A metal plate production quality detection system and method based on the Internet of Things

Abstract

The application discloses a metal plate production quality detection system and method based on the Internet of Things, and relates to the technical field of the Internet of Things.The application collects infrared thermal imaging and visible light image data of the metal plate by setting a processing detection cycle, realizes time alignment based on a time stamp, corrects coordinate system deviation by using inherent geometric characteristics, forms a multi-modal image data set consistent in time and space, constructs an initial noise shielding dictionary library, identifies abnormal areas by combining feature matching and residual analysis of current data, dynamically updates the real-time noise shielding dictionary library based on access frequency trend prediction, improves the accuracy and timeliness of detection, establishes a cross-modal feature mapping relationship, respectively adopts corresponding repair strategies for infrared single mode, visible light single mode and double mode abnormalities, ensures the pertinence of abnormal area repair, provides high-quality image data without noise shielding for a metal plate production quality detection and analysis module, and comprehensively improves the accuracy and efficiency of quality detection.

Description

Technical Field

[0001] This invention relates to the field of Internet of Things (IoT) technology, specifically to an IoT-based system and method for quality inspection in the production of metal sheets. Background Technology

[0002] Quality inspection in metal sheet production is a core link in ensuring the safety and efficiency of the industrial chain. It directly determines whether key indicators such as the mechanical properties and dimensional accuracy of the sheet meet the stringent standards of downstream industries. Furthermore, it serves as a crucial barrier to prevent substandard products from entering the market and to avoid potential engineering safety hazards. Simultaneously, accurate quality inspection can provide real-time feedback on production process defects, offering data support for optimizing rolling parameters and reducing raw material losses. This plays an irreplaceable role in improving production efficiency, reducing manufacturing costs, and enhancing product market competitiveness, making it a vital guarantee for the high-quality development of the metal sheet industry.

[0003] In the hot rolling stage of sheet metal production, due to the special environment of high temperature, high humidity, and the splashing of iron oxide scale and the pervasive moisture during rolling, infrared thermal imaging and visible light imaging equipment are often used to acquire images to analyze the surface defects, temperature distribution, and morphological characteristics of the sheet metal. However, this process is easily affected by factors such as thermal noise generated by the sensor, local occlusion caused by the adhesion of iron oxide scale, and image blurring caused by moisture evaporation. These factors generate a lot of noise and various obstructions, interfering with the identification of the true quality characteristics of the sheet metal, and thus affecting the accuracy and reliability of the final sheet metal production quality analysis. Therefore, there is an urgent need for an Internet of Things-based sheet metal production quality inspection system and method. Summary of the Invention

[0004] The purpose of this invention is to provide a metal sheet production quality inspection system and method based on the Internet of Things to solve the problems raised in the prior art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a method for quality inspection of metal sheet production based on the Internet of Things, the method comprising the following steps:

[0006] Step S1: Set the metal sheet processing and inspection cycle, collect infrared thermal image data and visible light image data of the metal sheet to be processed according to the metal sheet processing and inspection cycle, and perform time alignment of the collected infrared thermal image data and visible light image data based on the timestamp data to obtain a multimodal image data set of the metal sheet to be processed.

[0007] Step S1-1: Obtain infrared radiation image data of the metal sheet being processed by installing an infrared thermal imaging sensor; obtain visible light reflection image data of the metal sheet being processed by installing a visible light image sensor; adjust the viewing angles of the infrared thermal imaging sensor and the visible light image sensor to be consistent.

[0008] Steps S1-2: During the acquisition process, the infrared thermal imaging sensor and the visible light image sensor synchronously record the timestamp data corresponding to the acquisition time for each frame of image data generated, and the storage format of the timestamp data is the same.

[0009] Steps S1-3: Extract timestamp data from infrared radiation image data to form an infrared time series; extract timestamp data from visible light reflectance image data to form a visible light time series; align the timestamps in the infrared time series and the visible light time series, and pair infrared radiation image data and visible light reflectance image data at the same time.

[0010] Step S1-4: For the paired image data, select the inherent geometric features of the processed metal sheet to correct the coordinate system deviation between the infrared radiation image data and the visible light reflection image data, forming a multimodal image data set. The inherent geometric features of the processed metal sheet are represented as the contour features of the processed metal sheet. The specific correction process is to adjust the pixel coordinates of the infrared radiation image data and the visible light reflection image data so that the pixel coordinates of the inherent geometric features are at the same coordinate position.

[0011] By setting processing and inspection cycles to collect infrared thermal imaging and visible light image data, recording and aligning timestamps synchronously, and combining the contour features of metal sheets to correct coordinate system deviations, a multimodal image data set is formed. This achieves spatiotemporal consistency integration of multi-source image data, providing a unified benchmark image data foundation for subsequent quality inspection.

[0012] Step S2: Analyze the multimodal image data sets from multiple historical metal sheet processing and inspection cycles to obtain an initial noise occlusion dictionary; and analyze the multimodal image data sets from the current metal sheet processing and inspection cycle, using feature matching and multimodal residual analysis to obtain abnormal image regions in the multimodal image data sets.

[0013] Step S2-1: Retrieve multimodal image data sets from multiple historical metal sheet processing and inspection cycles, and extract noise and occlusion features from infrared radiation image data and visible light reflection image data in each multimodal image data set for analysis; the noise feature is represented by pixel clusters formed by sensor circuit interference; the occlusion feature is represented by the range value of the occlusion area projected by a fixed object in the environment; store and record the pixel coordinate range, gray value distribution interval, and texture feature parameters of the noise and occlusion features to form an initial noise and occlusion dictionary;

[0014] Step S2-2: Extract feature points from the infrared radiation image data and visible light reflection image data in the multimodal image data set of the current metal sheet processing and inspection cycle, respectively. The feature points include pixel coordinates, gray values ​​and texture feature parameters.

[0015] Step S2-3: For each feature point in the infrared radiation image data, delineate a feature sub-region centered on that feature point. The boundary of the feature sub-region is determined by judging the gradient change rate of the gray values ​​of surrounding pixels. When the gradient change rate exceeds a preset threshold, the pixel is included in the feature sub-region. For each feature point in the visible light reflectance image data, delineate the feature sub-region in the same way. Extract the gray values ​​and texture feature parameters of all pixels in each feature sub-region in the two modal images respectively, and calculate the gray value distribution range and texture feature parameter set in the two modal images.

