative feature vector into a classification prediction model to obtain a defect type identification includes:

[0035] Performing feature space mapping on the production-defect collaborative feature vector to obtain a high-dimensional feature encoding vector;

[0036] The high-dimensional feature encoding vector is input into a multi-layer classification network to perform category probability calculation, and the defect type identifier associated with the defect type is output.

[0037] Preferably, generating the defect detection analysis result including the defect location coordinates according to the defect type identifier includes:

[0038] Performing pattern matching on the defect type identifier and the historical defect database to obtain defect location reference parameters;

[0039] The defect type identifier is spatially mapped based on the defect location reference parameter to generate the defect position coordinates with area markings.

[0040] Compared with the prior art, the present invention has the following beneficial effects:

[0041] This method extracts features from production process data collected by the ERP system and defect detection data collected by defect detection devices, performs cross-domain correlation on the two types of data, and generates production-defect collaborative feature vectors, thereby generating defect detection analysis results. This method effectively integrates multi-dimensional data from production and testing, breaking down data silos and establishing deep correlations between production process parameters and product defects, providing a comprehensive and accurate analytical basis for defect tracing.

[0042] During feature extraction, production process data and defect detection data are structured and arranged according to the time dimension and parameter / detection indicator dimension, respectively. Time coding and group feature extraction techniques enable in-depth exploration of both types of data at different time scales and feature levels. A group feature extraction network based on a sliding time window captures the dynamic changes in production process data over time series, while a feature extraction network using temporal association coding accurately extracts the temporal dependency features of defect detection data, thereby enhancing the richness and accuracy of feature representation.

[0043] During the cross-data source fusion process, the collaborative attention fusion network generates association weights by calculating feature similarity and difference vectors. This weights are then optimized using a rule-based correction module, enabling adaptive fusion of production process features and defect detection features based on their actual degree of correlation. The feature enhancement network further optimizes both types of features through operations such as dimensionality compression and restoration, feature normalization, and weighted modulation. This enhances the expressive power of key features, effectively suppresses interference from noise and irrelevant information, and thus improves the quality of the collaborative feature vector.

[0044] During the defect analysis phase, the classification prediction model achieves high-precision identification of defect types through feature space mapping and multi-layer classification networks. Combined with pattern matching and spatial coordinate mapping of the historical defect database, it can generate defect detection and analysis results containing precise location coordinates, providing direct guidance for defect location and process optimization during the production process. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 This is a working principle diagram of the ERP data collaborative analysis method for photovoltaic product defect detection according to the present invention;

[0046] Figure 2 Design a flow chart for cross-domain fusion of production-defect features;

[0047] Figure 3 Design drawings generated from defect detection analysis results;

[0048] Figure 4 Design diagram of the feature extraction network for the production process. DETAILED DESCRIPTION

[0049] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0050] See also Figures 1-4 The ERP data collaborative analysis method for photovoltaic product defect detection according to the present invention is specifically implemented in the following steps:

[0051] Acquire photovoltaic production process data collected by the ERP system for multiple inspection batches within a predetermined period, as well as photovoltaic module defect detection data collected by the defect detection device for these multiple inspection batches. Production process data includes raw material batch numbers, lamination temperature records, welding process parameters, and packaging pressure values; defect detection data includes electroluminescence image grayscale values, hot spot distribution area coordinates, and hidden crack length parameters.

[0052] Feature extraction and cross-domain correlation are performed on the production process data and defect detection data of the multiple inspection batches to obtain a production-defect collaborative feature vector, specifically including:

[0053] a. Arrange the production process data of multiple inspection batches according to the time dimension and parameter category dimension as production process structured input data. At the same time, arrange the defect detection data of multiple inspection batches according to the time dimension and detection indicator dimension as defect detection structured input data.

[0054] b. Perform time coding and group feature extraction on the production process structured input data and the defect detection structured input data respectively to obtain multi-level feature data of the production process and multi-level feature data of the defect detection.

[0055] c. Perform cross-data source fusion on the multi-level feature data of the production process and the multi-level feature data of defect detection to obtain the production-defect collaborative feature vector.

[0056] Based on the production-defect collaborative feature vector, a defect detection analysis result is generated.

[0057] The technical solution of the present invention is further described in detail below with reference to specific embodiments.

[0058] Example 1:

[0059] During the photovoltaic product production process, the specific composition and processing of the production process data and defect detection data involved in this embodiment are as follows:

[0060] The collection and processing of production process data must cover key processes in PV module manufacturing. Raw material batch numbers serve as key identifiers for tracing raw material sources. The ERP system records data such as supplier information, arrival time, and inspection reports for each batch of raw materials. During data processing, the raw material batch numbers are encoded and mapped into multidimensional feature vectors to facilitate subsequent analysis of the correlation between different batches of raw materials and product defects. For example, one-hot encoding or hash encoding can be used to convert discrete batch numbers into continuous numerical features that can be recognized and processed by computer systems.

[0061] The acquisition frequency of the lamination temperature record is set to once per minute, continuously recording the temperature changes throughout the lamination process. In the data preprocessing stage, the temperature data is first smoothed and filtered to remove random noise interference; then the characteristic parameters of the temperature curve are extracted, such as peak temperature, heating rate, holding time, etc. For abnormal temperature curves, the occurrence time and fluctuation amplitude of the temperature anomaly are identified by comparing the temperature fluctuation threshold range established by historical normal production data. For example, when the temperature of a batch of lamination is monitored to fluctuate by more than ±5°C during the holding stage and the duration exceeds 30 seconds, the system automatically marks it as an abnormal temperature event.

