Dynamic data acquisition and intelligent screening analysis method and system based on multi-dimensional feature fusion

Abstract

The invention belongs to the technical field of data analysis, and provides a dynamic data acquisition and intelligent screening analysis method and system based on multi-dimensional feature fusion. The method comprises the steps of multi-dimensional working data collection, abnormal problem judgment, abnormal process matching, potential problem equipment determination, secondary data collection, abnormal multi-dimensional data screening and abnormal data analysis model analysis. According to the invention, the abnormal production identification sensitivity is improved through the abnormal judgment model; through the default frequency and the fault analysis frequency, the data volume during the abnormal traceability period is optimized, and the abnormal data analysis reliability is improved; through abnormal multi-dimensional data screening, the data structure is optimized, the proportion weight of abnormal potential data is improved, and the model calculation amount is reduced; and through the abnormal data analysis model, the early-stage manual labeling process is avoided, and the universality and flexibility of the model are improved.

Description

Technical Field

[0001] This invention relates to the field of data analysis technology, and in particular to a dynamic data acquisition and intelligent filtering analysis method and system based on multi-dimensional feature fusion. Background Technology

[0002] Currently, anomaly analysis in photovoltaic module production typically begins by identifying appearance defects, performance non-compliance with standards, and deviations in process parameters to determine if any abnormalities exist. These are then categorized and traced back to their source. For appearance anomalies, the manufacturing execution system (MES) is used to trace the corresponding processes, such as stringing and lamination, to investigate potential issues with materials like glass or encapsulant, or equipment malfunctions. For performance anomalies, electroluminescence (EL) testing is used to locate internal defects such as cold solder joints or broken grids. Inductively coupled plasma (IVP) testing is employed to measure power degradation, and fishbone diagrams are used to analyze factors such as human operation, cell batches, and soldering temperatures. For process parameter anomalies, equipment operating curves are retrieved, and the 5 Whys analysis is used to trace calibration deviations or sensor malfunctions. Finally, comparative experiments are conducted to verify the effectiveness of corrective measures, and these measures are then formalized into standardized processing procedures.

[0003] However, existing component anomaly analysis methods rely excessively on the subjective judgment and troubleshooting abilities of process engineers, resulting in slow and inefficient detection of abnormal components. Currently, some automated process inspection technologies exist; however, these technologies generally use fixed-frequency data collection, leading to poor accuracy in locating abnormal components. Furthermore, existing automated inspection technologies often rely on threshold judgments for individual devices, resulting in a single data source and insufficient comprehensive anomaly tracing. While some image analysis methods can achieve accurate anomaly matching, the small size of the abnormal area within the component leads to excessive waste of computing power and increased detection time. Moreover, most models used for identification require extensive pre-labeling; furthermore, the stable operating state of equipment is not fixed, causing models trained on fixed pre-labeled data to experience a continuous decline in accuracy as production progresses, with significant accuracy fluctuations. Summary of the Invention

[0004] In order to overcome the shortcomings of the existing technology, the purpose of this invention is to provide a dynamic data acquisition and intelligent screening analysis method and system based on multi-dimensional feature fusion, which solves the problems of poor reliability, low positioning accuracy, waste of computing power and poor model flexibility of the existing methods.

[0005] To achieve the above objectives, the present invention provides the following solution:

[0006] A dynamic data acquisition and intelligent filtering analysis method based on multi-dimensional feature fusion includes:

[0007] Work data is collected from each type of equipment on the production line at a preset default frequency to obtain multi-dimensional work data;

[0008] The collected abnormal component data is analyzed using a pre-built anomaly detection model to obtain the anomaly problem detection results;

[0009] Based on the pre-built problem source index library, the problem process is initially determined according to the abnormal problem determination result, and the abnormal process matching result is obtained;

[0010] The equipment in the area corresponding to the abnormal process matching result is set as potential problem equipment;

[0011] The default frequency of the potentially problematic equipment is switched to the fault analysis frequency, and the working data of the potentially problematic equipment is collected according to the fault analysis frequency to obtain the latest multi-dimensional working data.

[0012] Based on the abnormal process matching results, the multi-dimensional work data is filtered to obtain abnormal multi-dimensional data;

[0013] The abnormal multidimensional data is processed using an anomaly data analysis model to obtain the anomaly data analysis results.

[0014] Preferably, operational data is collected from each type of equipment on the production line according to a preset default frequency to obtain multi-dimensional operational data, including:

[0015] The working data of the first glass loading machine is collected to obtain the first loading machine data; the loading machine data includes: loading time, loading speed, and pressure data of each suction cup of the glass loading machine during loading;

[0016] The working data of the first film-laying machine is collected to obtain the first laying data; the laying data includes: traction force data, traction distance, and laying time;

[0017] The working data of the spot welding machine is collected to obtain spot welding data; the spot welding data includes: spot welding depth, spot welding duration, and spot welding temperature;

[0018] The stringing welding machine is used to collect data, including: upper suction cup pressure data, upper moving speed, tooling data, and lamp box temperature.

[0019] Data is collected from the string-spinning machine to obtain string-spinning data; the string-spinning data includes: extracted suction cup pressure data, robotic arm moving speed, and placement position data;

[0020] Data is collected from the lap welding machine to obtain lap welding data; the lap welding data includes: suction cup position data, descent depth data, and weld strip position data;

[0021] Data is collected from the tape machine to obtain tape addition data; the tape addition data includes: tape position data and tape descent depth data;

[0022] The working data of the second film-laying machine is collected to obtain the second laying data;

[0023] The working data of the second glass loading machine is collected to obtain the data of the second loading machine;

[0024] Data is collected from the EL testing equipment to obtain EL testing data; the EL testing data includes: component internal defect judgment data.

