A method and training system for identifying and troubleshooting production safety hazards / risks based on visual recognition

Abstract

The present invention relates to the fields of industrial automation and intelligent control, and in particular to a visual recognition-based method and training system for identifying and troubleshooting production safety hazards and risks. This method collects hyperspectral images, infrared thermal images, and vibration sensor data, performs data cleaning, feature extraction, and normalization processing to generate fused feature data. This fused feature data is monitored and compared in real time using edge computing devices and deep reinforcement learning algorithms. Sparse representation and multi-level spatiotemporal analysis are used for anomaly detection and hazard assessment, ultimately generating a comprehensive hazard assessment report. The present invention can efficiently identify and assess production hazards in real time, improving production safety and reliability.

Description

Technical Field

[0001] The present invention relates to the fields of industrial automation and intelligent control, and in particular to a method and training system for identifying and troubleshooting production safety hazards / risks based on visual recognition. Background Art

[0002] With the advancement of industrial automation, ensuring the safety and reliability of production processes has become a top priority. Using visual recognition technology to automatically identify and troubleshoot safety hazards and risks can significantly improve production efficiency, reduce human error, and enhance production safety.

[0003] The following problems exist in current production safety:

[0004] Traditional safety inspections rely on manual labor, which has problems such as slow recognition speed and limited coverage, and cannot detect potential hidden dangers in a timely manner; existing systems are inefficient when processing large-scale and complex production data, and have difficulty providing real-time safety assessments; traditional technologies often lack intelligent analysis capabilities and cannot effectively use data for in-depth hidden danger prediction and risk assessment.

[0005] These problems are mainly caused by the following reasons:

[0006] Manual inspection is time-consuming, inaccurate, and easily affected by subjective factors; traditional data processing technologies are unable to cope with the large amount of high-dimensional data generated in modern production environments; and the lack of advanced algorithms and models for intelligent analysis makes it impossible to fully tap the potential information in the data.

[0007] The existing technology (Chinese invention patent, publication number: CN116523908B, title: Safety production method, system, equipment and medium based on coil coating production line) has the following shortcomings or defects in addressing these issues:

[0008] Existing technologies rely on a single type of data (such as image data), making it difficult to integrate multi-source data for comprehensive analysis; data processing and analysis are slow and cannot meet the needs of real-time monitoring and assessment; simple algorithms are used for hidden danger assessment, which cannot deeply mine hidden danger information in the data; there is a lack of effective feedback and adjustment mechanisms, and dynamic adjustments cannot be made based on real-time data. Summary of the Invention

[0009] In response to the many problems existing in the above-mentioned existing technologies, the present invention provides a method and training system for identifying and troubleshooting safety hazards / risks in production based on visual recognition. The present invention collects and fuses multi-source data (hyperspectral images, infrared thermal images, and vibration sensor data) in real time, utilizes deep reinforcement learning algorithms for data monitoring and optimization, combines multi-level spatiotemporal analysis and causal analysis techniques to conduct hazard assessment, and ultimately generates a comprehensive hazard assessment report. The present invention can identify and assess hazards in the production process in real time and efficiently, significantly improving production safety and reliability.

[0010] A method for identifying and troubleshooting production safety hazards / risks based on visual recognition, comprising the following steps:

[0011] Real-time collection of hyperspectral image data, infrared thermal image data, and vibration sensor data on the production line, and cleaning, feature extraction, and normalization processing to generate fused feature data;

[0012] Using edge computing devices and deep reinforcement learning algorithms to monitor and compare the fused feature data in real time, including preliminary comparison, cluster analysis, and deep reinforcement learning optimization, to generate optimized comparison result data;

[0013] The optimized comparison result data is subjected to anomaly detection and hidden danger assessment through sparse representation, multi-level spatiotemporal analysis and causal analysis, a comprehensive hidden danger assessment report is generated, and the benchmark model and detection parameters are adjusted according to the feedback of the comprehensive hidden danger assessment report.

[0014] Preferably, the cleaning of the hyperspectral image data, infrared thermal image data and vibration sensor data includes: noise removal and outlier processing, wherein noise removal is achieved by a filtering algorithm, and outlier processing is achieved by detecting and eliminating abnormal data points through statistical analysis.

[0015] Preferably, the feature extraction includes: extracting edge and contour features from the cleaned hyperspectral image data, extracting temperature features from the cleaned infrared thermal image data, and extracting vibration mode features from the cleaned vibration sensing data, and generating hyperspectral feature data, infrared thermal feature data and vibration feature data respectively.

[0016] Preferably, the normalization processing includes: normalizing the hyperspectral feature data, infrared thermal feature data and vibration feature data to ensure that the data are within the same magnitude range, and the normalization processing is achieved by a linear normalization method.

[0017] Preferably, the real-time monitoring and comparison includes: performing a preliminary comparison on the fused feature data through an edge computing device to generate preliminary comparison result data, wherein the preliminary comparison is achieved by calculating the Euclidean distance between the real-time feature data and the dynamic benchmark model, and the Euclidean distance calculation formula is:

[0018] ,

[0019] in, represents the Euclidean distance; Indicates the number of feature dimensions; Indicates the first eigenvalues; represents the first eigenvalues.

[0020] Preferably, the cluster analysis includes: performing cluster analysis on the preliminary comparison result data using K-means clustering technology to generate cluster result data, wherein K-means clustering is achieved by iteratively optimizing cluster centers, and the objective function of the iterative optimization is:

[0021] ,

[0022] in, represents the objective function value; represents the number of clusters; Indicates the number of data points; Indicates that it belongs to Class data points; Indicates the The cluster center of the class.

[0023] Preferably, the deep reinforcement learning optimization includes: optimizing the clustering result data using a deep reinforcement learning algorithm to generate optimized comparison result data, wherein the deep reinforcement learning algorithm performs strategy optimization through a deep Q network, and the loss function of the deep Q network is:

[0024] ,

[0025] in, represents the loss function; Indicates the current network parameters; Indicates the current state; Indicates the current action; Indicates reward; Indicates the next state; represents the discount factor; represents the Q-value function; represents the target network parameters; Indicates the next action.

[0026] Preferably, the anomaly detection and hidden danger assessment includes: performing anomaly pattern recognition on the optimized comparison result data by a sparse representation algorithm to generate anomaly pattern data, wherein the sparse representation algorithm realizes anomaly detection by minimizing the reconstruction error, and the calculation formula of the reconstruction error is:

[0027] ,

[0028] in, represents the reconstruction error; Represents the optimized alignment result data matrix; represents the dictionary matrix; represents a sparse coefficient vector; represents the regularization parameter.