[0062] The collection of welding process parameters involves the coordinated operation of multiple sensors. A temperature sensor monitors the welding head temperature in real time, a pressure sensor records the welding pressure, and a time controller precisely records the welding duration. These parameters are synchronously collected with millisecond-level accuracy and transmitted to the ERP system. During data processing, time series analysis is used to correlate and integrate the temperature, pressure, and time parameters of the same weld point to form a multidimensional feature vector. Furthermore, the dynamic changes in welding parameters are modeled, such as analyzing the slope of the temperature-pressure curve during welding, to capture potential trends in weld quality.

[0063] Package pressure values ​​are collected using a distributed pressure sensor array, with multiple pressure measurement points placed at key locations in the packaging mold to ensure comprehensive pressure distribution data. The raw pressure data collected is subject to measurement errors and requires Kalman filtering. First, a state-space model of pressure variation is established, treating the pressure value as the system state variable and the sensor measurement value as the observation variable. Then, through an iterative prediction-update process, the pressure estimate is continuously optimized to reduce the impact of random errors. For example, if the pressure value at a certain measurement point experiences a sudden change, the Kalman filter combines historical data and changes at adjacent measurement points to smooth the sudden change, improving data reliability.

[0064] The processing of defect detection data requires integrating the characteristics of multiple detection technologies. A multi-scale feature extraction method is used to analyze the grayscale values ​​of electroluminescence (EL) images. First, the original EL image is decomposed using a Gaussian pyramid to obtain image layers of varying resolutions. Then, the local binary pattern (LBP) operator is applied at each scale to extract texture features, and the grayscale co-occurrence matrix is ​​calculated to obtain the statistical characteristics of the image. Principal component analysis (PCA) is used to reduce the dimensionality of the extracted grayscale features, removing redundant information and retaining the principal components that best reflect the defect characteristics. For example, when identifying hidden cracks in cell wafers, PCA analysis revealed that the first three principal components can explain over 95% of the image variation, effectively reducing the computational complexity of subsequent analysis.

[0065] The coordinates of the hot spot distribution area are determined using a segmentation algorithm based on region growing. First, the seed point selection rules are set, and the point with the largest temperature gradient change is selected as the seed point; then the region growth criteria are defined, including the temperature similarity threshold and the spatial continuity condition; during the growth process, adjacent pixels that meet the conditions are continuously merged to form a complete hot spot area. In order to improve the accuracy of coordinate positioning, sub-pixel positioning technology is introduced. By fitting the Gaussian curve of the edge of the hot spot area, the positioning accuracy is improved from the pixel level to the sub-pixel level. For example, for tiny hot spots with a diameter of less than 10 pixels, sub-pixel positioning technology can improve the coordinate accuracy to 0.1 pixel level.

[0066] The measurement of hidden crack length parameters utilizes a method that combines deep learning with traditional image processing. First, a U-Net network is used to semantically segment the hidden crack region in the EL image, obtaining a binary mask of the hidden crack. Morphological processing is then performed on the binary image to fill holes and remove noise. Finally, a chain code tracking algorithm is used to extract the outline of the hidden crack and calculate its pixel length. To eliminate the effects of image distortion on the measurement results, camera calibration is performed during the preprocessing stage to establish a mapping between image coordinates and actual physical coordinates. For example, a checkerboard calibration plate is used to obtain the camera's intrinsic and extrinsic parameters, accurately converting the hidden crack length in the image to its actual physical length.

[0067] During the data fusion phase, aligning the timestamps of production process data and defect detection data is a key step. Because different data sources have different sampling frequencies, linear interpolation is used to resample the data. For example, the ERP system collects production parameters once a minute, while defect detection data is sampled once a second. Linear interpolation is used to resample the production parameters onto the time scale of the defect detection data. During the alignment process, a time window mechanism is introduced, allowing a ±5-second time tolerance to accommodate time synchronization errors in actual production.

[0068] When establishing a correlation matrix between production parameters and defect characteristics, mutual information theory is used to calculate the correlation between the two. For continuous production parameters (such as lamination temperature) and discrete defect types (such as bubbles and delamination), the maximum information coefficient (MIC) is used for correlation analysis. For two continuous variables (such as welding temperature and crack length), the distance correlation coefficient (DistanceCorrelation) is used to measure their nonlinear correlation. Each element of the correlation matrix represents the strength of the association between a production parameter and a defect characteristic, with values ​​ranging from [0 to 1], where larger values ​​indicate stronger correlation. For example, the correlation strength between lamination temperature and bubble defects was calculated to be 0.82, indicating a strong correlation between the two.

[0069] For data storage, a combination of time-series databases and graph databases is employed. Time-series databases (such as InfluxDB) store time series information for production process data and defect detection data, leveraging their high write performance and time window query advantages to quickly retrieve data within a specific time period. Graph databases (such as Neo4j) store the relationships between production parameters and defect types, using nodes to represent entities (such as lamination temperature and hidden crack defects) and edges to represent relationships (such as causal relationships and correlations). This facilitates complex associative queries and reasoning analysis. For example, if a batch of products is found to have a large number of hot spot defects, the graph database can quickly trace the associated production parameters (such as abnormal packaging pressure values), supporting root cause analysis of quality issues.