[0025] Preferably, operational data is collected from each type of equipment on the production line according to a preset default frequency to obtain multi-dimensional operational data, including:

[0026] Data is collected from the edge banding machine to obtain edge banding data;

[0027] Data is collected from the laminator to obtain lamination data; the lamination data includes: temperature data, temperature rise data, and vacuum level data;

[0028] Data is collected from the edge trimming machine to obtain edge trimming data;

[0029] Data is collected from the framing machine to obtain framing data;

[0030] Data is acquired using IV testing equipment to obtain IV test data; the IV test data includes: component electrical performance data;

[0031] The multi-dimensional working data is obtained by integrating the first laying data, the spot welding data, the string welding data, the swaying string data, the overlapping welding data, the tape addition data, the second laying data, the second loading machine data, the EL detection data, the edge sealing data, the lamination data, the edge trimming data, the framing data, and the IV test data.

[0032] Preferably, the collected abnormal component data is analyzed using a pre-built anomaly detection model to obtain an anomaly detection result, which includes:

[0033] Abnormal data is extracted from the EL detection data and the IV test data using a preset abnormal data range to obtain the abnormal component data;

[0034] Extract the EL detection image corresponding to the abnormal component data to obtain the abnormal component image;

[0035] The abnormal component image is detected using a pre-trained normal battery cell detection model to obtain normal battery cell data; the expression of the normal battery cell detection model includes:

[0036]

[0037] Among them, D w,max =max(|y act (x)-y std (x)|); S break ={(x,y)|G(x,y)≤μ G -kσ G};

[0038] ΔG adj =μ busbar -μ adjacent ;P normal This represents the overall normality assessment result of the solar cell; D, B, V, H, CD, F, and S are the assessment values ​​for solder strip misalignment, cell breakage, poor solder joint, visible / hidden cell, microcrack, broken grid, and poor busbar welding, respectively; D w,max T represents the maximum offset of the solder strip. d The maximum offset threshold is represented by Area(·), which is the area extraction formula; S break For fragmented areas; T b The fragment area threshold; μ solder T represents the average gray level. g1 T g2 These are the average grayscale threshold and the average gradient threshold, respectively; μ grad Δμ is the average gradient; i is the sub-region index; Δμ i T represents the grayscale difference of the i-th sub-region; h S represents the threshold for the difference in brightness. i S represents the i-th sub-region; area B is the total area of ​​the solar cell; B(x,y) is the microcrack branch region; W crack (x,y) represents the width of the hidden crack; T c The width threshold is defined by Length(·), which is the length extraction function; p and n are the broken grid index and the total number of broken grids, respectively; I(·) is the indicator function; L p T is the length of the p-th broken gate; f1 T f2 These are the thresholds for the length of a single broken gate and the number of broken gates, respectively; μ busbar T represents the average gray level of the busbar region. s The grayscale threshold for the busbar; ΔGadj The difference in grayscale between the busbar and the adjacent area; The average gradient of the confluence region; The average gradient threshold; y act (x), y std (x,y) represents the actual position coordinates and the standard position coordinates of the solder strip, respectively; (x,y) represents the pixel coordinates; G(x,y) represents the grayscale value at position (x,y); μ G The overall grayscale mean; σ G S is the overall grayscale standard deviation; k is the strictness constant; S solder For the welding area; μ G,i μ is the average gray value of the i-th sub-region; G,total This represents the overall grayscale average.

[0039] Preferably, the collected abnormal component data is analyzed using a pre-built anomaly detection model to obtain an anomaly detection result, which includes:

[0040] Remove the battery cell regions in the abnormal component images that correspond to the normal battery cell data to obtain the region of interest image;

[0041] The image of the region of interest and the IV test data are bound together, and the defect judgment data inside the component is used as a label to add labels to the bound data to obtain a training dataset;

[0042] The ResNet network layers are used to extract features from the region of interest images in the training dataset to obtain image features;

[0043] The IV test data in the training dataset is mapped using a fully connected layer to obtain data features;

[0044] The image features and the data features are weighted and fused to obtain fused features;

[0045] The probability distribution of the fused features is calculated using a preset classifier to obtain the network output result;

[0046] Based on the network output and the defect determination data within the component, the ResNet network layer and the preset classifier are iteratively trained using the cross-entropy loss function to obtain the trained anomaly determination model.

[0047] Preferably, the default frequency of the potentially problematic equipment is switched to the fault analysis frequency, and the operating data of the potentially problematic equipment is collected according to the fault analysis frequency to obtain the latest multi-dimensional operating data, including:

[0048] The fault analysis frequency is set to ten times the default frequency;

[0049] When the potentially problematic device receives a fault check instruction, data is collected from the potentially problematic device at the fault analysis frequency to obtain data on the potentially problematic device. Data is also collected from devices other than the potentially problematic device at the default frequency to obtain data on non-abnormal devices.

[0050] By integrating the data of potentially faulty devices and the data of non-abnormal devices, the latest multi-dimensional working data is obtained.

[0051] Preferably, the multi-dimensional work data is filtered based on the abnormal process matching results to obtain abnormal multi-dimensional data, including:

[0052] Data in the multi-dimensional work data that is removed from the list of data where the distance between the equipment process and the abnormal process exceeds a set process threshold is used to obtain preliminary filtered data.

[0053] The abnormal multidimensional data is obtained by removing data collected from devices whose correlation with the potential problem device is less than the impact threshold from the multidimensional working data based on the pre-built device impact association library.