[0016] Step S2-4: Using feature vector similarity matching technology, the pixel relative coordinate distribution, gray value distribution range, gray value gradient change features, and texture feature parameter set of each feature sub-region in the infrared radiation image data and visible light reflectance image data are compared with the corresponding features stored in the initial noise occlusion dictionary after being dimensionless, and the comprehensive matching degree is calculated. The comprehensive matching degree is obtained by weighting pixel distribution similarity, gray value range overlap, gradient curve matching degree, and texture feature matching degree. Pixel distribution similarity is the proportion of pixel coordinate range overlap between the feature sub-region and the corresponding region in the initial noise occlusion dictionary; gray value range overlap is the proportion of the intersection of the gray value distribution ranges of the two regions to their union; gradient curve matching degree is the reciprocal of the mean square error between the gray value gradient change curve in the feature sub-region and the corresponding curve in the initial noise occlusion dictionary; and texture feature matching degree is the reciprocal of the Euclidean distance between the texture feature parameter set of the feature sub-region and the corresponding parameter set in the initial noise occlusion dictionary. Based on a preset comprehensive matching degree threshold, feature sub-regions that do not exceed the comprehensive matching degree threshold are classified as abnormal image regions.

[0017] The overall matching degree is calculated using the following formula:

[0018] M=w1×S1+w2×S2+w3×S3+w4×S4;

[0019] In the formula, M represents the comprehensive matching degree of the feature sub-region, and the larger the value of the comprehensive matching degree, the higher the matching degree; S1 represents the pixel distribution similarity of the feature sub-region; w1 represents the pixel distribution similarity weight value of the feature sub-region; S2 represents the gray-level interval overlap; w2 represents the gray-level interval overlap weight value; S3 represents the gradient curve matching degree; w3 represents the gradient curve matching degree weight value; S4 represents the texture feature matching degree; w4 represents the texture feature matching degree weight value.

[0020] The preset comprehensive matching degree threshold is the critical value for distinguishing between normal and abnormal areas. It is determined by combining the characteristics of metal sheet production scenarios and historical data: Based on the normal noise and occlusion area data of multiple historical processing and inspection cycles, the comprehensive matching degree distribution is statistically analyzed. The mean minus three times the standard deviation can be taken as the initial lower limit of the threshold through the three sigma principle. At the same time, plate image samples with known quality defects are introduced for verification to ensure that the matching degree of defective areas is lower than the preset comprehensive matching degree threshold. Adaptability adjustments are made according to production accuracy requirements, and periodic calibration and manual intervention correction are supported to adapt to environmental changes.

[0021] Pixel distribution similarity is used to represent the proportion of overlap between the pixel coordinate ranges of a feature sub-region and the corresponding region in the initial noise occlusion dictionary. It is calculated by dividing the area of ​​the overlapping pixel coordinate ranges of the feature sub-region and the corresponding region in the initial noise occlusion dictionary by the total area of ​​the pixel coordinate ranges of these two regions, with the area measured in pixels. The pixel distribution similarity value ranges from 0 to 1; a higher value indicates a more consistent coordinate distribution between the two regions. Pixel distribution similarity is achieved through histogram analysis, specifically by calculating the histogram intersection or Bach coefficient of the feature sub-region and the dictionary features to quantify the degree of overlap in the probability distribution of pixel values.

[0022] Gray-scale interval overlap is used to represent the proportion of overlap between the gray-scale value distribution intervals of a feature sub-region and the corresponding region in the initial noise occlusion dictionary. It is calculated as follows: the intersection length of the gray-scale value distribution intervals of the feature sub-region and the corresponding region in the initial noise occlusion dictionary is divided by the union length of these two intervals. The intersection length is the length of the overlapping portion of the two intervals, specifically the smaller of the two interval maximum values ​​minus the larger of the two interval minimum values; if the result is negative, it is set to 0. The union length is the total length covered after the two intervals are merged, specifically the larger of the two interval maximum values ​​minus the smaller of the two interval minimum values. The gray-scale interval overlap value ranges from 0 to 1; the larger the value, the more consistent the gray-scale distributions of the two regions. Gray-scale interval overlap is based on interval analysis technology and is calculated by dividing the length of the overlapping interval of the gray-scale value ranges of the two regions by the total coverage length.

[0023] The gradient curve fit is calculated by extracting gradient features using the gradient operator, specifically by the reciprocal of the mean square error of the gradient curve.

[0024] The formula for calculating the mean squared error is as follows:

[0025] ;

[0026] In the formula, E represents the mean square error of the gradient curve; N represents the total number of sampling points of the gray-level gradient curve within the feature sub-region; y i Y is represented as the gray-level gradient value of the i-th sampling point in the feature sub-region; i This is represented as the grayscale gradient value of the i-th sampling point in the corresponding feature sub-region in the initial noise occlusion dictionary;

[0027] Texture feature matching degree is obtained by extracting texture feature vectors through gray-level co-occurrence matrix and local binary mode, and then calculating similarity using cosine similarity or the inverse of Euclidean distance;

[0028] An initial noise occlusion dictionary was constructed by analyzing multimodal image datasets from multiple historical metal sheet processing and inspection cycles. Feature points were extracted from the current multimodal image dataset and feature sub-regions were delineated. Feature vector similarity matching technology was used to compare the feature sub-regions with features in the initial noise occlusion dictionary to calculate the comprehensive matching degree. Abnormal image regions were determined based on preset thresholds, thus achieving accurate identification of abnormal regions in multimodal image data.

[0029] Step S3: Construct a real-time dictionary update pool to store the identifiers and corresponding access counts of each atom accessed from the initial noise occlusion dictionary during the current metal sheet processing and inspection cycle; and make trend predictions by analyzing the access counts, and sort and update the atoms in the initial noise occlusion dictionary according to the trend predictions to obtain the real-time noise occlusion dictionary.

[0030] Step S3-1: Initialize the real-time dictionary database update pool, and set the data storage fields to include the unique atomic identifier, the cumulative number of accesses, and the access timestamp. The atom is represented as follows, where the unique atomic identifier corresponds one-to-one with the noise features and occlusion features stored in the initial noise occlusion dictionary database.

[0031] Step S3-2: During the current metal sheet processing and inspection cycle, whenever an atom in the initial noise masking dictionary is accessed once, the real-time dictionary update pool increments the cumulative access count of that atom by one and updates the timestamp of the most recent access to the timestamp of the current access.

[0032] Step S3-3: Select any atom in the initial noise masking dictionary as the research object, obtain the real-time dictionary update pool data in each historical metal sheet processing and inspection cycle, and predict the number of times the atom will be accessed in the current metal sheet processing and inspection cycle by combining the access count of the atom in each real-time dictionary update pool through deep machine learning. This number is recorded as the predicted access count.

[0033] Deep machine learning involves training convolutional neural networks on the time series of atomic access counts in historical detection cycles to learn their periodicity, trends, and other patterns, thereby predicting the access count in the current cycle.

[0034] Step S3-4: Extract the predicted access count of atoms in the real-time dictionary update pool, select the predicted access count as the primary sorting criterion, and select the most recent access timestamp as the secondary sorting criterion;

[0035] Step S3-5: Reorder the atoms in the initial noise occlusion dictionary according to the sorting criteria to generate a real-time noise occlusion dictionary; among them, atoms with more predicted access times are arranged first, and atoms with the same number of predicted access times are arranged first according to the more recent access timestamp.