[0029] Preferably, the hidden danger assessment includes: performing time and space feature analysis on the abnormal pattern data using multi-level time and space analysis technology to generate time and space feature data, performing causal relationship analysis on the time and space feature data using causal analysis technology to generate causal relationship data, and finally generating a comprehensive hidden danger assessment report, wherein the causal analysis is implemented by a Granger causality model, and the calculation formula of the Granger causality model is:

[0030] ,

[0031] in, represents the explained variable; represents the explanatory variable; represents the lag coefficient of the explained variable; represents the lag coefficient of the explanatory variable; The explained variable hysteresis value; The explanatory variable hysteresis value; represents the error term; represents the lag order of the explained variable; represents the lag order of the explanatory variable.

[0032] A training system for the visual recognition-based production safety hazard / risk identification and troubleshooting method, comprising:

[0033] A data acquisition module is used to collect hyperspectral image data, infrared thermal image data, and vibration sensor data on the production line in real time through hyperspectral imaging sensors, infrared thermal imaging sensors, and vibration sensors;

[0034] A data preprocessing and fusion module is used to clean, extract features and normalize the hyperspectral image data, infrared thermal image data and vibration sensor data to generate fused feature data;

[0035] A real-time monitoring and comparison module is used to monitor and compare the fused feature data in real time using edge computing devices and deep reinforcement learning algorithms, including preliminary comparison, cluster analysis, and deep reinforcement learning optimization, to generate optimized comparison result data;

[0036] Anomaly detection and hidden danger assessment module, used to perform anomaly detection and hidden danger assessment on the optimized comparison result data through sparse representation, multi-level spatiotemporal analysis and causal analysis, generate a comprehensive hidden danger assessment report, and adjust the benchmark model and detection parameters based on the feedback of the comprehensive hidden danger assessment report;

[0037] A model training module is used to train and optimize the benchmark model using the fused feature data and the optimized comparison result data to generate a dynamic benchmark model;

[0038] A model validation module, configured to validate and evaluate the dynamic benchmark model using a validation dataset to ensure the accuracy and robustness of the model;

[0039] The deployment and update module is used to deploy the verified dynamic benchmark model to the production environment and dynamically update the model parameters based on real-time data feedback.

[0040] Compared with the prior art, the advantages and beneficial effects of the present invention are:

[0041] The present invention achieves efficient and comprehensive hidden danger identification by introducing multi-source data fusion technology;

[0042] The present invention achieves real-time data monitoring and rapid response effects by utilizing edge computing devices and deep reinforcement learning algorithms;

[0043] This invention achieves intelligent hidden danger identification and risk assessment through deep reinforcement learning and sparse representation algorithm;

[0044] The present invention achieves dynamic feedback and real-time adjustment effects of hidden danger assessment through multi-level spatiotemporal analysis and causal analysis technology;

[0045] The present invention improves the comprehensiveness and accuracy of hidden danger identification through multimodal data fusion;

[0046] The present invention improves the efficiency and quality of data processing through efficient data cleaning, feature extraction and normalization;

[0047] Through the scalability and robustness of the system, the present invention is suitable for identifying and troubleshooting safety hazards and risks in different industrial scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1Schematic diagram of the process of the present invention;

[0049] Figure 2 Schematic diagram of multi-source data fusion in the present invention;

[0050] Figure 3 This is a flow chart of real-time monitoring and comparison in the present invention;

[0051] Figure 4 This is a flowchart for deep reinforcement learning optimization in the present invention;

[0052] Figure 5 This is a flowchart of anomaly detection and hidden danger assessment in the present invention;

[0053] Figure 6 It is a structural block diagram of the system of the present invention. DETAILED DESCRIPTION

[0054] Hereinafter, embodiments of the present disclosure will be described with reference to the accompanying drawings. However, it should be understood that these descriptions are merely exemplary and are not intended to limit the scope of the present disclosure. In the following detailed description, for ease of explanation, many specific details are set forth to provide a comprehensive understanding of the embodiments of the present disclosure. However, it is apparent that one or more embodiments may also be implemented without these specific details. In addition, in the following description, descriptions of well-known structures and technologies are omitted to avoid unnecessary confusion of the concepts of the present disclosure.

[0055] The terms used herein are only for describing specific embodiments and are not intended to limit the present disclosure. The terms "comprise," "include," etc. used herein indicate the presence of the features, steps, operations, and / or components, but do not exclude the presence or addition of one or more other features, steps, operations, or components.

[0056] All terms used herein (including technical and scientific terms) have the meanings commonly understood by those skilled in the art unless otherwise defined. It should be noted that the terms used herein should be interpreted as having a meaning consistent with the context of this specification and should not be interpreted in an idealized or overly rigid manner.

[0057] When expressions such as “at least one of A, B, and C, etc.” are used, they should generally be interpreted in accordance with the meaning commonly understood by those skilled in the art (for example, “a system having at least one of A, B, and C” should include but is not limited to systems having A alone, B alone, C alone, A and B, A and C, B and C, and / or A, B, and C, etc.). When expressions such as “at least one of A, B, or C, etc.” are used, they should generally be interpreted in accordance with the meaning commonly understood by those skilled in the art (for example, “a system having at least one of A, B, or C” should include but is not limited to systems having A alone, B alone, C alone, A and B, A and C, B and C, and / or A, B, and C, etc.).

[0058] The accompanying drawings illustrate some block diagrams and / or flow charts. It should be understood that some blocks in the block diagrams and / or flow charts, or combinations thereof, may be implemented by computer program instructions. These computer program instructions may be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing device, so that when executed by the processor, these instructions may create a device for implementing the functions / operations described in these block diagrams and / or flow charts. The techniques of the present disclosure may be implemented in the form of hardware and / or software (including firmware, microcode, etc.). In addition, the techniques of the present disclosure may take the form of a computer program product on a computer-readable storage medium having stored thereon instructions, which may be used by or in conjunction with an instruction execution system.

[0059] like Figure 1 As shown, a method for identifying and troubleshooting production safety hazards / risks based on visual recognition includes the following steps:

[0060] like Figure 2 As shown, hyperspectral image data, infrared thermal image data and vibration sensor data on the production line are collected in real time, and then cleaned, feature extracted and normalized to generate fused feature data;

[0061] Preferably, the cleaning of the hyperspectral image data, infrared thermal image data and vibration sensor data includes: noise removal and outlier processing, wherein noise removal is achieved by a filtering algorithm, and outlier processing is achieved by detecting and eliminating abnormal data points through statistical analysis.