[0054] Preferably, the abnormal multi-dimensional data is processed using an anomaly data analysis model to obtain anomaly data analysis results, including:

[0055] Set the multi-dimensional working data within a preset time period before the abnormal multi-dimensional data as normal working data;

[0056] The normal working data is cleaned, a fill ratio is set according to the default frequency of the device, and the cleaned normal working data is filled and standardized according to the fill ratio.

[0057] The normal working data after standardization is cut into sliding window samples of fixed length, and all the sliding window samples are arranged in order into an input matrix;

[0058] The input matrix is ​​sequentially processed by the encoder and decoder of the LSTM-AE model to extract features and reconstruct data, resulting in a reconstructed image.

[0059] Based on the reconstruction results and the input matrix, the encoder and decoder of the LSTM-AE model are iteratively trained using mean square error to obtain the abnormal data analysis model.

[0060] Preferably, the abnormal multi-dimensional data is processed using an anomaly data analysis model to obtain anomaly data analysis results, including:

[0061] The input matrix is ​​forward-propagated using the anomaly data analysis model to obtain an error list. The distribution of the error list is then calculated to obtain the overall threshold.

[0062] Based on the error list, the errors corresponding to each component within the preset time period are integrated to obtain a feature error matrix. The distribution of each column of the feature error matrix is ​​calculated to obtain the feature-level threshold.

[0063] The abnormal multidimensional data is input into the abnormal data analysis model to obtain the reconstructed output;

[0064] Error calculation is performed on the reconstructed output to obtain the overall reconstruction error. If the overall reconstruction error exceeds the overall threshold, error calculation is performed on each component to obtain the feature reconstruction error. Components whose feature reconstruction errors exceed the feature level threshold are matched to obtain the abnormal data analysis results.

[0065] Preferably, a dynamic data acquisition and intelligent filtering analysis system based on multi-dimensional feature fusion includes:

[0066] The work data acquisition module is used to collect work data from each type of equipment on the production line according to a preset default frequency, and obtain multi-dimensional work data.

[0067] The anomaly detection module is used to analyze the collected anomaly component data using a pre-built anomaly detection model to obtain the anomaly problem detection result;

[0068] The abnormal process matching module is used to perform a preliminary judgment of the abnormal process based on the abnormal problem judgment result according to the pre-built problem source index library, and obtain the abnormal process matching result;

[0069] The potential problem equipment matching module is used to set the equipment in the area corresponding to the abnormal process matching result as potential problem equipment;

[0070] The secondary data acquisition module is used to switch the default frequency of the potentially problematic equipment to the fault analysis frequency, and to collect the working data of the potentially problematic equipment according to the fault analysis frequency to obtain the latest multi-dimensional working data.

[0071] The data filtering module is used to filter the multi-dimensional work data according to the abnormal process matching results to obtain abnormal multi-dimensional data;

[0072] The anomaly analysis module is used to process the abnormal multi-dimensional data using an anomaly data analysis model to obtain anomaly data analysis results.

[0073] The present invention discloses the following technical effects:

[0074] This invention provides a dynamic data acquisition and intelligent filtering analysis method and system based on multi-dimensional feature fusion. Through an anomaly judgment model, it solves the problem of incomplete anomaly identification caused by conventional methods using single data points for device anomaly judgment, achieving accurate anomaly judgment using multi-dimensional data. By using default frequency and fault analysis frequency, it solves the problem of poor positioning accuracy caused by using a single frequency, optimizing the proportion of fault data. Through multi-dimensional anomaly data filtering, it solves the problem of wasted computing power due to data redundancy, increasing the weight of potential anomaly data. Through an anomaly data analysis model, it solves the problems of poor flexibility and versatility of conventional pre-trained models, achieving anomaly source tracing and identification without manual annotation. Attached Figure Description

[0075] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0076] Figure 1 A schematic diagram of a dynamic data acquisition and intelligent filtering analysis process based on multi-dimensional feature fusion provided in an embodiment of the present invention;

[0077] Figure 2 This is a flowchart of the normal battery cell data determination process provided in an embodiment of the present invention;

[0078] Figure 3 A flowchart illustrating the construction process of the anomaly detection model provided in this embodiment of the invention;

[0079] Figure 4 A flowchart illustrating the construction process of the abnormal data analysis model provided in this embodiment of the invention;

[0080] Figure 5 A flowchart for abnormal data analysis provided in an embodiment of the present invention. Detailed Implementation

[0081] 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.

[0082] The purpose of this invention is to provide a dynamic data acquisition and intelligent screening analysis method and system based on multi-dimensional feature fusion, which solves the problems of poor reliability, low positioning accuracy, waste of computing power, and poor model flexibility of existing methods.

[0083] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0084] Figure 1 This is a schematic diagram of a dynamic data acquisition and intelligent filtering analysis process based on multi-dimensional feature fusion provided in an embodiment of the present invention, such as... Figure 1 As shown, this invention provides a dynamic data acquisition and intelligent filtering analysis method based on multi-dimensional feature fusion, including:

[0085] Step 100: Collect working data for each type of equipment on the production line according to the preset default frequency to obtain multi-dimensional working data;

[0086] Step 200: Analyze the collected abnormal component data using a pre-built anomaly detection model to obtain the anomaly problem detection result;

[0087] Step 300: Based on the pre-built problem source index library, perform a preliminary determination of the problem process based on the abnormal problem determination results to obtain the abnormal process matching results;

[0088] Step 400: Set the equipment in the area corresponding to the abnormal process matching result as potential problem equipment;

[0089] Step 500: Switch the default frequency of the potentially problematic device to the fault analysis frequency, and collect the working data of the potentially problematic device according to the fault analysis frequency to obtain the latest multi-dimensional working data;

[0090] Step 600: Filter the multi-dimensional work data according to the abnormal process matching results to obtain abnormal multi-dimensional data;

[0091] Step 700: Process the abnormal multidimensional data using an anomaly data analysis model to obtain the anomaly data analysis results.