[0036] By constructing a real-time dictionary update pool to store the identifiers and corresponding access counts of each atom in the initial noise occlusion dictionary, and using historical data to predict the access count of atoms in the current detection cycle through deep machine learning, the atoms in the initial noise occlusion dictionary are reordered based on the predicted access count as the primary basis and the most recent access timestamp as the secondary basis, thus obtaining the real-time noise occlusion dictionary and realizing the dynamic updating of the noise occlusion dictionary.

[0037] Step S4: Use the real-time noise occlusion dictionary to perform real-time detection on the processed metal sheet, traverse and read the abnormal image regions in the multimodal image data set, and at the same time select the normal image regions to perform cross-modal feature mapping relationship analysis to obtain the cross-modal feature mapping relationship.

[0038] A real-time noise occlusion dictionary is used to detect multimodal image datasets of processed metal sheets in real time. Abnormal image regions are read through the data, while normal image regions are extracted from infrared radiation image data and visible light reflection image data. The coordinates of the two are compared, and the region with completely overlapping coordinate ranges is selected as the cross-modal analysis sample region. The sample region is divided into analysis units according to fixed pixel blocks. The average gray value, texture entropy value, and gradient direction histogram of infrared radiation image data are extracted as infrared modal feature vectors for each analysis unit, and the average gray value, texture entropy value, and gradient direction histogram of visible light reflection image data are extracted as visible light modal feature vectors.

[0039] After dimensionless processing of the infrared modal feature vectors and visible light modal feature vectors of all analysis units, they are paired to form a set of feature vector pairs. The set of feature vector pairs is trained by a least squares support vector machine to obtain a bidirectional mapping function between the infrared modal feature vectors and the visible light modal feature vectors, thus forming a cross-modal feature mapping relationship.

[0040] A real-time noise occlusion dictionary is used to detect and traverse abnormal image regions in the multimodal image dataset of the processed metal sheet in real time. Normal image regions are extracted from infrared and visible light images. Regions with completely overlapping coordinate ranges are selected as cross-modal analysis sample regions and divided into analysis units. Infrared and visible light modal feature vectors of each unit are extracted and paired after dimensionless processing to form a set of feature vector pairs. A bidirectional mapping function is obtained by training with a least squares support vector machine to form a cross-modal feature mapping relationship.

[0041] Step S5: Based on the cross-modal feature mapping relationship, noise occlusion repair is performed on the abnormal image regions detected in real time to obtain noise-free image data, which is then transmitted to the metal sheet production quality inspection and analysis module.

[0042] Step S5-1: Extract the pixel coordinate range of the abnormal image region detected in real time, and analyze it using infrared radiation image data and visible light reflection image data respectively. If only the infrared radiation image data has an abnormal image region and the corresponding coordinate region of the visible light reflection image data is normal, it is determined to be an infrared single-mode anomaly; if only the visible light reflection image data has an abnormal image region and the corresponding coordinate region of the infrared radiation image data is normal, it is determined to be a visible light single-mode anomaly; if both the corresponding coordinate regions have abnormal image regions, it is determined to be a dual-mode anomaly.

[0043] Step S5-2: For infrared single-mode anomalies, extract the feature vector of the normal region that overlaps with the coordinate range of the abnormal region in the visible light reflectance image data, input it into the visible light to infrared mapping function obtained in step S4, and generate an infrared repair feature vector; replace the pixel value of the abnormal region in the infrared radiation image data with the pixel value corresponding to this vector to complete the infrared single-mode anomaly repair; For visible light single-mode anomalies, extract the feature vector of the normal region that overlaps with the coordinate range of the abnormal region in the infrared radiation image data, input it into the infrared to visible light mapping function obtained in step S4, and generate a visible light repair feature vector; replace the pixel value of the abnormal region in the visible light reflectance image data with the pixel value corresponding to this vector to complete the visible light single-mode anomaly repair.

[0044] Step S5-3: For bimodal anomalies, determine the normal region within the pixel coordinate range surrounding the anomaly region. The normal region is the region whose distance from the edge of the anomaly region is less than n pixels and which was detected as normal in step S4. Extract the gray value distribution and texture feature parameters of the infrared radiation image data within the normal region, and use bilinear interpolation to fill the pixels in the infrared anomaly region to obtain the initial infrared repair image. Extract the gray value distribution and texture feature parameters of the visible light reflection image data within the normal region, and use bilinear interpolation to fill the pixels in the visible light anomaly region to obtain the initial visible light repair image.

[0045] Step S5-4: Calculate the pixel residuals in the abnormal region between the initial infrared restoration image and the initial visible light restoration image, i.e., the absolute value of the difference in grayscale values ​​of the corresponding pixels; extract the pixel residuals of the infrared and visible light images in the normal region from step S5-3 as the normal residual benchmark; adjust the pixel values ​​of the initial infrared restoration image and the initial visible light restoration image according to the cross-modal feature mapping relationship obtained in step S4, so that the pixel residuals in the abnormal region are less than the normal residual benchmark, and the pixel value adjustment is completed.

[0046] The pixel residual is calculated using the following formula:

[0047] ;

[0048] In the formula, R(x, y) represents the pixel residual at coordinates (x, y), with a value range of [0, 255]. The smaller the pixel residual value, the closer the pixels at that coordinate are; (x, y) represents the coordinates of a pixel in the image, I ir Represented as the grayscale value of the infrared radiation image at coordinates (x, y); I vis This is represented as the gray value of the visible light reflectance image at coordinates (x, y);

[0049] Step S5-5: After the pixel value adjustment is completed, the final repaired image data of the initial infrared repair image and the initial visible light repair image are obtained and transmitted to the metal sheet production quality inspection and analysis module.

[0050] Based on the cross-modal feature mapping relationship, the abnormal image regions detected in real time are first determined as infrared single-mode anomalies, visible light single-mode anomalies, or dual-mode anomalies. Then, for infrared single-mode anomalies, a repair feature vector is generated using the visible light to infrared mapping function to replace the abnormal pixel values. For visible light single-mode anomalies, a repair feature vector is generated using the infrared to visible light mapping function to replace the abnormal pixel values. For dual-mode anomalies, the initial repair image is obtained through bilinear interpolation, and the pixel values ​​are adjusted based on the normal residual benchmark. Finally, the repaired noise-free occlusion image data is transmitted to the metal sheet production quality inspection and analysis module.

[0051] Furthermore, an IoT-based metal sheet production quality inspection system includes a multimodal image acquisition module, an abnormal region identification module, a real-time dictionary update module, a cross-modal feature analysis module, and an abnormal image repair module.