[0070] In actual production environments, data acquisition systems must be highly reliable and resistant to interference. Production process data acquisition terminals utilize industrial-grade PLC controllers, communicating with field equipment via the Modbus / TCP protocol to ensure stable data transmission. Defect detection data acquisition equipment is equipped with redundant power supplies and data buffering modules, enabling temporary storage of inspection data during network outages and automatic retransmission upon network restoration. Furthermore, all acquisition equipment is equipped with electromagnetic shielding to minimize the impact of electromagnetic interference in industrial environments on data quality.

[0071] Encrypted tunneling technology is used during data transmission to ensure data security. Edge computing nodes at production sites desensitize collected data before sending it to cloud servers via the MQTT protocol. Data is encrypted end-to-end using the TLS 1.3 encryption protocol to prevent data theft or tampering during transmission. The cloud server performs integrity checks on received data, calculating a hash value and comparing it with the sender's hash value to ensure the data has not been modified during transmission.

[0072] To address the surge in data volume associated with expanded production, the system adopts a distributed architecture. The data collection layer deploys multiple edge computing nodes to process local data, reducing data transmission pressure. The data processing layer uses a Spark cluster for distributed computing, enabling parallel processing of large-scale production and defect data. The data storage layer utilizes a distributed file system (such as Ceph) and a distributed database cluster to ensure scalability and fault tolerance. For example, if the number of production sites expands from three to ten, data processing needs can be met by simply deploying new edge computing nodes in the corresponding areas and adjusting cluster resource allocation, without requiring a major overhaul of the system architecture.

[0073] During the data preprocessing stage, missing values ​​are handled using a predictive filling method based on a gradient boosting tree (XGBoost). For missing values ​​in the laminate temperature records, other production parameters from the same batch (such as packaging pressure and soldering time) are used as input features to train an XGBoost model to predict the missing temperature values. Compared with traditional mean filling or linear interpolation methods, this method can better capture nonlinear relationships between data and improve filling accuracy. Outlier detection uses the Isolation Forest algorithm, which constructs a random binary tree and calculates the anomaly score of data points, enabling rapid identification of abnormal fluctuations in production parameters. For example, if the soldering temperature of a batch is detected to have suddenly increased to three times the normal range, the system automatically marks it as an outlier and triggers an alarm.

[0074] During feature engineering, when constructing features for production process data, in addition to extracting statistical features (such as mean and variance) and trend features (such as slope and intercept), time window features are also introduced. For example, the moving average of the laminate temperature change rate for the previous 10 time points is calculated as a new feature for the current moment. For defect detection data, a multi-scale feature fusion method is used to weightedly fuse features extracted at different resolutions through an attention mechanism to highlight key defect characteristics. During the feature selection stage, a recursive feature elimination (RFE) algorithm is combined with a random forest model to screen out the feature subset that contributes most to defect type prediction, reducing feature dimensionality and improving model training efficiency.

[0075] During model training, various regularization techniques were employed to prevent overfitting. In the classification prediction model, L2 regularization constraints were applied to the weight parameters of the fully connected layers to limit model complexity. Dropout was also used to randomly ignore some neurons, enhancing the model's generalization capabilities. Training data was divided using time series cross-validation, chronologically partitioning historical data into multiple training and validation sets to ensure the model maintains good performance on data from different time periods. For example, data from January to June 2023 was used as the training set, data from July to August as the validation set, and data from September as the test set to assess the model's temporal stability.

[0076] After system deployment, a comprehensive monitoring and maintenance mechanism was established. The system monitors the operating status of data acquisition equipment in real time, using a heartbeat packet mechanism to check device online status. If an EL detector fails to send data for 10 consecutive minutes, a fault alarm is automatically triggered. Model performance is regularly evaluated, using Cumulative Gains Curve and Lift Charts to analyze the model's predictions. A model update process is triggered if the model accuracy drops by more than 5%. Furthermore, a production-defect association knowledge base is established, continuously accumulating new association rules and cases. Knowledge graph technology is used to visualize and apply knowledge, providing continuously optimized decision support for photovoltaic product quality control.

[0077] Example 2:

[0078] The feature extraction process of this embodiment achieves deep feature mining of production process data and defect detection data through time encoding and grouping network architecture. For structured input data of the production process, a grouping feature extraction network based on a sliding time window adopts a multi-scale time granularity analysis strategy. The network first divides the input data into slice units according to process segments, such as the raw material inspection segment, the lamination segment, and the welding segment. Each slice unit contains all parameter time series data within that process segment. For each slice unit, a sliding time window of different lengths (such as 5 minutes, 15 minutes, and 1 hour) is set, and the window slides across the time series data with a fixed step size (such as 1 minute).

[0079] Within each time window, three types of features are extracted: statistical features, trend features, and temporal relationship features. Statistical features calculate descriptive statistics such as the mean, variance, skewness, and kurtosis of the parameters within the window. Trend features use linear regression to fit the data trends within the window, obtaining parameters such as the slope and intercept. Temporal relationship features use the autocorrelation function to analyze the periodicity of the parameters within the window. For example, for laminate temperature records within a 15-minute time window, the standard deviation of the temperature values ​​is calculated to assess temperature stability, the heating rate is obtained through linear regression, and the autocorrelation function is used to detect the presence of periodic temperature fluctuations.