[0092] Specifically, operational data is collected from each type of equipment on the production line at a preset default frequency to obtain multi-dimensional operational data, including:

[0093] The working data of the first glass loading machine is collected to obtain the first loading machine data; the loading machine data includes: loading time, loading speed, and pressure data of each suction cup of the glass loading machine during loading;

[0094] The working data of the first film-laying machine is collected to obtain the first laying data; the laying data includes: traction force data, traction distance, and laying time;

[0095] The working data of the spot welding machine is collected to obtain spot welding data; the spot welding data includes: spot welding depth, spot welding duration, and spot welding temperature;

[0096] The stringing welding machine is used to collect data, including: upper suction cup pressure data, upper moving speed, tooling data, and lamp box temperature.

[0097] Data is collected from the string-spinning machine to obtain string-spinning data; the string-spinning data includes: extracted suction cup pressure data, robotic arm moving speed, and placement position data;

[0098] Data is collected from the lap welding machine to obtain lap welding data; the lap welding data includes: suction cup position data, descent depth data, and weld strip position data;

[0099] Data is collected from the tape machine to obtain tape addition data; the tape addition data includes: tape position data and tape descent depth data;

[0100] The working data of the second film-laying machine is collected to obtain the second laying data;

[0101] The working data of the second glass loading machine is collected to obtain the data of the second loading machine;

[0102] Data is collected from the EL testing equipment to obtain EL testing data; the EL testing data includes: component internal defect judgment data.

[0103] Furthermore, operational data is collected from each type of equipment on the production line at a preset default frequency to obtain multi-dimensional operational data, including:

[0104] Data is collected from the edge banding machine to obtain edge banding data;

[0105] Data is collected from the laminator to obtain lamination data; the lamination data includes: temperature data, temperature rise data, and vacuum level data;

[0106] Data is collected from the edge trimming machine to obtain edge trimming data;

[0107] Data is collected from the framing machine to obtain framing data;

[0108] Data is acquired using IV testing equipment to obtain IV test data; the IV test data includes: component electrical performance data;

[0109] The multi-dimensional working data is obtained by integrating the first laying data, the spot welding data, the string welding data, the swaying string data, the overlapping welding data, the tape addition data, the second laying data, the second loading machine data, the EL detection data, the edge sealing data, the lamination data, the edge trimming data, the framing data, and the IV test data.

[0110] refer to Figure 2 The collected abnormal component data is analyzed using a pre-built anomaly detection model to obtain the anomaly problem detection result, which includes:

[0111] Step 201: Extract abnormal data from the EL detection data and the IV test data using a preset abnormal data range to obtain the abnormal component data;

[0112] Step 202: Extract the EL detection image corresponding to the abnormal component data to obtain the abnormal component image;

[0113] Step 203: Detect the abnormal component image using a pre-trained normal battery cell detection model to obtain normal battery cell data; the expression of the normal battery cell detection model includes:

[0114]

[0115] Among them, D w,max =max(|y act (x)-y std (x)|); S break ={(x,y)|G(x,y)≤μ G -kσ G};

[0116] ΔG adj =μ busbar -μ adjacent ;P normal This represents the overall normality assessment result of the solar cell; D, B, V, H, CD, F, and S are the assessment values ​​for solder strip misalignment, cell breakage, poor solder joint, bright / dark area, microcrack, broken grid, and poor busbar welding, respectively; D w,max T represents the maximum offset of the solder strip. d The maximum offset threshold is represented by Area(·), which is the area extraction formula; S break For fragmented areas; T b The fragment area threshold; μ solder T represents the average gray level. g1 T g2These are the average grayscale threshold and the average gradient threshold, respectively; μ grad Δμ is the average gradient; i is the sub-region index; Δμ i T represents the grayscale difference of the i-th sub-region; h S represents the threshold for the difference in brightness. i S represents the i-th sub-region; area B is the total area of ​​the solar cell; B(x,y) is the microcrack branch region; W crack (x,y) represents the width of the hidden crack; T c The width threshold is defined by Length(·), which is the length extraction function; p and n are the broken grid index and the total number of broken grids, respectively; I(·) is the indicator function; L p T is the length of the p-th broken gate; f1 T f2 These are the thresholds for the length of a single broken gate and the number of broken gates, respectively; μ busbar T represents the average gray level of the busbar region. s The grayscale threshold for the busbar; ΔG adj The difference in grayscale between the busbar and the adjacent area; The average gradient of the confluence region; The average gradient threshold; y act (x), y std (x,y) represents the actual position coordinates and the standard position coordinates of the solder strip, respectively; (x,y) represents the pixel coordinates; G(x,y) represents the grayscale value at position (x,y); μ G The overall grayscale mean; σ G S is the overall grayscale standard deviation; k is the strictness constant; S solder For the welding area; μ G,i μ is the average gray value of the i-th sub-region; G,total This represents the overall grayscale average.