[0062] In the present invention, noise removal is achieved through a filtering algorithm. Filtering algorithms are a mature technical means in the field of signal processing, and generally include low-pass filtering, high-pass filtering, and band-pass filtering. The principle is to suppress or remove noise signals of specific frequencies by processing data in the frequency domain or time domain. Low-pass filters are used to remove high-frequency noise, high-pass filters are used to remove low-frequency noise, and band-pass filters are used to retain signals within a specific frequency band. After data is processed by the filtering algorithm, the noise component is significantly reduced, making subsequent feature extraction and analysis more accurate and reliable. Specific embodiments include the commonly used mean filtering and median filtering in hyperspectral image data processing. Mean filtering smoothes out random noise by averaging the image pixel values; median filtering effectively removes salt and pepper noise by selecting the median value within a window as the pixel value. In infrared thermal image data processing, Gaussian filtering can be used to remove high-frequency noise and maintain the smoothness of the temperature distribution. In vibration sensor data processing, low-pass filtering can remove high-frequency vibration signals, making the main vibration mode clearer.

[0063] The main purpose of outlier processing is to detect and remove abnormal data points introduced by sensor failures, data transmission errors, and other factors. These outliers can significantly affect data analysis results and must be addressed. This approach uses statistical analysis methods to identify values ​​that deviate from the normal range. Common methods include Z-scores, boxplots, and density-based local outlier factors (LOFs). After outliers are detected, they are processed through interpolation, deletion, or replacement to ensure data consistency and continuity. This outlier processing eliminates extreme deviations in the data, making it more representative and stable, and ensuring the reliability of subsequent analysis. Specific examples include: in hyperspectral image data processing, the Z-score method can be used to detect pixel value anomalies. Pixels outside a certain standard deviation range are marked as outliers and then replaced with the average value of the surrounding pixels. In infrared thermal image data processing, the boxplot method can be used to detect abnormally high or low temperatures in the temperature distribution, and interpolation methods can be used to address these points. In vibration sensor data processing, the density-based local outlier factor (LOF) method can effectively detect abnormal vibration signals caused by sensor failure and ensure data continuity by deleting these outliers.

[0064] By removing noise and processing outliers from hyperspectral image data, infrared thermal image data, and vibration sensor data, we ensure high data quality and consistency. This process provides a reliable foundation for subsequent data fusion, feature extraction, real-time monitoring and comparison, anomaly detection, and hazard assessment. The application of filtering algorithms and statistical analysis methods not only improves data accuracy and stability, but also effectively reduces the probability of false positives and missed negatives, making the entire production safety hazard / risk identification and investigation system more efficient and reliable.

[0065] Preferably, the feature extraction includes: extracting edge and contour features from the cleaned hyperspectral image data, extracting temperature features from the cleaned infrared thermal image data, and extracting vibration mode features from the cleaned vibration sensing data, and generating hyperspectral feature data, infrared thermal feature data and vibration feature data respectively.

[0066] The principle of extracting edge and contour features from cleaned hyperspectral image data is to use image processing algorithms, such as Canny edge detection and the Sobel operator, to extract edges by detecting areas in the image where pixel values ​​change significantly. These algorithms identify edges and contours in the image by calculating the gradient and second-order derivative of the image. These edge and contour features can reveal the geometric shape and structural changes of the workpiece and are an important basis for identifying potential hidden dangers. The data after edge and contour feature extraction is called hyperspectral feature data, which has the effect of accurately reflecting the surface characteristics of the workpiece and helping to identify possible physical defects or damage. For example, in hyperspectral image processing, the use of the Canny edge detection algorithm can effectively extract cracks and scratches on the surface of the workpiece. Through these feature data, the length and direction of the cracks can be further analyzed, providing reliable data support for hidden danger assessment.

[0067] The principle of extracting temperature features from infrared thermal image data after cleaning is to use the pixel values ​​in the image to represent temperature information. By extracting these pixel values, the temperature distribution of the workpiece can be obtained. The temperature features in infrared thermal images can reveal thermal anomalies generated by the workpiece during operation. These thermal anomalies may be caused by equipment overheating, friction or other failures. The data after temperature feature extraction is called infrared thermal feature data, which has the effect of accurately capturing the temperature changes of the workpiece and helping to discover potential thermal hazards. For example, in infrared thermal image processing, the temperature threshold segmentation method can be used to extract high-temperature areas, identify overheated parts of the workpiece, and further analyze the cause and impact range of overheating through temperature feature data, providing key data for preventing equipment failure.

[0068] The principle of extracting vibration pattern features from vibration sensor data after cleaning is to use signal processing algorithms, such as Fourier transform and wavelet transform, to convert the vibration signal in the time domain into frequency domain features, and identify the main frequency components and energy distribution in the vibration signal. Vibration pattern features can reflect the vibration state of the workpiece during operation. These vibration states can reveal mechanical failures, looseness, or other abnormal conditions of the equipment. The data after vibration pattern feature extraction is called vibration feature data. Its effect is to accurately reflect the vibration characteristics of the workpiece and help identify potential mechanical hazards. For example, in vibration sensor data processing, the main vibration frequency components are extracted through Fourier transform, which can determine whether the workpiece has resonance. Through this feature data, the cause of the resonance and possible solutions can be further analyzed, providing an important basis for mechanical maintenance.

[0069] By extracting edge and contour features, temperature features, and vibration mode features from hyperspectral image data, infrared thermal image data, and vibration sensor data, hyperspectral feature data, infrared thermal feature data, and vibration feature data are generated respectively. These feature data provide rich basic information for subsequent real-time monitoring and comparison, anomaly detection, and hidden danger assessment, ensuring that the system can accurately identify and assess safety hazards in production.

[0070] Preferably, the normalization processing includes: normalizing the hyperspectral feature data, infrared thermal feature data and vibration feature data to ensure that the data are within the same magnitude range, and the normalization processing is achieved by a linear normalization method.

[0071] Normalization is achieved through linear normalization, which works by scaling the data to a predefined range, usually between 0 and 1.

[0072] The calculation formula for linear normalization is:

[0073] ,

[0074] in, represents the normalized value, Represents the original value, and Represent the minimum and maximum values ​​of the original data, respectively. Through linear normalization, all feature data are scaled to the same magnitude range, which allows different types of feature data to be compared and analyzed on the same scale, avoiding deviations and errors caused by differences in data ranges.

[0075] In the present invention, the specific application of normalization processing to hyperspectral feature data, infrared thermal feature data and vibration feature data is as follows:

[0076] For hyperspectral feature data, normalization ensures that edge and contour features extracted from different spectral channels can be compared and analyzed at the same level. The feature values ​​of each channel of a hyperspectral image may vary greatly. Through normalization, these differences can be reduced, making subsequent feature fusion and analysis more accurate. For example, in hyperspectral image processing, the spectral reflectance values ​​of some channels may be between 0 and 100, while the values ​​of other channels may be between 0 and 1. Through normalization, the values ​​of all channels are scaled to the range of 0 to 1, so that the feature values ​​of different channels can be fused and analyzed on the same scale, thereby improving the accuracy of the overall analysis.