[0080] To capture the temporal dependencies between different process stages, the network introduces a gated recurrent unit (GRU) to perform temporal encoding on the feature vectors of each slice unit. The GRU selectively memorizes and forgets historical information through an update gate and reset gate mechanism, effectively processing dependencies in long-sequence data. For example, fluctuations in raw material batch parameters may not affect product quality until the subsequent lamination process. The GRU can capture this delayed effect across process stages. The multi-level feature data of the production process output by the network includes micro-features (such as short-term fluctuations in individual process parameters), meso-features (such as coordinated changes between parameters within a process stage), and macro-features (such as temporal dependencies across process stages).

[0081] For structured defect detection input data, a feature extraction network based on temporal correlation coding employs a multimodal fusion strategy. The network first performs feature preprocessing on the EL image grayscale values, hot spot distribution coordinates, and hidden crack length parameters. For the EL image grayscale values, Gaussian pyramid decomposition is used to convert the image into a multi-layer representation of varying resolutions. Local binary pattern (LBP) is applied to each layer to extract texture features, followed by principal component analysis (PCA) for dimensionality reduction. For the hot spot distribution coordinates, a spatial pyramid is constructed, dividing the PV module surface into grid regions at different levels. Within each region, the frequency and area percentage of hot spots are analyzed.

[0082] During the temporal correlation encoding stage, the network employs a Transformer architecture to capture the evolution of defects across different inspection batches. The preprocessed feature sequence for each inspection batch is used as input, and a self-attention mechanism is used to calculate the temporal correlation weights between features. For example, a hidden crack defect in the current batch may be associated with abnormal welding process parameters in previous batches. The self-attention mechanism automatically identifies these cross-batch correlation patterns. To enhance the representation of temporal information, the network incorporates positional encoding into the input features, embedding temporal information into the feature vector using sine and cosine functions.

[0083] The network also incorporates a contrastive learning mechanism. By constructing pairs of positive samples (representations of the same defect type in adjacent batches) and negative samples (samples of different defect types), the network maximizes the similarity between positive pairs and minimizes the similarity between negative pairs. This training approach enables the network to learn more discriminative temporal feature representations. The resulting multi-level feature data for defect detection includes spatial features (such as the geometric distribution of defects), temporal features (such as the evolution of defects), and semantic features (such as the inherent correlations between defect types).

[0084] During the feature fusion stage, multi-level feature data from the production process and multi-level feature data from defect detection are first mapped to the same feature space via a linear projection layer. To preserve feature information at different levels, a hierarchical fusion strategy is employed: microscopic features of the production process are fused with spatial features of defect detection, mesoscopic features are fused with temporal features, and macroscopic features are fused with semantic features. The fusion operation utilizes an attention mechanism, dynamically assigning weights based on feature importance. For example, when identifying hot spot defects, the short-term fluctuations in lamination temperature (microscopic features) may be more strongly correlated with the spatial distribution of hot spots (spatial features of defect detection). The attention mechanism will accordingly increase the fusion weight of these two sets of features.

[0085] To further enhance feature representation, a feature interaction module is introduced. This module uses a multi-layer perceptron (MLP) to learn nonlinear interactions between features and generate high-order feature combinations. For example, the combined features of welding process parameters and grayscale anomalies in EL images may have a higher discriminative power for identifying cold solder defects. During this interaction process, residual connections and layer normalization techniques are employed to ensure effective information transfer and training stability.

[0086] The resulting production-defect collaborative feature vectors contain multi-dimensional information: process parameter characteristics, defect spatial characteristics, temporal evolution characteristics, and the interactions between them. These feature vectors serve as the fundamental input for subsequent defect classification and location analysis, providing comprehensive and in-depth data support for photovoltaic product quality control. The entire feature extraction process, through the collaborative work of time encoding and grouping networks, effectively mines potential correlation patterns between production processes and defect detection data, laying the foundation for accurate defect tracing and quality prediction.

[0087] Example 3:

[0088] The cross-data source fusion process in this embodiment achieves deep correlation analysis between production process data and defect detection data through a collaborative attention mechanism and feature enhancement strategy. When the multi-level feature data of the production process and the multi-level feature data of the defect detection are input into the collaborative attention fusion network, feature alignment is first performed. The dynamic time warping (DTW) algorithm is used to align the temporal dimensions of the two feature sets, addressing the time offset caused by differences in production cycle time and inspection frequency. For example, temperature fluctuations in the lamination process may not manifest as hidden cracks in the cell until several hours later through EL inspection. DTW can find the optimal matching path between the two data sets while preserving the temporal characteristics.

[0089] The implementation of the collaborative attention mechanism consists of three steps: feature similarity calculation, difference analysis, and weight generation. In feature similarity calculation, a cosine similarity matrix is ​​used to calculate the similarity between production process feature vectors, forming a production feature association graph. For example, the edge weights of raw material batch parameters and welding temperature parameters in the association graph represent the degree of temporal coordinated variation between the two. Simultaneously, a difference matrix is ​​constructed by calculating the Euclidean distance between defect detection feature vectors to highlight the characteristic differences between different defect types. For example, the differences in the spatial distribution and grayscale characteristics of hot spot defects and hidden crack defects are quantified in the matrix.