[0117] refer to Figure 3 The collected abnormal component data is analyzed using a pre-built anomaly detection model to obtain the anomaly problem detection result, which includes:

[0118] Step 204: Remove the battery cell regions in the abnormal component image that correspond to the normal battery cell data to obtain the region of interest image;

[0119] Step 205: Bind the image of the region of interest and the IV test data, and use the defect judgment data inside the component as labels to add labels to the bound data to obtain a training dataset;

[0120] Step 206: Use ResNet network layers to extract features from the region of interest images in the training dataset to obtain image features;

[0121] Step 207: Use a fully connected layer to map the IV test data in the training dataset to obtain data features;

[0122] Step 208: Perform weighted fusion processing on the image features and the data features to obtain fused features;

[0123] Step 209: Calculate the probability distribution of the fused features using a preset classifier to obtain the network output result;

[0124] Step 210: Based on the network output results and the component internal defect determination data, the ResNet network layer and the preset classifier are iteratively trained using the cross-entropy loss function to obtain the trained anomaly determination model.

[0125] Specifically, the default frequency of the potentially problematic equipment is switched to the fault analysis frequency, and the operating data of the potentially problematic equipment is collected according to the fault analysis frequency to obtain the latest multi-dimensional operating data, including:

[0126] The fault analysis frequency is set to ten times the default frequency;

[0127] When the potentially problematic device receives a fault check instruction, data is collected from the potentially problematic device at the fault analysis frequency to obtain data on the potentially problematic device. Data is also collected from devices other than the potentially problematic device at the default frequency to obtain data on non-abnormal devices.

[0128] By integrating the data of potentially faulty devices and the data of non-abnormal devices, the latest multi-dimensional working data is obtained.

[0129] Furthermore, based on the abnormal process matching results, the multi-dimensional work data is filtered to obtain abnormal multi-dimensional data, including:

[0130] Data in the multi-dimensional work data that is removed from the list of data where the distance between the equipment process and the abnormal process exceeds a set process threshold is used to obtain preliminary filtered data.

[0131] The abnormal multidimensional data is obtained by removing data collected from devices whose correlation with the potential problem device is less than the impact threshold from the multidimensional working data based on the pre-built device impact association library.

[0132] refer to Figure 4 The abnormal multi-dimensional data is processed using an anomaly data analysis model to obtain anomaly data analysis results, including:

[0133] Step 701: Set the multi-dimensional working data within a preset time period before the abnormal multi-dimensional data as normal working data;

[0134] Step 702: Clean the normal working data, set the fill ratio according to the default frequency of the device, and fill and standardize the cleaned normal working data according to the fill ratio;

[0135] Step 703: Cut the normal working data after standardization into sliding window samples of fixed length, and arrange all the sliding window samples in order into an input matrix;

[0136] Step 704: Use the encoder and decoder of the LSTM-AE model to extract features and reconstruct data from the input matrix in sequence to obtain the reconstruction result;

[0137] Step 705: Based on the reconstruction results and the input matrix, the encoder and decoder of the LSTM-AE model are iteratively trained using mean square error to obtain the abnormal data analysis model.

[0138] refer to Figure 5 The abnormal multi-dimensional data is processed using an anomaly data analysis model to obtain anomaly data analysis results, including:

[0139] Step 706: Use the anomaly data analysis model to perform forward propagation on the input matrix to obtain an error list, and perform distribution calculation on the error list to obtain the overall threshold;

[0140] Step 707: Integrate the errors corresponding to each component within the preset time period according to the error list to obtain a feature error matrix. Perform distribution calculation on each column of the feature error matrix to obtain the feature-level threshold.

[0141] Step 708: Input the abnormal multi-dimensional data into the abnormal data analysis model to obtain the reconstruction output;

[0142] Step 709: Perform error calculation on the reconstructed output to obtain the overall reconstruction error. If the overall reconstruction error exceeds the overall threshold, perform error calculation on each component to obtain the feature reconstruction error, and match the components whose feature reconstruction errors exceed the feature level threshold to obtain the abnormal data analysis results.

[0143] Specifically, the main equipment in a photovoltaic module production line includes: a first glass loading machine, a first film laying machine, a spot welding machine, a string welding machine, a string swivel machine, a stacking welding machine, a tape machine, a second film laying machine, a second glass loading machine, EL testing equipment, an edge sealing machine, a laminator, an edge trimming machine, a framing machine, IV testing equipment, and other equipment. The glass loading machine is mainly used to place the upper and lower layers of glass. Its key detection data is concentrated on the suction cups for transporting the glass and related data of the robotic arm. The film laying machine is used to lay the film between the two sides of the glass. The film laying has a significant impact on the overall module and needs to be ensured to be laid flat and in place. Therefore, it is necessary to collect data such as traction force, traction distance, and laying time. The stringing machine is used to assemble individual cells into a battery string. This equipment is the most critical equipment in module production. It is roughly divided into a feeding area and a stringing area. The feeding area needs to focus on the feeding suction cups and transmission equipment, while the stringing area mainly focuses on related equipment such as welding strips, tooling, and light boxes. The stringing machine is similar to the glass loading machine, but it also needs to detect the stringing position to avoid problems such as string spacing. In addition to the washing of the trays, the stacking machine also needs to focus on the busbar. The laminator mainly focuses on data such as temperature and vacuum during operation. The tape machine, edge sealing machine, edge trimming machine, and framing machine collect corresponding data according to their control parameters. The EL detection equipment and IV testing equipment are mainly used for the judgment of routine anomalies and to assist the work of the anomaly judgment model.