[0052] The multimodal image acquisition module is used to acquire infrared thermal imaging and visible light image data of metal plates according to the detection cycle, and form a multimodal image data set after time alignment; the abnormal region identification module is used to construct an initial noise occlusion dictionary library using historical multimodal image data, and to obtain abnormal image regions through current data; the real-time dictionary update module is used to construct a real-time dictionary library update pool, and to obtain a real-time noise occlusion dictionary library by predicting and updating the initial noise occlusion dictionary library through access frequency trends; the cross-modal feature analysis module is used to detect abnormal image regions in real time using the real-time noise occlusion dictionary library, and to analyze normal regions to obtain cross-modal feature mapping relationships; the abnormal image repair module is used to repair noise occlusion in abnormal image regions according to cross-modal feature mapping relationships, and to transmit the repaired image data to the detection and analysis module;

[0053] The output of the multimodal image acquisition module is electrically connected to the input of the abnormal region identification module; the output of the abnormal region identification module is electrically connected to the input of the real-time dictionary update module; the output of the real-time dictionary update module is electrically connected to the input of the cross-modal feature analysis module; and the output of the cross-modal feature analysis module is electrically connected to the input of the abnormal image repair module.

[0054] The multimodal image acquisition module includes an image sensing acquisition unit and a spatiotemporal calibration processing unit; the image sensing acquisition unit is used to acquire infrared radiation image data and visible light reflection image data of the metal plate through an infrared thermal imaging sensor and a visible light image sensor; the spatiotemporal calibration processing unit is used to perform time alignment of the infrared and visible light image data based on timestamp data and correct coordinate system deviations.

[0055] The abnormal region identification module includes an initial dictionary construction unit and an abnormal region detection unit; the initial dictionary construction unit is used to analyze the multimodal image data set of multiple historical detection periods to obtain an initial noise occlusion dictionary library; the abnormal region detection unit is used to obtain abnormal image regions from the multimodal image data set of the current detection period through feature matching and multimodal residual analysis.

[0056] The real-time dictionary update module includes an access data storage unit and a dictionary sorting update unit. The access data storage unit is used to store the identifiers and corresponding access counts of each atom in the initial noise occlusion dictionary library within the current detection period. The dictionary sorting update unit is used to analyze the access counts to make trend predictions and to sort and update the atoms in the initial noise occlusion dictionary library according to the predictions.

[0057] The cross-modal feature analysis module includes a real-time image detection unit and a feature mapping construction unit. The real-time image detection unit is used to traverse and read abnormal image regions in the multimodal image dataset using a real-time noise occlusion dictionary. The feature mapping construction unit is used to select normal image regions and perform cross-modal feature mapping relationship analysis to obtain cross-modal feature mapping relationships.

[0058] The abnormal image repair module includes an image repair processing unit and a detection data transmission unit; the image repair processing unit is used to repair noise occlusion in the abnormal image area detected in real time to obtain noise-free image data; the detection data transmission unit is used to transmit the noise-free image data to the metal sheet production quality inspection and analysis module.

[0059] Compared with the prior art, the beneficial effects of the present invention are:

[0060] 1. This invention collects infrared thermal imaging and visible light image data of metal sheets by setting a processing and inspection cycle, achieves time alignment based on timestamps, and uses inherent geometric features to correct coordinate system deviations to form a multimodal image data set. This solves the problem of spatiotemporal inconsistency of image data of different modalities, and provides a unified benchmark image data foundation for subsequent anomaly identification, feature analysis and repair, ensuring the accuracy of multi-source data fusion analysis.

[0061] 2. This invention constructs an initial noise occlusion dictionary, combines feature matching and multimodal residual analysis of current detection period data to identify abnormal regions, and simultaneously constructs a real-time dictionary update pool, dynamically updating the dictionary ranking based on access frequency trends. This makes noise occlusion identification more adaptable to real-time production changes, improves the timeliness and accuracy of abnormal region detection, and reduces the problem of insufficient adaptability of fixed dictionary libraries to dynamic scenes.

[0062] 3. This invention establishes a cross-modal feature mapping relationship and adopts corresponding repair strategies for infrared single-mode, visible light single-mode, and dual-mode anomalies. For single-mode anomalies, cross-modal mapping is used to generate repair feature vectors, and for dual-mode anomalies, bilinear interpolation and residual adjustment are combined to optimize the repair effect. This ensures the targetedness and reliability of anomaly area repair and provides high-quality image data without noise occlusion for the quality detection and analysis module. Attached Figure Description

[0063] Figure 1 This is a schematic flowchart of a metal sheet production quality inspection method based on the Internet of Things according to the present invention.

[0064] Figure 2 This is a schematic diagram of the structure of a metal sheet production quality inspection system based on the Internet of Things according to the present invention. Detailed Implementation

[0065] 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.

[0066] Example 1: As Figure 1 As shown, the present invention provides a technical solution: a method for quality inspection of metal sheet production based on the Internet of Things. The method for quality inspection of metal sheet production includes the following steps:

[0067] Step S1: Set the metal sheet processing and inspection cycle, collect infrared thermal image data and visible light image data of the metal sheet to be processed according to the metal sheet processing and inspection cycle, and perform time alignment of the collected infrared thermal image data and visible light image data based on the timestamp data to obtain a multimodal image data set of the metal sheet to be processed.

[0068] Step S1-1: Obtain infrared radiation image data of the metal sheet being processed by installing an infrared thermal imaging sensor; obtain visible light reflection image data of the metal sheet being processed by installing a visible light image sensor; adjust the viewing angles of the infrared thermal imaging sensor and the visible light image sensor to be consistent.

[0069] Steps S1-2: During the acquisition process, the infrared thermal imaging sensor and the visible light image sensor synchronously record the timestamp data corresponding to the acquisition time for each frame of image data generated, and the storage format of the timestamp data is the same.

[0070] Steps S1-3: Extract timestamp data from infrared radiation image data to form an infrared time series; extract timestamp data from visible light reflectance image data to form a visible light time series; align the timestamps in the infrared time series and the visible light time series, and pair infrared radiation image data and visible light reflectance image data at the same time.

[0071] Step S1-4: For the paired image data, select the inherent geometric features of the processed metal sheet to correct the coordinate system deviation between the infrared radiation image data and the visible light reflection image data, forming a multimodal image data set. The inherent geometric features of the processed metal sheet are represented as the contour features of the processed metal sheet. The specific correction process is to adjust the pixel coordinates of the infrared radiation image data and the visible light reflection image data so that the pixel coordinates of the inherent geometric features are at the same coordinate position.

[0072] In practical implementation, a synchronized multimodal image acquisition and calibration mechanism is used to achieve spatiotemporal unification of data from different sensors. During operation, it is essential to ensure strict consistency of the viewing angles of the infrared thermal imaging sensor and the visible light image sensor, accurate synchronization of timestamp recordings, and coordinate system correction based on the contour features of the metal sheet. This ensures a reliable image data foundation for subsequent analysis.