[0090] The similarity matrix and the difference matrix are input into the attention calculation layer, where a softmax function is used to generate production-defect association weights. This weight not only reflects the strength of the association between features but also incorporates temporal information. For example, if an abnormal grayscale is detected in the EL image of a batch of products, the association weight automatically emphasizes the lamination temperature parameter features within the 24 hours prior to that moment, as the accumulated thermal stress during the lamination process may be a potential cause of the grayscale anomaly.

[0091] The rule-based correction module uses a hybrid reasoning mechanism. It encodes photovoltaic production domain knowledge into production rules, such as "If the lamination temperature continuously exceeds the threshold and the packaging pressure fluctuates by more than 15%, the probability of bubble defects increases by 0.7." A Bayesian network is also constructed to capture the uncertainty associations between parameters through prior probability distributions. For example, prolonged welding time does not necessarily lead to cold solder joints, but it will affect the probability of cold solder joints through conditional probability. When the association weights output by the collaborative attention network conflict with domain rules, the correction module initiates a conflict resolution algorithm, fusing the two information sources through evidence theory. For example, if the data-driven association weights show that welding pressure is strongly correlated with fragmentation defects, but domain knowledge indicates that this parameter primarily affects solder joint quality, the correction module will reduce the confidence level of this association weight.

[0092] During the feature enhancement phase, a variational autoencoder (VAE) is used to compress multi-level feature data from the production process. By introducing latent variables, VAEs are able to preserve the probabilistic distribution characteristics of the data during the dimensionality reduction process. For example, for the time series characteristics of lamination temperature, the VAE not only learns statistical properties such as the mean and variance of the temperature, but also captures the probabilistic distribution pattern of temperature fluctuations. The optimized association weight data is used as conditional input to the decoder, guiding the feature reconstruction process and ensuring that the compressed feature vectors are more focused on information dimensions that are strongly correlated with defects.

[0093] Adaptive Batch Normalization (ABL) is used to normalize multi-level feature data for defect detection. This method dynamically adjusts normalization parameters based on the characteristic distribution of different types of defects, avoiding the over-smoothing of abnormal features by traditional normalization methods. For example, for the edge features of tiny hidden crack defects, ABL retains the sharpness of their gradient changes while suppressing noise features. During the weighted modulation process, the optimized correlation weight data is decomposed into spatial weights and temporal weights, which act on the spatial and temporal dimensions of the defect features, respectively. For example, for the evolution of hot spot defects, the temporal weight will enhance the characteristic expression of early abnormal temperature points, while the spatial weight will highlight the grayscale change characteristics at the edge of the hot spot area.

[0094] During the joint feature encoding stage, a multi-scale feature fusion network is employed. The network consists of multiple parallel encoding branches, each processing feature information at a different granularity. The short-term branch uses a gated recurrent unit (GRU) to capture short-term dependencies between production-defect features, such as the correlation between welding parameter fluctuations and real-time EL image anomalies. The long-term branch employs a Transformer architecture to exploit long-term evolutionary patterns across batches, such as the cumulative impact of raw material batch differences on product defects. The outputs of each branch are fused using an attention aggregation mechanism, with attention weights dynamically determined by the temporal importance of the features and their classification contribution. For example, when identifying defects caused by the packaging process, the temporal features of the packaging pressure in the short-term branch are given higher weight; whereas, when analyzing raw material batch issues, the historical data features in the long-term branch are prioritized.

[0095] The entire cross-data source fusion process achieves adaptive learning of feature associations through a collaborative attention mechanism, improves model interpretability through rule-constrained correction, and highlights key defect patterns through feature enhancement. The resulting production-defect collaborative feature vector not only incorporates the statistical characteristics of the original data but also incorporates temporal dependencies and domain knowledge constraints, providing a more discriminative feature representation for subsequent defect classification and location. This fusion method, while preserving the original physical meaning of the data, uncovers deep production-defect association patterns, providing a more accurate data foundation for photovoltaic product quality control.

[0096] Example 4:

[0097] The defect detection analysis result generation process in this embodiment utilizes a classification prediction model and a spatial mapping algorithm to achieve defect type identification and precise location. When the production-defect collaborative feature vector is input into the classification prediction model, feature space mapping is first performed. This process utilizes a multi-layer residual network structure, with each layer comprising convolution operations, batch normalization, and a ReLU activation function. Convolution operations extract local patterns of features using convolution kernels of varying sizes. Batch normalization accelerates model convergence and enhances stability. The ReLU activation function introduces nonlinear transformations, enabling the model to learn complex feature relationships.

[0098] To capture multi-scale information about features, the network adopts a parallel multi-branch structure. Each branch processes the input features using convolution kernels of different sizes (e.g., 1×1, 3×3, and 5×5). The outputs of each branch are then fused through a concatenation operation. For example, a 1×1 convolution kernel extracts inter-channel relationships between features, a 3×3 convolution kernel captures local spatial features, and a 5×5 convolution kernel focuses on broader contextual information. This multi-scale feature extraction approach preserves both detailed feature information and global semantics.

[0099] During the feature mapping process, an attention mechanism is introduced to enhance key features. This approach combines channel attention and spatial attention. Channel attention calculates the importance weight of each channel through global average pooling and fully connected layers, while spatial attention generates a spatial weight map through convolution operations. These two attention mechanisms complement each other, enabling the model to focus on the feature regions most relevant to the defect type. For example, when identifying subtle cracks in solar cells, the attention mechanism enhances the feature representation of the crack edge in the EL image and suppresses interference from irrelevant background areas.