[0144] Furthermore, to address the issue of having only EL detection results but lacking anomaly region annotations, or to improve the efficiency of anomaly region annotation in EL images, this embodiment employs a reverse method. It indirectly extracts anomaly regions by reversely eliminating high-requirement normal battery cell regions, avoiding the increased cost and time associated with constructing an additional anomaly region detection model. The normal battery cell detection model constructed in this embodiment is as follows:

[0145]

[0146] D w,max =max(|y act (x)-y std (x)|)

[0147]

[0148] S break ={(x,y)|G(x,y)≤μ G -kσ G}

[0149]

[0150]

[0151] Δμ i =|μG,i -μ G,total |

[0152]

[0153] ΔG adj =μ busbar -μ adjacent

[0154]

[0155] This embodiment only identifies a few abnormalities, such as solder strip misalignment, broken pieces, poor solder joints, visible / hidden pieces, microcracks, broken grids, and poor busbar welding. D, B, V, H, CD, F, and S correspond to the judgment values ​​for these abnormalities, and the judgment value contains only one of 0 and 1. Finally, all judgment values ​​are negated and multiplied to obtain the final judgment result. Only when the test results do not contain any of the above abnormalities will the corresponding battery cell be judged to be normal.

[0156] Furthermore, regions corresponding to normal cell data are removed from abnormal component images to obtain regions of interest (ROI) images. To eliminate image noise and enhance the characteristics of abnormal regions, ROI images are preprocessed, including conversion to grayscale, Gaussian filtering for noise reduction, and adaptive contrast enhancement. IV test data is preprocessed, including outlier removal and Z-score normalization to ensure the data distribution conforms to the model's requirements. The preprocessed ROI images are then temporally and spatially correlated with the IV test data. Then, using the component internal defect judgment data as a standard, detailed labels are added to the correlated data to construct the training dataset. ResNet50 network layers are used to extract features from the ROI images in the training dataset. Specifically, basic features such as edges and textures are extracted through convolutional layers, and higher-level semantic features are extracted through subsequent convolutional layers and global average pooling layers to obtain image feature vectors. Fully connected layers are used to map the IV test data in the training dataset to obtain data feature vectors. The image feature vectors and data feature vectors are then weighted and fused, where the weight parameters are learnable and dynamically improved through backpropagation. The fused features are input into a pre-defined classifier to calculate the probability distribution and output the probability values ​​of various anomalies, thus obtaining the network's output. The ResNet network layers, fully connected layers, and the pre-defined classifier are then jointly and iteratively trained to finally obtain a converged anomaly detection model.

[0157] Preferably, this embodiment designs a problem source index library, some of the contents of which are shown in Table 1.

[0158] Table 1

[0159]

[0160] By performing a preliminary judgment on the problematic process based on the pre-built problem source index library, the abnormal process matching result can be obtained, and the equipment in the area corresponding to the abnormal process matching result can be set as potential problematic equipment.

[0161] Furthermore, the fault analysis frequency is set to ten times the default frequency. When a suspected faulty device receives a fault inspection instruction, data for that device is collected at the fault analysis frequency to obtain suspected faulty device data; simultaneously, data from other devices is collected at the default frequency to obtain normal device data. The suspected faulty device data and normal device data are integrated to form the latest multi-dimensional work data. Preliminary data screening: Data in the multi-dimensional work data where the mismatch between the device's process and the abnormal process is too large (exceeding a set threshold) is deleted. Abnormal multi-dimensional data: Based on a pre-established equipment impact correlation library, data collected from devices in the multi-dimensional work data that are not closely related to the potential faulty device (correlation degree below the impact threshold) is deleted.

[0162] Specifically, multi-dimensional working data from a period preceding the abnormal multi-dimensional data is designated as normal working data. This normal working data is cleaned, and a fill ratio is set according to the device's default frequency. Based on this, the cleaned normal working data is filled and standardized. The standardized normal working data is then segmented into fixed-length sliding window samples and arranged sequentially into an input matrix. The encoder and decoder of the LSTM-AE model are used to sequentially extract features and reconstruct the data from the input matrix, obtaining the reconstruction result. Based on the reconstruction result and the input matrix, the encoder and decoder of the LSTM-AE model are iteratively trained using mean squared error to obtain the abnormal data analysis model.

[0163] Furthermore, an anomaly analysis model is used to process the input matrix forward, obtaining an error list and calculating its distribution to determine the overall threshold. Errors from each component within a preset time period are integrated to form a feature error matrix, and the distribution of each column is calculated to obtain the feature-level threshold. Anomaly multi-dimensional data is input into the anomaly analysis model to obtain the reconstruction output. The reconstruction error is calculated to obtain the overall reconstruction error. If the overall threshold is exceeded, the feature reconstruction error of each component is calculated, identifying components that exceed the feature-level threshold, thus obtaining the final anomaly analysis result.

[0164] As an optional implementation, this embodiment also provides a dynamic data acquisition and intelligent filtering analysis system based on multi-dimensional feature fusion, including:

[0165] The work data acquisition module is used to collect work data from each type of equipment on the production line according to a preset default frequency, and obtain multi-dimensional work data.

[0166] The anomaly detection module is used to analyze the collected anomaly component data using a pre-built anomaly detection model to obtain the anomaly problem detection result;

[0167] The abnormal process matching module is used to perform a preliminary judgment of the abnormal process based on the abnormal problem judgment result according to the pre-built problem source index library, and obtain the abnormal process matching result;

[0168] The potential problem equipment matching module is used to set the equipment in the area corresponding to the abnormal process matching result as potential problem equipment;

[0169] The secondary data acquisition module is used to switch the default frequency of the potentially problematic equipment to the fault analysis frequency, and to collect the working data of the potentially problematic equipment according to the fault analysis frequency to obtain the latest multi-dimensional working data.

[0170] The data filtering module is used to filter the multi-dimensional work data according to the abnormal process matching results to obtain abnormal multi-dimensional data;

[0171] The anomaly analysis module is used to process the abnormal multi-dimensional data using an anomaly data analysis model to obtain anomaly data analysis results.