[0073] Step S2: Analyze the multimodal image data sets from multiple historical metal sheet processing and inspection cycles to obtain an initial noise occlusion dictionary; and analyze the multimodal image data sets from the current metal sheet processing and inspection cycle, using feature matching and multimodal residual analysis to obtain abnormal image regions in the multimodal image data sets.

[0074] Step S2-1: Retrieve multimodal image data sets from multiple historical metal sheet processing and inspection cycles, and extract noise and occlusion features from infrared radiation image data and visible light reflection image data in each multimodal image data set for analysis; the noise feature is represented by pixel clusters formed by sensor circuit interference; the occlusion feature is represented by the range value of the occlusion area projected by a fixed object in the environment; store and record the pixel coordinate range, gray value distribution interval, and texture feature parameters of the noise and occlusion features to form an initial noise and occlusion dictionary;

[0075] Step S2-2: Extract feature points from the infrared radiation image data and visible light reflection image data in the multimodal image data set of the current metal sheet processing and inspection cycle, respectively. The feature points include pixel coordinates, gray values ​​and texture feature parameters.

[0076] Step S2-3: For each feature point in the infrared radiation image data, delineate a feature sub-region centered on that feature point. The boundary of the feature sub-region is determined by judging the gradient change rate of the gray values ​​of surrounding pixels. When the gradient change rate exceeds a preset threshold, the pixel is included in the feature sub-region. For each feature point in the visible light reflectance image data, delineate the feature sub-region in the same way. Extract the gray values ​​and texture feature parameters of all pixels in each feature sub-region in the two modal images respectively, and calculate the gray value distribution range and texture feature parameter set in the two modal images.

[0077] Step S2-4: Using feature vector similarity matching technology, the pixel relative coordinate distribution, gray value distribution range, gray value gradient change features, and texture feature parameter set of each feature sub-region in the infrared radiation image data and visible light reflectance image data are compared with the corresponding features stored in the initial noise occlusion dictionary after being dimensionless, and the comprehensive matching degree is calculated. The comprehensive matching degree is obtained by weighting pixel distribution similarity, gray value range overlap, gradient curve matching degree, and texture feature matching degree. Pixel distribution similarity is the proportion of pixel coordinate range overlap between the feature sub-region and the corresponding region in the initial noise occlusion dictionary; gray value range overlap is the proportion of the intersection of the gray value distribution ranges of the two regions to their union; gradient curve matching degree is the reciprocal of the mean square error between the gray value gradient change curve in the feature sub-region and the corresponding curve in the initial noise occlusion dictionary; and texture feature matching degree is the reciprocal of the Euclidean distance between the texture feature parameter set of the feature sub-region and the corresponding parameter set in the initial noise occlusion dictionary. Based on a preset comprehensive matching degree threshold, feature sub-regions that do not exceed the comprehensive matching degree threshold are classified as abnormal image regions.

[0078] In practice, anomalies are identified by combining a historical noise occlusion feature database with real-time feature matching. In practical applications, the gradient change rate threshold for feature sub-regions must be set to match the material characteristics of the metal sheet, and the weights of each component in the overall matching score should be adjusted according to the actual impact of noise and occlusion in the production environment to ensure the accuracy of anomaly region identification.

[0079] Step S3: Construct a real-time dictionary update pool to store the identifiers and corresponding access counts of each atom accessed from the initial noise occlusion dictionary during the current metal sheet processing and inspection cycle; and make trend predictions by analyzing the access counts, and sort and update the atoms in the initial noise occlusion dictionary according to the trend predictions to obtain the real-time noise occlusion dictionary.

[0080] Step S3-1: Initialize the real-time dictionary database update pool, and set the data storage fields to include the unique atomic identifier, the cumulative number of accesses, and the access timestamp. The atom is represented as follows, where the unique atomic identifier corresponds one-to-one with the noise features and occlusion features stored in the initial noise occlusion dictionary database.

[0081] Step S3-2: During the current metal sheet processing and inspection cycle, whenever an atom in the initial noise masking dictionary is accessed once, the real-time dictionary update pool increments the cumulative access count of that atom by one and updates the timestamp of the most recent access to the timestamp of the current access.

[0082] Step S3-3: Select any atom in the initial noise masking dictionary as the research object, obtain the real-time dictionary update pool data in each historical metal sheet processing and inspection cycle, and predict the number of times the atom will be accessed in the current metal sheet processing and inspection cycle by combining the access count of the atom in each real-time dictionary update pool through deep machine learning. This number is recorded as the predicted access count.

[0083] Step S3-4: Extract the predicted access count of atoms in the real-time dictionary update pool, select the predicted access count as the primary sorting criterion, and select the most recent access timestamp as the secondary sorting criterion;

[0084] Step S3-5: Reorder the atoms in the initial noise occlusion dictionary according to the sorting criteria to generate a real-time noise occlusion dictionary; among them, atoms with more predicted access times are arranged first, and atoms with the same number of predicted access times are arranged first according to the more recent access timestamp.

[0085] In practical implementation, a dynamic update mechanism is used to improve the real-time adaptability of the dictionary database. During implementation, it is essential to ensure that the real-time dictionary database update pool accurately records the number of atomic accesses. The training data for the deep machine learning model must cover historical access records under different production conditions, and the sorting rules must strictly follow the priority of predicted access counts and the most recent access timestamp.

[0086] Step S4: Use the real-time noise occlusion dictionary to perform real-time detection on the processed metal sheet, traverse and read the abnormal image regions in the multimodal image data set, and at the same time select the normal image regions to perform cross-modal feature mapping relationship analysis to obtain the cross-modal feature mapping relationship.

[0087] A real-time noise occlusion dictionary is used to detect multimodal image datasets of processed metal sheets in real time. Abnormal image regions are read through the data, while normal image regions are extracted from infrared radiation image data and visible light reflection image data. The coordinates of the two are compared, and the region with completely overlapping coordinate ranges is selected as the cross-modal analysis sample region. The sample region is divided into analysis units according to fixed pixel blocks. The average gray value, texture entropy value, and gradient direction histogram of infrared radiation image data are extracted as infrared modal feature vectors for each analysis unit, and the average gray value, texture entropy value, and gradient direction histogram of visible light reflection image data are extracted as visible light modal feature vectors.

[0088] After dimensionless processing of the infrared modal feature vectors and visible light modal feature vectors of all analysis units, they are paired to form a set of feature vector pairs. The set of feature vector pairs is trained by a least squares support vector machine to obtain a bidirectional mapping function between the infrared modal feature vectors and the visible light modal feature vectors, thus forming a cross-modal feature mapping relationship.

[0089] In practical implementation, the focus is on constructing cross-modal feature mapping relationships. First, anomaly regions need to be accurately located using a real-time noise occlusion dictionary. Then, normal regions with completely overlapping coordinates are selected as analysis samples. Feature vector extraction should fully include the average gray value, texture entropy value, and gradient direction histogram. The training process of the least squares support vector machine must ensure the representativeness of the feature vectors to the set in order to establish a reliable bidirectional mapping function.