[0100] The multi-layer classification network uses softmax regression to calculate defect type probabilities. The network consists of multiple fully connected layers, each followed by a dropout layer to prevent overfitting. The dropout layer randomly ignores some neurons with a certain probability, forcing the model to learn more robust feature representations. For example, during training, a dropout rate of 0.5 means that each neuron has a 50% probability of being ignored in each iteration. This reduces the model's reliance on specific features and improves generalization.

[0101] To optimize the training process of the classification model, FocalLoss is used as the loss function. FocalLoss addresses the imbalance between positive and negative samples in photovoltaic defect detection by reducing the weight of easy samples. In photovoltaic production, the number of normal product samples is typically much larger than the number of defective samples. The traditional cross-entropy loss function causes the model to favor the majority class (normal samples). FocalLoss uses the modulation factor (1-pt)γ to reduce the loss contribution of high-confidence samples (easy samples), allowing the model to focus more on difficult-to-classify samples (defective samples). For example, when γ = 2, the loss for a sample with a classification probability of 0.9 is reduced to 1 / 100 of its original value, while the loss for a sample with a classification probability of 0.1 is only reduced to 0.81 times its original value.

[0102] During the defect localization phase, the defect type identifier is pattern-matched against a historical defect database. This database contains a large number of known defect samples, along with their feature vectors, type labels, and corresponding spatial locations. A nearest neighbor search algorithm is used to find the K samples in the database that are most similar to the current defect type. To improve search efficiency, a KD tree index is applied to the feature vectors in the database, dividing the high-dimensional space into multiple hyperrectangular regions, each storing a certain number of sample points. For example, for a 128-dimensional feature vector, the KD tree can quickly locate the K sample points closest to the query vector without searching the entire database.

[0103] Based on the K retrieved similar samples, defect location reference parameters are constructed. Statistical parameters of the spatial distribution of these samples, such as mean, variance, and principal direction, are calculated. For hot spot defects, the hot spot center coordinates, area size, and shape characteristics of similar samples are statistically analyzed. For hidden crack defects, the crack starting point, direction, and length distribution are analyzed. These statistical parameters constitute the prior knowledge for defect location and guide the subsequent coordinate mapping process.

[0104] During the spatial coordinate mapping process, the Thin Plate Spline (TPS) interpolation algorithm is used to convert the defect type identification into specific spatial coordinates. TPS interpolation is an elastic transformation method that can construct a smooth deformation field through the correspondence between control point sets. First, a set of control points are defined on the defect image, and the positions of these control points correspond to the reference positions in the historical defect database. The deformation field of the entire image is then calculated using the TPS algorithm, and the abstract defect type identification is mapped to specific image coordinates. For example, when a hot spot defect is identified in a certain area, the TPS algorithm will map the defect type identification of the area to a specific coordinate position in the image based on the distribution pattern of historical hot spot samples.

[0105] To improve positioning accuracy, sub-pixel positioning technology has been introduced. A Gaussian fit is performed on the defect edge area, and the sub-pixel boundary of the defect is determined by calculating the peak position of the Gaussian function. For example, for the edge of a hidden crack defect, a two-dimensional Gaussian function is used to fit the grayscale variation curve. The crack boundary is then determined with sub-pixel accuracy by solving for the Gaussian function's extreme points. This method can improve positioning accuracy from the pixel level (typically tens of microns) to the sub-pixel level (typically several microns), meeting the requirements for high-precision inspection of photovoltaic modules.

[0106] The resulting defect location coordinates with regional markers are represented as polygonal regions, with the vertex coordinates of the polygons accurate to the sub-pixel level. These coordinates contain not only the defect location but also attribute information such as the defect type and severity. For example, for a hot spot defect, the output coordinate information will include parameters such as the hot spot's center position, area size, and temperature distribution gradient, providing detailed data support for subsequent defect repair and quality analysis. The entire defect detection and analysis result generation process achieves a precise conversion from abstract features to specific physical locations through the organic combination of classification prediction and spatial mapping, providing an intuitive and actionable decision-making basis for photovoltaic product quality control.

[0107] Example 5:

[0108] During the photovoltaic product production process, this embodiment's full-process integration achieves deep collaboration between production processes and quality inspection through a data acquisition synchronization mechanism, a feature fusion network architecture, and a defect traceability decision system. The data acquisition phase utilizes a distributed multi-source heterogeneous data acquisition framework, integrating the ERP system with automated inspection equipment via an Industrial Internet of Things gateway. The ERP system automatically collects production process data at preset intervals (e.g., every 15 minutes), including raw material batch numbers, lamination temperature curves, and welding parameter sequences. Simultaneously, a machine vision system captures EL and infrared thermal images at a rate of 10 frames per second, extracting defect feature parameters in real time through edge computing nodes.

[0109] To address clock skew between different data sources, a global clock synchronization protocol was introduced. GPS timing modules were deployed on each data collection terminal to ensure that timestamp errors across all devices were within ±100 microseconds. When quality issues were identified within a batch of products, unified timestamps could be used to accurately correlate production parameters with test results. For example, if EL testing detected a hidden crack in a cell at 10:05:30, the system could quickly locate the lamination temperature records from 10:00:00 to 10:05:00 and analyze whether the temperature fluctuations during that period exceeded the process window.