[0172] The beneficial effects of this invention are as follows:

[0173] This invention improves the sensitivity of abnormal production identification through an anomaly determination model; optimizes the amount of data during anomaly tracing by default frequency and fault analysis frequency, thereby improving the reliability of anomaly data analysis; optimizes the data structure by filtering anomaly data from multiple dimensions, increasing the proportion and weight of potential anomaly data and reducing the computational load of the model; and avoids the manual annotation process in the early stage through an anomaly data analysis model, thereby improving the versatility and flexibility of the model.

[0174] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0175] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A dynamic data acquisition and intelligent screening analysis method based on multi-dimensional feature fusion, characterized in that, The method comprises the following steps: Collecting work data of each device on the production line according to a preset default frequency to obtain multi-dimensional work data; Analyzing the collected abnormal component data by using a pre-constructed abnormality judgment model to obtain an abnormal problem judgment result; Performing problem process preliminary judgment on the abnormal problem judgment result according to a pre-constructed problem source index library to obtain an abnormal process matching result; Setting the device in the region corresponding to the abnormal process matching result as a potential problem device; Switching the default frequency of the potential problem device to a fault analysis frequency, and collecting work data of the potential problem device according to the fault analysis frequency to obtain the latest multi-dimensional work data; Filtering the multi-dimensional work data according to the abnormal process matching result to obtain abnormal multi-dimensional data; Processing the abnormal multi-dimensional data by using an abnormal data analysis model to obtain an abnormal data analysis result. 2.The dynamic data acquisition and intelligent screening analysis method based on multi-dimensional feature fusion according to claim 1, characterized in that, Collecting work data of each device on the production line according to a preset default frequency to obtain multi-dimensional work data, comprising: Collecting work data of a first glass sheeting machine to obtain first sheeting machine data; the sheeting machine data comprises sheeting time, sheeting speed and pressure data of each suction cup of the glass sheeting machine during sheeting; Collecting work data of a first film laying machine to obtain first laying data; the laying data comprises traction force data, traction distance and laying time; Collecting work data of a spot welding machine to obtain spot welding data; the spot welding data comprises spot welding depth, spot welding time length and spot welding temperature; Collecting work data of a string welding integrated machine to obtain string welding data; the string welding data comprises sheeting suction cup pressure data, sheeting movement speed, tooling data and light box temperature; Collecting work data of a string placing machine to obtain string placing data; the string placing data comprises extraction suction cup pressure data, mechanical arm movement speed and placing position data; Collecting work data of a lap welding machine to obtain lap welding data; the lap welding data comprises suction cup position data, lowering depth data and welding strip position data; Collecting work data of a tape machine to obtain tape adding data; the tape adding data comprises tape position data and tape lowering depth data; Collecting work data of a second film laying machine to obtain second laying data; Collecting work data of a second glass sheeting machine to obtain second sheeting machine data; Collecting data of an EL detection device to obtain EL detection data; the EL detection data comprises internal defect judgment data of a component. 3.The dynamic data acquisition and intelligent screening analysis method based on multi-dimensional feature fusion according to claim 2, characterized in that, Collecting work data of each device on the production line according to a preset default frequency to obtain multi-dimensional work data, comprising: Collecting data of an edge sealing machine to obtain edge sealing data; Collecting data of a laminating machine to obtain laminating data; the laminating data comprises temperature data, temperature rising data and vacuum degree data; Collecting data of an edge trimming machine to obtain edge trimming data; Collecting data of a frame loading machine to obtain frame loading data; Collecting data of an IV test device to obtain IV test data; the IV test data comprises electrical performance data of a component; Integrate the first laying data, the spot welding data, the string welding data, the string swinging data, the overlay welding data, the tape adding data, the second laying data, the second sheet feeding machine data, the EL detection data, the edge sealing data, the laminating data, the edge cutting data, the frame assembling data, the IV test data, to obtain the multi-dimensional work data.

4. The dynamic data acquisition and intelligent screening analysis method based on multi-dimensional feature fusion according to claim 3, characterized in that, Analyze the collected abnormal component data by using a pre-constructed abnormality judgment model to obtain an abnormal problem judgment result, which includes: Extract abnormal data from the EL detection data and the IV test data by using a pre-set abnormal data range to obtain the abnormal component data; Extract an EL detection image corresponding to the abnormal component data to obtain an abnormal component image; The pre-trained normal battery piece detection model is used for detecting the abnormal component image, and normal battery piece data is obtained; an expression of the normal battery piece detection model comprises: where D w,max = max(|y act (x) - y std (x)|); S break = {(x, y) | G(x, y) < μ G - kσ G}; Δμ i = |μ G,i - μ G,total |; ΔG = μ adj - μ busbar ; P adjacent = D + B + V + H + CD + F + S; D normal is the solder strip offset determination value; B w,max is the maximum solder strip offset; T d is the maximum offset threshold; Area(·) is the area extraction formula; S break is the broken piece area; T b is the broken piece area threshold; μ solder is the average gray level; T g1 , T g2 are the average gray level threshold and the average gradient threshold, respectively; μ grad is the average gradient; i is the sub-region index; Δμ i is the region gray level difference of the i-th sub-region; T h is the light-dark difference threshold; S i is the i-th sub-region; S area is the total area of the battery piece; B(x, y) is the hidden crack branch area; W crack (x, y) is the hidden crack width; T c is the width threshold; Length(·) is the length extraction function; p, n are the broken grid index and the total number of broken grids, respectively; I(·) is the indicator function; L p is the length of the p-th broken grid; T f1 , T f2 are the single broken grid length threshold and the broken grid number threshold, respectively; μ busbar is the bus bar region average gray level; T s is the bus bar gray level threshold; ΔG adj is the bus bar and adjacent region gray level difference; is the bus bar region average gradient; is the average gradient threshold; y act (x), y std (x) are the solder strip actual position coordinates and the solder strip standard position coordinates, respectively; (x, y) is the pixel coordinate; G(x, y) is the gray level value of the (x, y) position; μ G is the overall gray level mean; σ G is the overall gray level standard deviation; k is the strictness constant; S solder is the welding area; μ G,i is the gray level mean of the i-th sub-region; μ G,total The overall gray mean value is taken as the whole.