[0077] For infrared thermal feature data, normalization ensures that data within different temperature ranges can be compared and analyzed on the same scale. The temperature data in infrared thermal images may have large differences in numerical range. Through normalization, the temperature data can be scaled to the range of 0 to 1, allowing the temperature feature data to be fused and analyzed on the same scale as other types of feature data. For example, in infrared thermal image processing, the temperature in different areas may vary between 20°C and 100°C. Through normalization, the temperature values ​​are scaled to the range of 0 to 1, allowing data from different temperature areas to be compared on the same scale, making it easier to identify temperature anomalies and hidden dangers.

[0078] For vibration feature data, normalization ensures that data within different frequency and amplitude ranges can be compared and analyzed on the same scale. The frequency and amplitude of vibration sensing data may vary significantly. Normalization can reduce these differences, allowing vibration pattern features to be integrated and analyzed on the same scale as other types of feature data. For example, in vibration sensing data processing, the vibration frequency and amplitude of different mechanical components may vary significantly. Normalization can be used to scale the frequency and amplitude values ​​to the range of 0 to 1, allowing vibration feature data from different mechanical components to be compared on the same scale, making it easier to identify mechanical failures and hidden dangers.

[0079] Through normalization, hyperspectral, infrared, and vibration data are compared and analyzed within the same magnitude range, ensuring the accuracy and consistency of data fusion and analysis. This process provides a reliable foundation for subsequent real-time monitoring and comparison, anomaly detection, and hazard assessment, avoiding analytical bias caused by differences in data magnitude and thus improving the overall accuracy and reliability of the system. The application of normalization not only enhances the comparability of different feature data but also effectively improves data processing efficiency, making the entire system more efficient and accurate in processing and analyzing multimodal data.

[0080] Using edge computing devices and deep reinforcement learning algorithms to monitor and compare the fused feature data in real time, including preliminary comparison, cluster analysis, and deep reinforcement learning optimization, to generate optimized comparison result data;

[0081] like Figure 3 As shown, preferably, the real-time monitoring and comparison includes: performing a preliminary comparison on the fused feature data through an edge computing device to generate preliminary comparison result data, wherein the preliminary comparison is achieved by calculating the Euclidean distance between the real-time feature data and the dynamic reference model, and the Euclidean distance calculation formula is:

[0082] ,

[0083] in, represents the Euclidean distance; Indicates the number of feature dimensions; Indicates the first eigenvalues; represents the first eigenvalues.

[0084] Euclidean distance is a common method for measuring the straight-line distance between two points. In this paper, it is used to compare the difference between real-time feature data and a dynamic baseline model. The dynamic baseline model represents the feature data under normal production conditions. By comparing real-time feature data with the baseline model, anomalies can be effectively identified.

[0085] In practice, the real-time monitoring and comparison process relies on edge computing devices processing collected fused feature data in real time and calculating the Euclidean distance between this data and a pre-built dynamic baseline model. A larger Euclidean distance indicates a significant difference between the real-time data and the baseline model, potentially indicating an anomaly; a smaller Euclidean distance indicates a similarity between the real-time data and the baseline model, indicating normal production.

[0086] Effectively, by calculating Euclidean distance in real time, the system can quickly identify data points that deviate significantly from the baseline model, alerting potential safety hazards. For example, in hyperspectral image data processing, if the reflectance of certain spectral channels in the real-time feature data deviates significantly from the baseline model, it may indicate the presence of new scratches or cracks on the workpiece surface. In infrared thermal image data processing, if the real-time temperature feature data differs significantly from the baseline model, it may indicate localized overheating of the equipment. In vibration sensor data processing, if the real-time vibration feature data differs significantly from the baseline model, it may indicate loose or faulty mechanical components.

[0087] Specific embodiments include:

[0088] Hyperspectral image data processing: In real-time hyperspectral image data, the reflectance of each spectral channel is calculated to compare with the baseline model to identify possible physical defects on the workpiece surface. For example, if the reflectance value of a channel is significantly higher than the baseline model, it indicates the presence of scratches on the surface.

[0089] Infrared thermal image data processing: This system uses real-time infrared thermal image data to identify thermal anomalies in equipment by calculating the difference between temperature signature data and a baseline model. For example, if the temperature in a certain area is significantly higher than the baseline model, this indicates possible overheating.

[0090] Vibration sensor data processing: This system uses real-time vibration sensor data to identify vibration anomalies in mechanical components by calculating the difference between the vibration signature data and a baseline model. For example, a vibration amplitude at a certain frequency that is significantly higher than the baseline model may indicate mechanical looseness.

[0091] Through the method of the present invention, the real-time monitoring and comparison steps play a key role in the production safety hazard / risk identification and investigation system based on visual recognition, which can timely and accurately identify various hidden dangers that may exist in the production process, thereby improving the safety and reliability of the production process.

[0092] Preferably, the cluster analysis includes: performing cluster analysis on the preliminary comparison result data using K-means clustering technology to generate cluster result data, wherein K-means clustering is achieved by iteratively optimizing cluster centers, and the objective function of the iterative optimization is:

[0093] ,

[0094] in, represents the objective function value; represents the number of clusters; Indicates the number of data points; Indicates that it belongs to Class data points; Indicates the The cluster center of the class.

[0095] The cluster analysis step involves clustering the preliminary alignment data using K-means clustering to generate clustered data. K-means clustering is achieved by iteratively optimizing cluster centers. The K-means clustering algorithm is a commonly used unsupervised learning method that is primarily used to assign data points into k clusters, ensuring that data points within the same cluster are as similar as possible and data points in different clusters are as dissimilar as possible.

[0096] K-means clustering works by iteratively optimizing cluster centers to minimize the distance between each data point and its cluster center. The process involves initializing k cluster centers, assigning each data point to the closest cluster center, and then updating the cluster center to the average value of all data points in that cluster. This process is repeated until the cluster center no longer changes significantly.

[0097] In the present invention, the application effect of the cluster analysis step is reflected in the following aspects:

[0098] First, cluster analysis can effectively classify the preliminary comparison result data to help identify different types of potential hidden dangers. For example, in hyperspectral image data processing, cluster analysis can divide workpieces with different surface features into multiple clusters, thereby identifying workpieces with cracks, scratches, or other physical defects. In infrared thermal image data processing, cluster analysis can divide areas with different temperature characteristics into multiple clusters, thereby identifying areas that may have overheating or abnormal temperature distribution. In vibration sensing data processing, cluster analysis can divide mechanical components with different vibration modes into multiple clusters, thereby identifying components that may have mechanical failures or abnormal vibrations.