[0110] A multi-scale spatiotemporal feature fusion network is used for feature extraction and cross-domain correlation. Production process data is segmented using sliding time windows, with the window length dynamically adjusted based on process characteristics: daily windows are used for raw material batch data, minute-level windows for lamination temperature data, and second-level windows for soldering parameters. Data within each window is processed using a gated recurrent unit (GRU) to extract temporal features, which are then weighted and aggregated using an attention mechanism. For example, for the lamination temperature curve, the attention mechanism automatically focuses on time segments with abnormal heating rates, which may be closely related to internal stress defects in the cell.

[0111] Feature extraction from defect detection data utilizes a multimodal fusion strategy. Texture features are extracted from EL images using a residual network, and candidate boxes are generated for hot spot distribution areas using a region proposal network (RPN). Feature alignment is then achieved through a RoIAlign operation. Features from different modalities are combined through a tensor fusion layer to generate a multidimensional defect feature vector. For example, when the EL image shows grayscale anomalies and the infrared thermal image detects localized high temperatures, the tensor fusion layer strengthens the correlation between these two features, highlighting the characteristic pattern of a combined hot spot and microcrack defect.

[0112] The cross-domain correlation process incorporates a causal inference mechanism, exploring potential causal relationships by constructing a production-defect Bayesian network. Network nodes represent production parameters and defect types, while edge weights indicate causal strength. For example, the edge weights between raw material batch parameters and cell conversion efficiency defects are determined through causal effect estimation. This estimation, based on a counterfactual reasoning framework, calculates the causal effect value of "if the raw material batch changes, the defect probability changes." This causal analysis distinguishes between correlation and causality, avoiding misinterpretation of incidental phenomena as causal relationships.

[0113] The defect detection and analysis results generation stage utilizes a cascaded classifier architecture. The first-level classifier uses a lightweight neural network to quickly screen normal samples, concentrating computing resources on suspicious samples. The second-level classifier employs a deep neural network for refined classification, identifying specific defect types. For example, for EL images, the first-level classifier uses MobileNetV3 to quickly determine the presence of abnormal areas. The second-level classifier uses ResNet50 to perform fine-grained classification of suspicious areas, distinguishing between different defect types, such as hidden cracks, fragments, and broken grids.

[0114] In the defect location process, a spatial mapping method based on physical model constraints is proposed. A three-dimensional geometric model of the photovoltaic module is established, and the defect detection results are mapped into the physical coordinate system. For hot spot defects, the heat source location is calculated using the heat conduction equation:

[0115]

[0116] in, represents the heat flux vector, is the thermal conductivity, is the temperature gradient. This formula describes the relationship between the direction of heat flow and the temperature gradient. By measuring the surface temperature distribution and combining it with the material's thermophysical properties, the location of the internal heat source can be inferred. This physical model constraint can improve the accuracy of defect location, especially for internal defects covered by packaging materials.

[0117] The system is deployed using a microservices architecture, encapsulating functions such as data collection, feature extraction, and model inference as independent services. Asynchronous communication between services is achieved through message queues, ensuring high system availability and scalability. For example, as production scales up, the number of feature extraction service instances can be scaled out to meet increased data processing needs. Furthermore, a service circuit breaker mechanism is introduced to automatically switch to a backup service if a service fails, ensuring continuous system operation.

[0118] In actual production environments, the system has established a real-time monitoring and early warning mechanism. By setting production parameter thresholds and defect probability thresholds, early warning of quality issues is achieved. For example, if the lamination temperature deviates from the target value by ±3°C for three consecutive sampling cycles, the system automatically triggers an alert, prompting adjustments to process parameters. If the probability of hidden crack defects in a batch of products exceeds 0.3, the system automatically intercepts that batch of products, preventing them from entering the next process.

[0119] The entire system uses a knowledge graph to achieve structured management of production knowledge. This system converts photovoltaic production process knowledge, equipment maintenance expertise, and defect handling experience into nodes and edges within a graph database. For example, the "Lamination Temperature Excessive" node is connected to the "Bubble Defect" node via the "Cause" edge, and then to the "Adjust Heating Power" node via the "Repair Method" edge. This knowledge graph supports complex reasoning queries. When a specific defect is discovered, relevant production parameter anomalies and treatment measures can be quickly retrieved.

[0120] In terms of data security, the system uses end-to-end encryption to protect sensitive data. Production process data is encrypted at the point of collection, securely transmitted using the TLS 1.3 protocol, and stored using blockchain technology to ensure data immutability. For example, raw material batch information is uniquely identified using a hash algorithm and recorded on the blockchain. Any modification to this information is detected in real time by the system.

[0121] The system also implements a continuous learning mechanism, regularly collecting new production data and defect samples for incremental model training. When model performance degradation is detected, retraining is automatically triggered. For example, 100,000 new production defect samples are collected each quarter. Using federated learning technology, model parameters are updated without compromising enterprise data privacy, ensuring the system maintains its ability to identify new defects.

[0122] This fully integrated approach achieves closed-loop management, from production parameter collection to precise defect location. The system monitors the production process in real time, providing early warning of potential quality issues and pinpointing defects, providing a comprehensive and efficient solution for photovoltaic product quality control.