5. The dynamic data acquisition and intelligent screening analysis method based on multi-dimensional feature fusion according to claim 4, characterized in that, Analyze the collected abnormal component data by using a pre-constructed abnormality judgment model to obtain an abnormal problem judgment result, which includes: Remove a battery piece region corresponding to the normal battery piece data in the abnormal component image to obtain a region of interest image; Bind the region of interest image and the IV test data, and add a label to the bound data by using the component internal defect judgment data as the label to obtain a training data set; Extract features of the region of interest image in the training data set by using a ResNet network layer to obtain image features; Map the IV test data in the training data set by using a fully connected layer to obtain data features; Perform weighted fusion processing on the image features and the data features to obtain fused features; Calculate a probability distribution of the fused features by using a pre-set classifier to obtain a network output result; Iteratively train the ResNet network layer and the pre-set classifier by using a cross-entropy loss function according to the network output result and the component internal defect judgment data to obtain the trained abnormality judgment model. 6.The dynamic data acquisition and intelligent screening analysis method based on multi-dimensional feature fusion according to claim 5, characterized in that, Switch a default frequency of the potential problem equipment to a fault analysis frequency, and collect work data of the potential problem equipment according to the fault analysis frequency to obtain the latest multi-dimensional work data, which includes: Determine the fault analysis frequency as ten times of the default frequency; After the potential problem equipment receives a fault check instruction, collect data of the potential problem equipment at the fault analysis frequency to obtain potential fault equipment data, and collect data of equipment other than the potential problem equipment at the default frequency to obtain non-abnormal equipment data; Integrate the potential fault equipment data and the non-abnormal equipment data to obtain the latest multi-dimensional work data.

7. The dynamic data acquisition and intelligent screening analysis method based on multi-dimensional feature fusion according to claim 6, characterized in that, Screen the multi-dimensional work data according to the abnormal process matching result to obtain abnormal multi-dimensional data, which includes: Remove data in the multi-dimensional work data whose distance from the abnormal process matching result exceeds a set process threshold to obtain preliminary screening data; Remove data collected by equipment with a correlation degree less than an influence threshold to the potential problem equipment from the multi-dimensional work data according to a pre-constructed equipment influence correlation library to obtain the abnormal multi-dimensional data. 8.The dynamic data acquisition and intelligent screening analysis method based on multi-dimensional feature fusion according to claim 7, characterized in that, The abnormal multi-dimensional data is processed by using an abnormal data analysis model to obtain an abnormal data analysis result, including: The multi-dimensional working data of the abnormal multi-dimensional data in a preset time period is set as normal working data; The normal working data is cleaned, a filling ratio is set according to the default frequency of the equipment, and the cleaned normal working data is filled and standardized according to the filling ratio; The normal working data after the standardization processing is cut into a fixed-length sliding window sample, and all the sliding window samples are arranged in order into an input matrix; The input matrix is sequentially subjected to feature extraction and data reconstruction by using an encoder and a decoder of an LSTM-AE model to obtain a reconstruction result; The encoder and the decoder of the LSTM-AE model are iteratively trained by using a mean square error according to the reconstruction result and the input matrix to obtain the abnormal data analysis model. 9.The dynamic data acquisition and intelligent screening analysis method based on multi-dimensional feature fusion of claim 8, characterized in that, The abnormal multi-dimensional data is processed by using an abnormal data analysis model to obtain an abnormal data analysis result, including: The input matrix is subjected to forward propagation by using the abnormal data analysis model to obtain an error list, and the overall threshold is obtained by distributing the error list; The error corresponding to each component in the preset time period is integrated according to the error list to obtain a feature error matrix, and the feature level threshold is obtained by distributing each column of the feature error matrix; The abnormal multi-dimensional data is input into the abnormal data analysis model to obtain a reconstruction output; The overall reconstruction error is obtained by error calculation on the reconstruction output, if the overall reconstruction error exceeds the overall threshold, the feature reconstruction error of each component is obtained by error calculation, and the components whose feature reconstruction error exceeds the feature level threshold are matched to obtain the abnormal data analysis result.

10. A dynamic data acquisition and intelligent screening analysis system based on multi-dimensional feature fusion, characterized in that, including: a working data acquisition module, configured to acquire working data of each device on a production line according to a preset default frequency to obtain multi-dimensional working data; an abnormality determination module, configured to analyze collected abnormal component data by using a pre-constructed abnormality determination model to obtain an abnormal problem determination result; an abnormal process matching module, configured to preliminarily determine a problem process of the abnormal problem determination result according to a pre-constructed problem source index library to obtain an abnormal process matching result; a potential problem equipment matching module, configured to set the equipment in a region corresponding to the abnormal process matching result as potential problem equipment; a secondary data acquisition module, configured to switch a default frequency of the potential problem equipment to a fault analysis frequency, and acquire working data of the potential problem equipment according to the fault analysis frequency to obtain the latest multi-dimensional working data; a data screening module, configured to screen the multi-dimensional working data according to the abnormal process matching result to obtain abnormal multi-dimensional data; an abnormality analysis module, configured to process the abnormal multi-dimensional data by using an abnormal data analysis model to obtain an abnormal data analysis result.

Images