[0090] Step S5: Based on the cross-modal feature mapping relationship, noise occlusion repair is performed on the abnormal image regions detected in real time to obtain noise-free image data, which is then transmitted to the metal sheet production quality inspection and analysis module.

[0091] Step S5-1: Extract the pixel coordinate range of the abnormal image region detected in real time, and analyze it using infrared radiation image data and visible light reflection image data respectively. If only the infrared radiation image data has an abnormal image region and the corresponding coordinate region of the visible light reflection image data is normal, it is determined to be an infrared single-mode anomaly; if only the visible light reflection image data has an abnormal image region and the corresponding coordinate region of the infrared radiation image data is normal, it is determined to be a visible light single-mode anomaly; if both the corresponding coordinate regions have abnormal image regions, it is determined to be a dual-mode anomaly.

[0092] Step S5-2: For infrared single-mode anomalies, extract the feature vector of the normal region that overlaps with the coordinate range of the abnormal region in the visible light reflectance image data, input it into the visible light to infrared mapping function obtained in step S4, and generate an infrared repair feature vector; replace the pixel value of the abnormal region in the infrared radiation image data with the pixel value corresponding to this vector to complete the infrared single-mode anomaly repair; For visible light single-mode anomalies, extract the feature vector of the normal region that overlaps with the coordinate range of the abnormal region in the infrared radiation image data, input it into the infrared to visible light mapping function obtained in step S4, and generate a visible light repair feature vector; replace the pixel value of the abnormal region in the visible light reflectance image data with the pixel value corresponding to this vector to complete the visible light single-mode anomaly repair.

[0093] Step S5-3: For bimodal anomalies, determine the normal region within the pixel coordinate range surrounding the anomaly region. The normal region is the region whose distance from the edge of the anomaly region is less than n pixels and which was detected as normal in step S4. Extract the gray value distribution and texture feature parameters of the infrared radiation image data within the normal region, and use bilinear interpolation to fill the pixels in the infrared anomaly region to obtain the initial infrared repair image. Extract the gray value distribution and texture feature parameters of the visible light reflection image data within the normal region, and use bilinear interpolation to fill the pixels in the visible light anomaly region to obtain the initial visible light repair image.

[0094] Step S5-4: Calculate the pixel residuals in the abnormal region between the initial infrared restoration image and the initial visible light restoration image, i.e., the absolute value of the difference in grayscale values ​​of the corresponding pixels; extract the pixel residuals of the infrared and visible light images in the normal region from step S5-3 as the normal residual benchmark; adjust the pixel values ​​of the initial infrared restoration image and the initial visible light restoration image according to the cross-modal feature mapping relationship obtained in step S4, so that the pixel residuals in the abnormal region are less than the normal residual benchmark, and the pixel value adjustment is completed.

[0095] Step S5-5: After the pixel value adjustment is completed, the final repaired image data of the initial infrared repair image and the initial visible light repair image are obtained and transmitted to the metal sheet production quality inspection and analysis module.

[0096] In practice, differentiated repair strategies are adopted for different types of abnormal regions. During implementation, it is crucial to accurately determine the types of infrared single-mode, visible light single-mode, and dual-mode anomalies. Single-mode repair requires strictly generating repair vectors based on cross-modal feature mapping relationships, while dual-mode repair must ensure the rationality of normal region selection and that pixel residual adjustments meet normal benchmarks, ultimately providing the quality inspection module with noise-free, occluded image data.

[0097] Example 2, as Figure 2 As shown, the present invention provides a metal sheet production quality inspection system based on the Internet of Things. The metal sheet production quality inspection system includes a multimodal image acquisition module, an abnormal area identification module, a real-time dictionary update module, a cross-modal feature analysis module, and an abnormal image repair module.

[0098] The multimodal image acquisition module is used to acquire infrared thermal imaging and visible light image data of metal plates according to the detection cycle, and form a multimodal image data set after time alignment; the abnormal region identification module is used to construct an initial noise occlusion dictionary library using historical multimodal image data, and to obtain abnormal image regions through current data; the real-time dictionary update module is used to construct a real-time dictionary library update pool, and to obtain a real-time noise occlusion dictionary library by predicting and updating the initial noise occlusion dictionary library through access frequency trends; the cross-modal feature analysis module is used to detect abnormal image regions in real time using the real-time noise occlusion dictionary library, and to analyze normal regions to obtain cross-modal feature mapping relationships; the abnormal image repair module is used to repair noise occlusion in abnormal image regions according to cross-modal feature mapping relationships, and to transmit the repaired image data to the detection and analysis module;

[0099] The output of the multimodal image acquisition module is electrically connected to the input of the abnormal region identification module; the output of the abnormal region identification module is electrically connected to the input of the real-time dictionary update module; the output of the real-time dictionary update module is electrically connected to the input of the cross-modal feature analysis module; and the output of the cross-modal feature analysis module is electrically connected to the input of the abnormal image repair module.

[0100] The multimodal image acquisition module includes an image sensing acquisition unit and a spatiotemporal calibration processing unit; the image sensing acquisition unit is used to acquire infrared radiation image data and visible light reflection image data of the metal plate through an infrared thermal imaging sensor and a visible light image sensor; the spatiotemporal calibration processing unit is used to perform time alignment of the infrared and visible light image data based on timestamp data and correct coordinate system deviations.

[0101] The abnormal region identification module includes an initial dictionary construction unit and an abnormal region detection unit; the initial dictionary construction unit is used to analyze the multimodal image data set of multiple historical detection periods to obtain an initial noise occlusion dictionary library; the abnormal region detection unit is used to obtain abnormal image regions from the multimodal image data set of the current detection period through feature matching and multimodal residual analysis.

[0102] The real-time dictionary update module includes an access data storage unit and a dictionary sorting update unit. The access data storage unit is used to store the identifiers and corresponding access counts of each atom in the initial noise occlusion dictionary library within the current detection period. The dictionary sorting update unit is used to analyze the access counts to make trend predictions and to sort and update the atoms in the initial noise occlusion dictionary library according to the predictions.

[0103] The cross-modal feature analysis module includes a real-time image detection unit and a feature mapping construction unit. The real-time image detection unit is used to traverse and read abnormal image regions in the multimodal image dataset using a real-time noise occlusion dictionary. The feature mapping construction unit is used to select normal image regions and perform cross-modal feature mapping relationship analysis to obtain cross-modal feature mapping relationships.

[0104] The abnormal image repair module includes an image repair processing unit and a detection data transmission unit; the image repair processing unit is used to repair noise occlusion in the abnormal image area detected in real time to obtain noise-free image data; the detection data transmission unit is used to transmit the noise-free image data to the metal sheet production quality inspection and analysis module.