[0099] Secondly, the iterative optimization process of K-means clustering ensures the stability and accuracy of the clustering results. By optimizing the cluster centers, data points within the same cluster are made more similar, and the differences between different clusters are more significant, thereby improving the effectiveness of cluster analysis. For example, during the cluster analysis process, iterative optimization can ensure that the edge and contour features of all hyperspectral image data points belonging to a cluster are similar, the temperature features of all infrared thermal image data points belonging to a cluster are similar, and the vibration mode features of all vibration sensor data points belonging to a cluster are similar, thereby improving the accuracy of hidden danger identification.

[0100] Specific embodiments include:

[0101] Hyperspectral image data processing: Using the K-means clustering algorithm, we group the edge and contour feature data points on the workpiece surface into multiple clusters, thereby identifying workpieces with different surface conditions. For example, clustering the hyperspectral data points indicating cracks on the workpiece surface allows for rapid identification of cracked workpieces.

[0102] Infrared thermal image data processing: Using the K-means clustering algorithm, data points with different temperature characteristics of the equipment are divided into multiple clusters, thereby identifying areas with different temperature distributions. For example, clustering data points in overheated areas of the equipment can quickly identify components that may have overheating problems.

[0103] Vibration sensor data processing: The K-means clustering algorithm is used to group data points representing different vibration modes of mechanical components into multiple clusters, thereby identifying components with different vibration states. For example, clustering data points indicating mechanical resonance can quickly identify components with resonance issues.

[0104] Through the above method, the cluster analysis step plays an important role in the safety production hazard / risk identification and investigation system based on visual recognition. It can effectively classify and identify different types of hazards and improve the detection accuracy and reliability of the system.

[0105] like Figure 4 As shown, preferably, the deep reinforcement learning optimization includes: using a deep reinforcement learning algorithm to optimize the clustering result data to generate optimized comparison result data, wherein the deep reinforcement learning algorithm performs strategy optimization through a deep Q network, and the loss function of the deep Q network is:

[0106] ,

[0107] in, represents the loss function; Indicates the current network parameters; Indicates the current state; Indicates the current action; Indicates reward; Indicates the next state; represents the discount factor; represents the Q-value function; represents the target network parameters; Indicates the next action.

[0108] The present invention uses a deep reinforcement learning algorithm to optimize the clustering result data and generate optimized comparison result data. Specifically, the deep reinforcement learning algorithm uses a deep Q-network to perform strategy optimization and uses the loss function of the deep Q-network to optimize the model's predictive performance.

[0109] The Deep Q-Network (DQN) is a commonly used deep reinforcement learning algorithm that combines Q-learning and deep neural networks to handle high-dimensional state spaces. Its core concept is to approximate the Q-value function using a neural network to find the optimal policy. In this paper, the Deep Q-Network's loss function is used to evaluate the model's prediction error and to perform parameter updates via the backpropagation algorithm.

[0110] The optimized comparison result data is subjected to anomaly detection and hidden danger assessment through sparse representation, multi-level spatiotemporal analysis and causal analysis, a comprehensive hidden danger assessment report is generated, and the benchmark model and detection parameters are adjusted according to the feedback of the comprehensive hidden danger assessment report.

[0111] In this invention, the principle of deep reinforcement learning optimization step is that through the deep Q network, the system can continuously learn and improve its strategy in the process of real-time monitoring and comparison. The specific process includes: the system collects the current state at each time step , select an action , get instant rewards based on the actions you choose and the next state The system calculates Q-values ​​through a deep Q-network and selects the optimal action by maximizing future rewards. During training, the system continuously optimizes the Q-value function, enabling it to accurately predict the value of different states and actions, thereby improving the system's decision-making ability.

[0112] Effectively, through deep reinforcement learning optimization, the system is able to more effectively identify and assess hidden dangers in complex production environments. Specifically, the system is able to dynamically adjust its strategies to respond to different types and levels of risks, improving the accuracy and timeliness of hidden danger identification. For example, in hyperspectral image data processing, through deep reinforcement learning, the system is able to more accurately identify tiny cracks on the surface of the workpiece and adjust the detection strategy according to the development trend of the cracks; in infrared thermal image data processing, the system is able to optimize its detection model based on real-time temperature changes and more quickly identify overheating problems of equipment; in vibration sensor data processing, the system is able to optimize its detection strategy based on changes in vibration patterns and more accurately identify loose or faulty mechanical components.

[0113] Specific embodiments include:

[0114] Hyperspectral image data processing: By optimizing the crack detection model through a deep Q-network, the system can adjust detection strategies based on real-time hyperspectral data, improving crack identification accuracy and response speed. For example, when the system identifies a new crack on a workpiece surface, it can dynamically adjust its detection parameters to ensure continuous crack monitoring and assessment.

[0115] Infrared thermal image data processing: By optimizing the temperature detection model through a deep Q-network, the system can optimize detection strategies based on the device's real-time temperature changes, quickly identifying overheating issues and taking appropriate measures. For example, if the system detects a rapid increase in temperature in a certain part of the device, it can promptly issue an alarm and adjust the detection frequency to ensure that the problem is promptly addressed.

[0116] Vibration sensor data processing: By optimizing the vibration detection model through a deep Q-network, the system can optimize detection strategies based on the real-time vibration patterns of mechanical components, accurately identifying potential faults and hidden dangers. For example, if the system identifies an abnormal change in the vibration frequency of a mechanical component, it can adjust its detection parameters and strategies to ensure that the fault is discovered and addressed promptly.

[0117] Through the above method, the deep reinforcement learning optimization step plays an important role in the visual recognition-based production safety hazard / risk identification and investigation system. It can dynamically adjust and optimize the detection strategy, improve the system's response speed and accuracy, and thus more effectively ensure the safety and reliability of the production process.

[0118] like Figure 5 As shown, preferably, the anomaly detection and hidden danger assessment includes: performing anomaly pattern recognition on the optimized comparison result data through a sparse representation algorithm to generate anomaly pattern data, wherein the sparse representation algorithm realizes anomaly detection by minimizing the reconstruction error, and the calculation formula of the reconstruction error is:

[0119] ,

[0120] in, represents the reconstruction error; Represents the optimized alignment result data matrix; represents the dictionary matrix; represents a sparse coefficient vector; represents the regularization parameter.

[0121] The anomaly detection and hidden danger assessment step is a key step, used to identify and evaluate abnormal situations and potential hidden dangers that occur during the production process. This step involves using a sparse representation algorithm to identify abnormal patterns in the optimized comparison data and generate abnormal pattern data. The sparse representation algorithm achieves anomaly detection by minimizing reconstruction error.