[0123] It should be noted that, in this document, relational terms such as first and second, etc., are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "comprises," "includes," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that includes a list of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such process, method, article, or apparatus.

[0124] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. ERP data collaborative analysis method for photovoltaic product defect detection, characterized by: include: Obtaining photovoltaic production process data of multiple inspection batches within a predetermined period collected by the ERP system, and photovoltaic module defect detection data of the multiple inspection batches collected by the defect detection device; performing feature extraction and cross-domain correlation on the production process data of the multiple inspection batches and the defect detection data of the multiple inspection batches to obtain a production-defect collaborative feature vector; generating defect detection analysis results based on the production-defect collaborative feature vector; The process of extracting features and performing cross-domain correlation on the production process data of the multiple inspection batches and the defect detection data of the multiple inspection batches to obtain a production-defect collaborative feature vector includes: Arranging the production process data of the multiple inspection batches according to the time dimension and the parameter category dimension into production process structured input data, and arranging the defect detection data of the multiple inspection batches according to the time dimension and the detection index dimension into defect detection structured input data; Performing time coding and group feature extraction on the production process structured input data and the defect detection structured input data respectively to obtain production process multi-level feature data and defect detection multi-level feature data; The multi-level feature data of the production process and the multi-level feature data of the defect detection are fused across data sources to obtain the production-defect collaborative feature vector.

2. The ERP data collaborative analysis method for photovoltaic product defect detection according to claim 1 is characterized in that: The production process data includes raw material batch number, lamination temperature record, welding process parameters and packaging pressure value.

3. The ERP data collaborative analysis method for photovoltaic product defect detection according to claim 2 is characterized in that: The defect detection data includes the grayscale value of the electroluminescent image, the coordinates of the hot spot distribution area and the length parameters of the hidden crack.

4. The ERP data collaborative analysis method for photovoltaic product defect detection according to claim 3 is characterized in that: Performing time coding and group feature extraction on the production process structured input data and the defect detection structured input data respectively to obtain production process multi-level feature data and defect detection multi-level feature data, including: Inputting the production process structured input data into a group feature extraction network based on a sliding time window to obtain multi-level feature data of the production process; The defect detection structured input data is input into a feature extraction network based on temporal association coding to obtain the defect detection multi-level feature data.

5. The ERP data collaborative analysis method for photovoltaic product defect detection according to claim 4 is characterized in that: Cross-data source fusion of the production process multi-level feature data and the defect detection multi-level feature data to obtain the production-defect collaborative feature vector includes: Inputting the production process multi-level feature data and the defect detection multi-level feature data into a collaborative attention fusion network to obtain production-defect association weight data; Inputting the production-defect association weight data into a rule-based constraint correction module to obtain optimized association weight data; Inputting the production process multi-level feature data and the defect detection multi-level feature data into a feature enhancement network to obtain enhanced production process multi-level feature data and enhanced defect detection multi-level feature data respectively; The enhanced production process multi-level feature data and the enhanced defect detection multi-level feature data are jointly encoded to obtain the production-defect collaborative feature vector.

6. The ERP data collaborative analysis method for photovoltaic product defect detection according to claim 5 is characterized in that: Inputting the multi-level feature data of the production process and the multi-level feature data of the defect detection into a collaborative attention fusion network to obtain production-defect association weight data, including: The feature similarity vector of the multi-level feature data of the production process and the feature difference vector of the multi-level feature data of the defect detection are calculated, and the production-defect association weight data are generated through a vector superposition operation.

7. The ERP data collaborative analysis method for photovoltaic product defect detection according to claim 6, characterized in that: Inputting the production process multi-level feature data and the defect detection multi-level feature data into a feature enhancement network to obtain enhanced production process multi-level feature data and enhanced defect detection multi-level feature data, respectively, includes: Performing dimension compression on the multi-level feature data of the production process to obtain a production process feature compression vector; Performing a dimensionality reduction operation after performing feature fusion on the optimized association weight data and the production process feature compression vector to obtain enhanced multi-level feature data of the production process; After feature normalization is performed on the defect detection multi-level feature data, weighted modulation is performed on the data with the optimized associated weight data to obtain enhanced defect detection multi-level feature data.

8. The ERP data collaborative analysis method for photovoltaic product defect detection according to claim 7, characterized in that: Based on the production-defect collaborative feature vector, defect detection analysis results are generated, including: Inputting the production-defect collaborative feature vector into a classification prediction model to obtain a defect type identification; The defect detection analysis result including the defect location coordinates is generated according to the defect type identification.

9. The ERP data collaborative analysis method for photovoltaic product defect detection according to claim 8, characterized in that: Inputting the production-defect collaborative feature vector into a classification prediction model to obtain a defect type identification includes: Performing feature space mapping on the production-defect collaborative feature vector to obtain a high-dimensional feature encoding vector; The high-dimensional feature encoding vector is input into a multi-layer classification network to perform category probability calculation, and the defect type identifier associated with the defect type is output.

10. The ERP data collaborative analysis method for photovoltaic product defect detection according to claim 9, characterized in that: Generating the defect detection analysis result including the defect location coordinates according to the defect type identifier includes: Performing pattern matching on the defect type identifier and the historical defect database to obtain defect location reference parameters; The defect type identifier is spatially mapped based on the defect location reference parameter to generate the defect position coordinates with area markings.

Images