[0105] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

Claims

1. A method for quality inspection of metal sheet production based on the Internet of Things, characterized in that: The method for testing the production quality of metal sheets includes the following steps: Step S1: Set the metal sheet processing and inspection cycle, collect infrared thermal image data and visible light image data of the metal sheet to be processed according to the metal sheet processing and inspection cycle, and perform time alignment of the collected infrared thermal image data and visible light image data based on the timestamp data to obtain a multimodal image data set of the metal sheet to be processed. Step S2: Analyze the multimodal image data sets from multiple historical metal sheet processing and inspection cycles to obtain an initial noise occlusion dictionary; and analyze the multimodal image data sets from the current metal sheet processing and inspection cycle, using feature matching and multimodal residual analysis to obtain abnormal image regions in the multimodal image data sets. Step S3: Construct a real-time dictionary update pool to store the identifiers and corresponding access counts of each atom accessed from the initial noise occlusion dictionary during the current metal sheet processing and inspection cycle; and make trend predictions by analyzing the access counts, and sort and update the atoms in the initial noise occlusion dictionary according to the trend predictions to obtain the real-time noise occlusion dictionary. Step S4: Use the real-time noise occlusion dictionary to perform real-time detection on the processed metal sheet, traverse and read the abnormal image regions in the multimodal image data set, and at the same time select the normal image regions to perform cross-modal feature mapping relationship analysis to obtain the cross-modal feature mapping relationship. Step S5: Based on the cross-modal feature mapping relationship, noise occlusion repair is performed on the abnormal image regions detected in real time to obtain noise-free image data, which is then transmitted to the metal sheet production quality inspection and analysis module. Step S5-1: Extract the pixel coordinate range of the abnormal image region detected in real time, and analyze it using infrared radiation image data and visible light reflection image data respectively. If only the infrared radiation image data has an abnormal image region and the corresponding coordinate region of the visible light reflection image data is normal, it is determined to be an infrared single-mode anomaly; if only the visible light reflection image data has an abnormal image region and the corresponding coordinate region of the infrared radiation image data is normal, it is determined to be a visible light single-mode anomaly; if both the corresponding coordinate regions have abnormal image regions, it is determined to be a dual-mode anomaly. Step S5-2: For infrared single-mode anomalies, extract the feature vector of the normal region that overlaps with the coordinate range of the abnormal region in the visible light reflectance image data, input it into the visible light to infrared mapping function obtained in step S4, and generate an infrared repair feature vector; replace the pixel value of the abnormal region in the infrared radiation image data with the pixel value corresponding to this vector to complete the infrared single-mode anomaly repair; For visible light single-mode anomalies, extract the feature vector of the normal region that overlaps with the coordinate range of the abnormal region in the infrared radiation image data, input it into the infrared to visible light mapping function obtained in step S4, and generate a visible light repair feature vector; replace the pixel value of the abnormal region in the visible light reflectance image data with the pixel value corresponding to this vector to complete the visible light single-mode anomaly repair. Step S5-3: For bimodal anomalies, determine the normal region within the pixel coordinate range surrounding the anomaly region. The normal region is the region whose distance from the edge of the anomaly region is less than n pixels and which was detected as normal in step S4. Extract the gray value distribution and texture feature parameters of the infrared radiation image data within the normal region, and use bilinear interpolation to fill the pixels in the infrared anomaly region to obtain the initial infrared repair image. Extract the gray value distribution and texture feature parameters of the visible light reflection image data within the normal region, and use bilinear interpolation to fill the pixels in the visible light anomaly region to obtain the initial visible light repair image. Step S5-4: Calculate the pixel residuals in the abnormal region between the initial infrared restoration image and the initial visible light restoration image, i.e., the absolute value of the difference in grayscale values ​​of the corresponding pixels; extract the pixel residuals of the infrared and visible light images in the normal region from step S5-3 as the normal residual benchmark; adjust the pixel values ​​of the initial infrared restoration image and the initial visible light restoration image according to the cross-modal feature mapping relationship obtained in step S4, so that the pixel residuals in the abnormal region are less than the normal residual benchmark, and the pixel value adjustment is completed. After step S5-5 and pixel value adjustment, the final repaired image data of the infrared initial repaired image and the visible light initial repaired image are obtained and transmitted to the metal sheet production quality inspection and analysis module.

2. The method for quality inspection of metal sheet production based on the Internet of Things according to claim 1, characterized in that: The specific steps of step S1 are as follows: Step S1-1: Obtain infrared radiation image data of the metal sheet being processed by installing an infrared thermal imaging sensor; obtain visible light reflection image data of the metal sheet being processed by installing a visible light image sensor; adjust the viewing angles of the infrared thermal imaging sensor and the visible light image sensor to be consistent. Steps S1-2: During the acquisition process, the infrared thermal imaging sensor and the visible light image sensor synchronously record the timestamp data corresponding to the acquisition time for each frame of image data generated, and the storage format of the timestamp data is the same. Steps S1-3: Extract timestamp data from infrared radiation image data to form an infrared time series; extract timestamp data from visible light reflectance image data to form a visible light time series; align the timestamps in the infrared time series and the visible light time series, and pair infrared radiation image data and visible light reflectance image data at the same time. Step S1-4: For the paired image data, select the inherent geometric features of the processed metal sheet to correct the coordinate system deviation between the infrared radiation image data and the visible light reflection image data, and form a multimodal image data set. The inherent geometric features of the processed metal sheet are represented as the contour features of the processed metal sheet. The specific process of correction involves adjusting the pixel coordinates of the infrared radiation image data and the visible light reflection image data so that the pixel coordinates of the inherent geometric features are at the same coordinate position.

3. The method for quality inspection of metal sheet production based on the Internet of Things according to claim 2, characterized in that: The specific steps of step S2 are as follows: Step S2-1: Retrieve multimodal image data sets from multiple historical metal sheet processing and inspection cycles, and extract noise and occlusion features from infrared radiation image data and visible light reflection image data in each multimodal image data set for analysis; the noise feature is represented by pixel clusters formed by sensor circuit interference; the occlusion feature is represented by the range value of the occlusion area projected by a fixed object in the environment; store and record the pixel coordinate range, gray value distribution interval, and texture feature parameters of the noise and occlusion features to form an initial noise and occlusion dictionary; Step S2-2: Extract feature points from the infrared radiation image data and visible light reflection image data in the multimodal image data set of the current metal sheet processing and inspection cycle, respectively. The feature points include pixel coordinates, gray values ​​and texture feature parameters. Step S2-3: For each feature point in the infrared radiation image data, delineate a feature sub-region centered on that feature point. The boundary of the feature sub-region is determined by judging the gradient change rate of the gray values ​​of surrounding pixels. When the gradient change rate exceeds a preset threshold, the pixel is included in the feature sub-region. For each feature point in the visible light reflectance image data, delineate the feature sub-region in the same way. Extract the gray values ​​and texture feature parameters of all pixels in each feature sub-region in the two modal images respectively, and calculate the gray value distribution range and texture feature parameter set in the two modal images. Step S2-4: Using feature vector similarity matching technology, the pixel relative coordinate distribution, gray value distribution range, gray value gradient change features, and texture feature parameter set of each feature sub-region in the infrared radiation image data and visible light reflectance image data are compared with the corresponding features stored in the initial noise occlusion dictionary after being dimensionless, and the comprehensive matching degree is calculated. The comprehensive matching degree is calculated using the following formula: M=w1×S1+w2×S2+w3×S3+w4×S4; In the formula, M represents the comprehensive matching degree of the feature sub-region, and the larger the value of the comprehensive matching degree, the higher the matching degree; S1 represents the pixel distribution similarity of the feature sub-region; w1 represents the pixel distribution similarity weight value of the feature sub-region; S2 represents the gray-level interval overlap; w2 represents the gray-level interval overlap weight value; S3 represents the gradient curve matching degree; w3 represents the gradient curve matching degree weight value; and S4 represents the texture feature matching degree. w4 represents the weight value of the texture feature matching degree; Feature sub-regions whose overall matching degree M does not exceed the overall matching degree threshold are classified as abnormal image regions, while feature sub-regions whose overall matching degree M exceeds the overall matching degree threshold are classified as normal image regions.