[0122] The principle of the sparse representation algorithm is to represent the data as a linear combination of dictionary matrices, where the sparse coefficient vector represents the weight of the linear combination. By minimizing the reconstruction error, the best approximation of the data can be achieved, thereby effectively identifying abnormal patterns in the data. Specifically, the sparse representation algorithm finds the sparsest coefficient vector so that the original data can be reconstructed as accurately as possible through the linear combination of the dictionary matrix, thereby achieving anomaly detection. Regularization parameter It is used to balance reconstruction error and sparsity, so that the model can accurately reconstruct the data while maintaining sparsity.

[0123] In the present invention, the application effects of the anomaly detection and hidden danger assessment steps are reflected in the following aspects:

[0124] First, sparse representation algorithms can effectively extract key features from large amounts of data and identify potential abnormal patterns. For example, in hyperspectral image data processing, sparse representation algorithms can identify spectral anomalies caused by surface defects. These abnormal pattern data can be used to further analyze the type and severity of the defects. In infrared thermal image data processing, sparse representation algorithms can identify temperature anomalies caused by equipment overheating. These abnormal pattern data can be used to further analyze the cause and scope of the overheating. In vibration sensor data processing, sparse representation algorithms can identify vibration anomalies caused by mechanical failures. These abnormal pattern data can be used to further analyze the location and cause of the failure.

[0125] Secondly, sparse representation algorithms offer significant advantages when processing high-dimensional data, ensuring computational efficiency while providing highly accurate anomaly detection results. By minimizing reconstruction errors, sparse representation algorithms can accurately locate abnormal patterns in high-dimensional data spaces, improving the accuracy and reliability of anomaly detection.

[0126] Specific embodiments include:

[0127] Hyperspectral image data processing: In hyperspectral image data processing, sparse representation algorithms generate hyperspectral anomaly pattern data by identifying anomalies in spectral features. For example, if the reflectance of certain spectral channels deviates significantly from the normal value, sparse representation algorithms can detect these anomalies through dictionary matrix reconstruction and mark them as potential surface defects.

[0128] Infrared thermal image data processing: In infrared thermal image data processing, sparse representation algorithms identify anomalies in temperature distribution and generate infrared thermal anomaly pattern data. For example, if the temperature in a certain area is significantly higher than the normal operating range, the sparse representation algorithm can detect these anomalies through dictionary matrix reconstruction and flag them as potential equipment overheating issues.

[0129] Vibration sensor data processing: In vibration sensor data processing, sparse representation algorithms identify anomalies in vibration characteristics and generate vibration anomaly pattern data. For example, if the vibration frequency and amplitude of a mechanical component deviate significantly from normal operation, the sparse representation algorithm can detect these anomalies through dictionary matrix reconstruction and flag them as potential mechanical failures.

[0130] Through the above method, the anomaly detection and hidden danger assessment steps play an important role in the production safety hidden danger / risk identification and investigation system based on visual recognition. It can efficiently and accurately identify and evaluate various abnormal situations and potential hidden dangers in the production process, thereby improving the safety and reliability of the system.

[0131] Preferably, the hidden danger assessment includes: performing time and space feature analysis on the abnormal pattern data using multi-level time and space analysis technology to generate time and space feature data, performing causal relationship analysis on the time and space feature data using causal analysis technology to generate causal relationship data, and finally generating a comprehensive hidden danger assessment report, wherein the causal analysis is implemented by a Granger causality model, and the calculation formula of the Granger causality model is:

[0132] ,

[0133] in, represents the explained variable; represents the explanatory variable; represents the lag coefficient of the explained variable; represents the lag coefficient of the explanatory variable; The explained variable hysteresis value; The explanatory variable hysteresis value; represents the error term; represents the lag order of the explained variable; represents the lag order of the explanatory variable.

[0134] The hazard assessment step is a crucial step in ensuring production process safety. This step involves analyzing the temporal and spatial characteristics of abnormal pattern data using multi-level spatiotemporal analysis techniques to generate spatiotemporal feature data. Causal analysis techniques are then used to analyze the causal relationships within this spatiotemporal feature data, ultimately generating a comprehensive hazard assessment report.

[0135] Multi-level spatiotemporal analysis technology identifies potential risk patterns by comprehensively analyzing abnormal pattern data across both time and space. Temporal feature analysis focuses on the temporal trends and periodicity of abnormal patterns, while spatial feature analysis focuses on the spatial distribution and aggregation of abnormal patterns. By combining temporal and spatial feature analysis, the system can comprehensively understand the occurrence, development, and propagation of abnormal patterns.

[0136] Causal analysis techniques further analyze causal relationships in spatiotemporal data using Granger causality models. Granger causality models are used to determine whether one time series can be used to predict another. The core concept is that if the past values ​​of one variable significantly improve the prediction accuracy of the future values ​​of another variable, the former is considered to be a Granger cause of the latter. Specifically, by performing a lagged analysis on the explained and explanatory variables, the causal relationship between the two is identified, providing a deeper understanding of the causes of abnormal patterns.

[0137] In the present invention, the application effect of the hidden danger assessment step is reflected in the following aspects:

[0138] First, through multi-level spatiotemporal analysis technology, the system can identify the temporal and spatial characteristics of abnormal patterns, improving the accuracy of hidden danger assessments. For example, in hyperspectral image data processing, temporal feature analysis can identify the development trend of surface defects, while spatial feature analysis can identify the distribution of surface defects. In infrared thermal image data processing, temporal feature analysis can identify the periodic changes in equipment overheating, while spatial feature analysis can identify the distribution of overheating areas. In vibration sensor data processing, temporal feature analysis can identify changes in the vibration patterns of mechanical failures, while spatial feature analysis can identify the locations of vibration anomalies.

[0139] Secondly, through causal analysis technology, the system can determine the causal relationship between abnormal patterns and potential hidden dangers, improving the depth and breadth of hidden danger assessment. For example, through the Granger causality model, the system can identify the causal relationship between surface defects and production process parameters in hyperspectral image data; the causal relationship between equipment overheating and ambient temperature changes in infrared thermal image data; and the causal relationship between mechanical failures and operating loads in vibration sensor data.

[0140] Specific embodiments include:

[0141] Hyperspectral image data processing: Through multi-level spatiotemporal analysis, the temporal variation trend and spatial distribution of cracks on the workpiece surface can be identified; through the Granger causality model, the causal relationship between cracks and specific production process parameters can be identified, thereby providing the causes of cracks and possible solutions.

[0142] Infrared thermal image data processing: Through multi-level spatiotemporal analysis, the temporal variation patterns and spatial distribution characteristics of equipment overheating are identified; through the Granger causality model, the causal relationship between overheating and ambient temperature fluctuations is identified, thereby providing the causes and preventive measures for equipment overheating.

[0143] Vibration sensor data processing: Through multi-level spatiotemporal analysis, the temporal variation pattern and spatial location of abnormal vibration of mechanical components are identified. Through the Granger causality model, the causal relationship between vibration anomalies and operating load changes is identified, thereby providing cause analysis of mechanical failures and maintenance recommendations.

[0144] Through the above method, the hidden danger assessment step plays an important role in the production safety hidden danger / risk identification and investigation system based on visual recognition. It can comprehensively and accurately assess various hidden dangers in the production process, provide in-depth cause-effect analysis and detailed hidden danger assessment reports, thereby effectively ensuring the safety and reliability of the production process.

[0145] like Figure 6 As shown, a training system for the method for identifying and troubleshooting production safety hazards / risks based on visual recognition includes:

[0146] The data acquisition module is used to collect hyperspectral image data, infrared thermal image data, and vibration sensing data from the production line in real time using hyperspectral imaging sensors, infrared thermal imaging sensors, and vibration sensors. Hyperspectral imaging sensors can capture surface details of workpieces on the production line and identify surface defects such as cracks and scratches by analyzing the reflectivity of different spectral channels. Infrared thermal imaging sensors capture the temperature distribution of workpieces and equipment, helping to identify overheating areas and potential thermal hazards. Vibration sensors monitor the vibration of mechanical equipment and, by analyzing vibration patterns, can detect mechanical failures at an early stage. Through data collection from these three sensors, the system can obtain comprehensive information about the production line, laying the foundation for subsequent data processing and analysis.

[0147] The data preprocessing and fusion module is used to clean, extract features and normalize the hyperspectral image data, infrared thermal image data and vibration sensor data to generate fused feature data. The data preprocessing and fusion module is used to clean, extract features and normalize the hyperspectral image data, infrared thermal image data and vibration sensor data to generate fused feature data. The data cleaning step includes noise removal and outlier processing to ensure data quality. The feature extraction step extracts edge and contour features from the hyperspectral image, temperature features from the infrared thermal image, and vibration mode features from the vibration sensor data. The normalization process standardizes different data types through linear normalization methods to make them within the same order of magnitude, which is convenient for subsequent analysis. The fused feature data integrates the above feature data into a comprehensive data set through multi-source data fusion technology to improve the comprehensiveness and accuracy of the data.

[0148] The real-time monitoring and comparison module is used to monitor and compare the fused feature data in real time using edge computing devices and deep reinforcement learning algorithms, including preliminary comparison, cluster analysis, and deep reinforcement learning optimization, to generate optimized comparison result data. The real-time monitoring and comparison module is used to monitor and compare the fused feature data in real time using edge computing devices and deep reinforcement learning algorithms, including preliminary comparison, cluster analysis, and deep reinforcement learning optimization, to generate optimized comparison result data. The preliminary comparison determines whether the current production status is normal by calculating the Euclidean distance between the real-time feature data and the dynamic benchmark model. Cluster analysis uses the K-means algorithm to classify the preliminary comparison results and identify different types of hidden dangers. Deep reinforcement learning optimization optimizes the comparison strategy through a deep Q network to improve the accuracy and timeliness of hidden danger identification. This module can monitor data changes during the production process in real time, quickly identify and classify potential hidden dangers, and ensure production safety.

[0149] The anomaly detection and hidden danger assessment module uses sparse representation, multi-level spatiotemporal analysis, and causal analysis to perform anomaly detection and hidden danger assessment on the optimized comparison result data, generate a comprehensive hidden danger assessment report, and adjust the baseline model and detection parameters based on feedback from the comprehensive hidden danger assessment report. The sparse representation algorithm identifies anomaly patterns by minimizing reconstruction error, while the multi-level spatiotemporal analysis combines temporal and spatial characteristics to analyze the changing trends and distribution patterns of anomaly data. Causal analysis uses the Granger causality model to identify the causal relationships of potential hidden dangers. Through these technologies, the system can comprehensively assess hidden dangers in the production process, generate detailed assessment reports, and adjust the detection model based on real-time feedback, improving the system's adaptability and accuracy.

[0150] The model training module is used to train and optimize the baseline model using the fused feature data and optimized comparison result data to generate a dynamic baseline model. This module uses a deep learning algorithm to train data, continuously updating and optimizing the baseline model to accurately reflect the normal state of the production process. This continuous model training enables the system to adapt to changes in different production conditions and improve the robustness and reliability of hazard identification.

[0151] The model validation module verifies and evaluates the dynamic baseline model using a validation dataset to ensure its accuracy and robustness. This module tests the dynamic baseline model using the validation dataset, evaluating its predictive performance and accuracy to ensure its reliability in practical applications. Through model validation, the system continuously improves and optimizes the baseline model, enhancing the accuracy of hazard identification.

[0152] The Deployment and Update Module deploys validated dynamic benchmark models into production environments and dynamically updates model parameters based on real-time data feedback. This module applies validated benchmark models to actual production processes, dynamically adjusting them based on real-time monitoring data to ensure the model remains optimal. Real-time updates enable the system to rapidly respond to changes in production, improving the real-time and accuracy of hazard identification.

[0153] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware. Furthermore, the present application may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0154] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems) and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0155] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0156] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0157] In a typical configuration, a computing device includes one or more processors (CPUs), input / output interfaces, network interfaces, and memory.

[0158] Memory may include non-permanent storage in a computer-readable medium, random access memory (RAM) and / or non-volatile memory in the form of read-only memory (ROM) or flash RAM. Memory is an example of a computer-readable medium.

[0159] Computer-readable media includes both permanent and non-permanent, removable and non-removable media that can be implemented using any method or technology for information storage. Information can be computer-readable instructions, data structures, program modules, or other data. Examples of computer storage media include, but are not limited to, phase-change RAM (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, compact disc read-only memory (CD-ROM), digital versatile disc (DVD) or other optical storage, magnetic cassettes, magnetic disk storage or other magnetic storage devices, or any other non-transmission media that can be used to store information that can be accessed by a computing device. As defined herein, computer-readable media does not include transitory computer-readable media, such as modulated data signals and carrier waves.

[0160] It should also be noted that the terms "comprises," "includes," or any other variations thereof are intended to encompass non-exclusive inclusion, such that a process, method, commodity, or apparatus that includes a series of elements includes not only those elements but also other elements not explicitly listed, or includes elements inherent to such process, method, commodity, or apparatus. In the absence of further limitations, an element defined by the phrase "comprises a ..." does not exclude the presence of other identical elements in the process, method, commodity, or apparatus that includes the element.

[0161] The above are merely embodiments of the present application and are not intended to limit the present application. For those skilled in the art, the present application may have various modifications and variations. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present application should be included within the scope of the claims of the present application.