4. The method for quality inspection of metal sheet production based on the Internet of Things according to claim 3, characterized in that: The specific steps of step S3 are as follows: Step S3-1: Initialize the real-time dictionary database update pool, and set the data storage fields to include the unique atomic identifier, the cumulative number of accesses, and the access timestamp. The atom is represented as follows, where the unique atomic identifier corresponds one-to-one with the noise features and occlusion features stored in the initial noise occlusion dictionary database. Step S3-2: During the current metal sheet processing and inspection cycle, whenever an atom in the initial noise masking dictionary is accessed once, the real-time dictionary update pool increments the cumulative access count of that atom by one and updates the timestamp of the most recent access to the timestamp of the current access. Step S3-3: Select any atom in the initial noise masking dictionary as the research object, obtain the real-time dictionary update pool data in each historical metal sheet processing and inspection cycle, and predict the number of times the atom will be accessed in the current metal sheet processing and inspection cycle by combining the access count of the atom in each real-time dictionary update pool through deep machine learning. This number is recorded as the predicted access count. Step S3-4: Extract the predicted access count of atoms in the real-time dictionary update pool, select the predicted access count as the primary sorting criterion, and select the most recent access timestamp as the secondary sorting criterion; Step S3-5: Reorder the atoms in the initial noise occlusion dictionary according to the sorting criteria to generate a real-time noise occlusion dictionary; the atoms with more predicted access times are arranged first, and the atoms with the same number of predicted access times are arranged first according to the more recent access timestamp.

5. The method for quality inspection of metal sheet production based on the Internet of Things according to claim 4, characterized in that: In step S4, the multimodal image data set of the processed metal sheet is detected in real time using a real-time noise occlusion dictionary library. Abnormal image regions are read through the data, and normal image regions are extracted from infrared radiation image data and visible light reflection image data. The coordinates of the two are compared, and the region with completely overlapping coordinate ranges is selected as the cross-modal analysis sample region. The sample region is divided into analysis units according to fixed pixel blocks. The average gray value, texture entropy value and gradient direction histogram of infrared radiation image data are extracted as infrared modal feature vectors for each analysis unit, and the average gray value, texture entropy value and gradient direction histogram of visible light reflection image data are extracted as visible light modal feature vectors. After dimensionless processing of the infrared modal feature vectors and visible light modal feature vectors of all analysis units, they are paired to form a set of feature vector pairs. The set of feature vector pairs is trained using a least squares support vector machine to obtain a bidirectional mapping function between the infrared modal feature vectors and the visible light modal feature vectors, thus forming a cross-modal feature mapping relationship.

6. A metal sheet production quality inspection system based on the Internet of Things (IoT), applied to the metal sheet production quality inspection method based on the IoT as described in any one of claims 1-5, characterized in that: The metal sheet production quality inspection system includes a multimodal image acquisition module, an abnormal region identification module, a real-time dictionary update module, a cross-modal feature analysis module, and an abnormal image repair module. The multimodal image acquisition module is used to acquire infrared thermal imaging and visible light image data of metal plates according to the detection cycle, and form a multimodal image data set after time alignment; the abnormal region identification module is used to construct an initial noise occlusion dictionary library using historical multimodal image data, and obtain abnormal image regions through current data; the real-time dictionary update module is used to construct a real-time dictionary library update pool, and obtain a real-time noise occlusion dictionary library by predicting and updating the initial noise occlusion dictionary library through access frequency trends. The cross-modal feature analysis module is used to detect abnormal image regions in real time using a real-time noise occlusion dictionary, and analyze normal regions to obtain cross-modal feature mapping relationships. The abnormal image repair module is used to repair noise occlusion in abnormal image regions according to cross-modal feature mapping relationships, and transmit the repaired image data to the detection and analysis module.

7. The IoT-based metal sheet production quality inspection system according to claim 6, characterized in that: The multimodal image acquisition module includes an image sensing acquisition unit and a spatiotemporal calibration processing unit; the image sensing acquisition unit is used to acquire infrared radiation image data and visible light reflection image data of the metal plate through an infrared thermal imaging sensor and a visible light image sensor; the spatiotemporal calibration processing unit is used to perform time alignment of the infrared and visible light image data based on timestamp data and correct coordinate system deviations. The abnormal region identification module includes an initial dictionary construction unit and an abnormal region detection unit; the initial dictionary construction unit is used to analyze the multimodal image data set of multiple historical detection periods to obtain an initial noise occlusion dictionary library; the abnormal region detection unit is used to obtain abnormal image regions from the multimodal image data set of the current detection period through feature matching and multimodal residual analysis. The real-time dictionary update module includes an access data storage unit and a dictionary sorting update unit. The access data storage unit is used to store the identifiers and corresponding access counts of each atom in the initial noise occlusion dictionary library within the current detection period. The dictionary sorting update unit is used to analyze the access counts to make trend predictions and to sort and update the atoms in the initial noise occlusion dictionary library according to the predictions.

8. The IoT-based metal sheet production quality inspection system according to claim 6, characterized in that: The cross-modal feature analysis module includes a real-time image detection unit and a feature mapping construction unit; the real-time image detection unit is used to traverse and read abnormal image regions in the multimodal image dataset using a real-time noise occlusion dictionary; The feature mapping construction unit is used to select normal image regions to perform cross-modal feature mapping relationship analysis to obtain cross-modal feature mapping relationships; The abnormal image repair module includes an image repair processing unit and a detection data transmission unit; the image repair processing unit is used to repair noise occlusion in the abnormal image region detected in real time to obtain noise-free image data; The detection data transmission unit is used to transmit noise-free image data to the metal sheet production quality detection and analysis module.

Images