Claims

1. A method for identifying and troubleshooting production safety hazards / risks based on visual recognition, characterized in that: The following steps are involved: Real-time collection of hyperspectral image data, infrared thermal image data, and vibration sensor data on the production line, and cleaning, feature extraction, and normalization processing to generate fused feature data; Using edge computing devices and deep reinforcement learning algorithms to monitor and compare the fused feature data in real time, including preliminary comparison, cluster analysis, and deep reinforcement learning optimization, to generate optimized comparison result data; The real-time monitoring and comparison includes: performing a preliminary comparison on the fused feature data through an edge computing device to generate preliminary comparison result data, wherein the preliminary comparison is achieved by calculating the Euclidean distance between the real-time feature data and a dynamic benchmark model. The dynamic benchmark model represents the feature data under normal production conditions. By comparing the real-time feature data with the benchmark model, abnormal conditions are identified. The Euclidean distance calculation formula is: Where D represents the Euclidean distance; n represents the number of feature dimensions; x i Represents the i-th eigenvalue in the real-time feature data; y i represents the i-th eigenvalue in the dynamic benchmark model; The optimized comparison result data is subjected to anomaly detection and hidden danger assessment through sparse representation, multi-level spatiotemporal analysis and causal analysis to generate a comprehensive hidden danger assessment report. The dynamic benchmark model and the detection parameters for collecting hyperspectral image data, infrared thermal image data and vibration sensor data are adjusted based on the feedback from the comprehensive hidden danger assessment report.

2. The method according to claim 1, characterized in that The cleaning of the hyperspectral image data, infrared thermal image data and vibration sensor data includes: noise removal and outlier processing, wherein the noise removal is achieved by a filtering algorithm, and the outlier processing is achieved by detecting and eliminating abnormal data points through statistical analysis.

3. The method according to claim 2, characterized in that The feature extraction includes: extracting edge and contour features from the cleaned hyperspectral image data, extracting temperature features from the cleaned infrared thermal image data, and extracting vibration mode features from the cleaned vibration sensing data, thereby generating hyperspectral feature data, infrared thermal feature data, and vibration feature data, respectively.

4. The method according to claim 3, characterized in that The normalization processing includes: performing normalization processing on the hyperspectral feature data, infrared thermal feature data and vibration feature data to ensure that the data are within the same magnitude range, and the normalization processing is achieved through a linear normalization method.

5. The method according to claim 1, characterized in that The cluster analysis includes: performing cluster analysis on the preliminary comparison result data using K-means clustering technology to generate cluster result data, wherein K-means clustering is achieved by iteratively optimizing cluster centers, and the objective function of the iterative optimization is: Where J represents the objective function value; k represents the number of clusters; n represents the number of data points; represents the i-th data point belonging to the j-th class; μ j represents the cluster center of the jth class.

6. The method according to claim 5, characterized in that The deep reinforcement learning optimization includes: optimizing the clustering result data using a deep reinforcement learning algorithm to generate optimized comparison result data, wherein the deep reinforcement learning algorithm performs strategy optimization through a deep Q network, and the loss function of the deep Q network is: Where L(θ) represents the loss function; θ represents the current network parameters; S represents the current state; a represents the current action; r represents the reward; s' represents the next state; γ represents the discount factor; Q represents the Q value function; θ - represents the target network parameters; a' represents the next action.

7. The method according to claim 1, characterized in that The anomaly detection and hidden danger assessment includes: performing anomaly pattern recognition on the optimized comparison result data using a sparse representation algorithm to generate anomaly pattern data, wherein the sparse representation algorithm realizes anomaly detection by minimizing reconstruction error, and the calculation formula of the reconstruction error is: Where E represents the reconstruction error; X represents the optimized comparison result data matrix; D represents the dictionary matrix; α represents the sparse coefficient vector; and λ represents the regularization parameter.

8. The method according to claim 7, characterized in that The hidden danger assessment includes: using multi-level spatiotemporal analysis technology to perform temporal and spatial feature analysis on the abnormal pattern data to generate spatiotemporal feature data, and using causal analysis technology to perform causal relationship analysis on the spatiotemporal feature data to generate causal relationship data, and finally generating a comprehensive hidden danger assessment report, wherein the causal analysis is implemented by the Granger causality model, and the calculation formula of the Granger causality model is: Among them, α i represents the lag coefficient of the explained variable; β j represents the lag coefficient of the explanatory variable; Y t-i represents the ith lag value of the explained variable; X t-j represents the jth lag value of the explanatory variable; ∈ t represents the error term; p represents the lag order of the explained variable; q represents the lag order of the explanatory variable.

9. A training system for the method for identifying and troubleshooting production safety hazards / risks based on visual recognition according to any one of claims 1 to 8, characterized in that: include: A data acquisition module is used to collect hyperspectral image data, infrared thermal image data, and vibration sensor data on the production line in real time through hyperspectral imaging sensors, infrared thermal imaging sensors, and vibration sensors; A data preprocessing and fusion module is used to clean, extract features and normalize the hyperspectral image data, infrared thermal image data and vibration sensor data to generate fused feature data; A real-time monitoring and comparison module is used to monitor and compare the fused feature data in real time using edge computing devices and deep reinforcement learning algorithms, including preliminary comparison, cluster analysis, and deep reinforcement learning optimization, to generate optimized comparison result data; The real-time monitoring and comparison includes: performing a preliminary comparison on the fused feature data through an edge computing device to generate preliminary comparison result data, wherein the preliminary comparison is achieved by calculating the Euclidean distance between the real-time feature data and a dynamic benchmark model. The dynamic benchmark model represents the feature data under normal production conditions. By comparing the real-time feature data with the benchmark model, abnormal conditions are identified. The Euclidean distance calculation formula is: Where D represents the Euclidean distance; n represents the number of feature dimensions; x i Represents the i-th eigenvalue in the real-time feature data; y i represents the i-th eigenvalue in the dynamic benchmark model; An anomaly detection and hidden danger assessment module, which is used to perform anomaly detection and hidden danger assessment on the optimized comparison result data through sparse representation, multi-level spatiotemporal analysis, and causal analysis, generate a comprehensive hidden danger assessment report, and adjust the detection parameters of the dynamic benchmark model and hyperspectral image data, infrared thermal image data, and vibration sensor data based on the feedback from the comprehensive hidden danger assessment report; A model training module is used to train and optimize the benchmark model using the fused feature data and the optimized comparison result data to generate a dynamic benchmark model; A model validation module, configured to validate and evaluate the dynamic benchmark model using a validation dataset to ensure the accuracy and robustness of the model; The deployment and update module is used to deploy the verified dynamic benchmark model to the production environment and dynamically update the model parameters based on real-time data feedback.

Images