Industrial safety early warning method based on image semantic large model

Abstract

The invention relates to the technical field of industrial park safety risk management. The invention provides an industrial safety early warning method based on an image semantic large model, which comprises the following steps: acquiring historical image data of an industrial site, and identifying a key change image containing potential risks; performing positive sample labeling on images with potential risks in the key change images, and performing negative sample labeling on images without potential risks in the key change images to obtain a labeled data set; constructing an industrial safety assessment initial model based on the image semantic large model, and training the industrial safety assessment model through the annotation data set to obtain an industrial safety assessment model; and inputting the real-time image data into the industrial safety evaluation model to obtain an evaluation prediction result, and performing corresponding safety early warning according to the evaluation prediction result. The problems that it is difficult to judge potential risks only by means of monitoring pictures, effective target locking and tracking are lacked, and risk sources cannot be rapidly determined in an existing monitoring means are solved.

Description

Technical Field

[0001] The present invention relates to the technical field of industrial park safety risk management, and in particular to an industrial safety early warning method based on a large image semantic model. Background Art

[0002] With the acceleration of the industrialization process, in order to comprehensively improve the governance capabilities of industrial parks in the field of safety risk management, the government has formulated a series of strict standards and specifications for production activities in different industries, aiming to effectively reduce the incidence of safety accidents through institutionalized and standardized management methods.

[0003] Against this backdrop, numerous industrial parks have responded positively to this call, deploying advanced risk monitoring systems and installing various sensor devices, including those for security, fire protection, and environmental monitoring. The widespread use of these devices has undoubtedly greatly enhanced the parks' safety risk monitoring and prevention capabilities, providing strong support for the timely identification and resolution of potential safety hazards.

[0004] However, in practical application, traditional monitoring methods still face numerous challenges. For one thing, it's often difficult to detect obvious signs of a security incident before it occurs, making it difficult for monitoring personnel to visually identify potential risks. Furthermore, even if unusual activity is detected in the surveillance footage, the lack of effective target detection and tracking technology often makes it difficult to quickly identify the source of the risk, thus delaying the optimal response. Summary of the Invention

[0005] The purpose of the present invention is to provide an industrial safety early warning method based on a large image semantic model, aiming to solve the problem that existing monitoring means are difficult to judge potential risks based solely on monitoring images, lack effective target locking and tracking, and cannot quickly determine the source of risk.

[0006] The present invention is achieved through the following technical solutions:

[0007] An industrial safety early warning method based on a large image semantic model comprises the following steps:

[0008] Acquire historical image data from industrial sites and pre-process the historical image dataset using an image key change monitoring algorithm based on inter-frame differences. This algorithm calculates the pixel differences between consecutive images and identifies key change images that may contain potential risks.

[0009] Positive samples are labeled for images with potential risks in key change images, and negative samples are labeled for images without potential risks in key change images to obtain a labeled dataset;

[0010] Build an initial industrial safety assessment model based on the image semantic model, and train the industrial safety assessment model through the labeled data set to obtain the industrial safety assessment model;

[0011] Acquire real-time image data of industrial sites and input the real-time image data into the industrial safety assessment model to obtain assessment prediction results, and issue corresponding safety warnings based on the assessment prediction results.

[0012] Optionally, the specific process of acquiring historical image data of an industrial site, preprocessing the historical image data set based on an image key change monitoring algorithm based on inter-frame differences, calculating pixel differences between consecutive images, and identifying key change images containing potential risks is as follows:

[0013] Collect historical image data from industrial field monitoring equipment to form a continuous sequence of image frames;

[0014] By using the image key change detection algorithm of inter-frame difference, the difference between the corresponding pixels of each pair of adjacent image frames in the image frame sequence is calculated to obtain the pixel difference matrix;

[0015] According to the preset pixel difference threshold S, the pixel difference matrix is binarized, and the pixels with a difference greater than the pixel difference threshold S are marked as potential change points, and the remaining pixels are marked as background points;

[0016] Perform morphological processing on the binarized image to eliminate small discrete pixels in the image and thicken the image outline;

[0017] The number of white pixels in the image after morphological processing is counted. When the number of white pixels is greater than the preset threshold T, it is determined that there is a key change containing potential risks between the pair of adjacent image frames, and the pair of image frames is treated as the key change image for subsequent processing.

[0018] Optionally, the specific process of performing positive labeling on images with potential risks in the key change images and performing negative labeling on images without potential risks in the key change images to obtain the labeled data set is as follows:

[0019] Based on industrial production safety standards, we build a safety scenario prompt word dataset targeting potential risks in various industrial production scenarios.

[0020] Select the corresponding safety scenario prompt words based on the industrial production scenes in the key change images;

[0021] The images with potential risks in the key change images are positively labeled using safety scenario prompt words, and the images without potential risks in the key change images are negatively labeled to obtain the labeled dataset.

[0022] Optionally, the specific process of constructing the initial industrial safety assessment model based on the image semantic large model and training the industrial safety assessment model by annotating the data set to obtain the industrial safety assessment model is as follows:

[0023] The image semantic model is used as the basic model. Combined with the professional knowledge in the field of industrial production safety, the basic model is fine-tuned to obtain the initial model for industrial safety assessment.

[0024] The labeled dataset is divided into a training set and a validation set, and the training set is input into the initial model of industrial safety assessment for training;

[0025] After several rounds of iterative training, the performance of the initial industrial safety assessment model on the validation set reaches the preset accuracy threshold, and a trained industrial safety assessment model is obtained.

[0026] Optionally, during the training process of the initial industrial safety assessment model, the LoRA fine-tuning algorithm is adopted to simulate the gradient matrix in training by introducing low-rank decomposition to adjust the number of parameters that need to be updated.

[0027] Optionally, the industrial safety assessment model is quantified using an AWQ quantification algorithm.

[0028] Optionally, the specific process of using the AWQ quantification algorithm to quantify the industrial safety assessment model is as follows:

[0029] The key weights in the industrial safety assessment model are located by activating the distribution; the channels of the key weights are amplified by using the equivalent transformation method; the floating-point data type in the industrial safety assessment model is converted into the integer data type; after the quantization processing is completed, the quantized industrial safety assessment model is verified.

[0030] Optionally, the LMdeploy model acceleration framework is used to accelerate the inference of the quantized industrial safety assessment model. During the inference acceleration process, the industrial safety assessment model is optimized through persistent batch processing, blocking KV cache, dynamic segmentation and fusion, tensor parallel computing, and high-performance CUDA kernel, and the performance of the optimized industrial safety assessment model is tested.

[0031] Optionally, a false positive rate threshold is set for the evaluation and prediction results, and risk categories of the industrial safety assessment model with a false positive rate greater than the threshold in the evaluation and prediction are screened and stored; the screened false positive risk categories are presented to domain experts for verification; based on the feedback from domain experts, the false positive image data is independently collected, and the false positive image data is added to the subsequent iterative training process of the industrial safety assessment model.

[0032] Optionally, the specific process of issuing a corresponding safety warning based on the evaluation prediction result is:

[0033] Analyze the assessment and prediction results of the industrial safety assessment model to identify potential safety risks;

[0034] Generate corresponding security warning information based on the identified potential security risks, combined with the preset security risk levels and response measures library;

[0035] Send security warning information to relevant security managers or systems in real time through preset communication channels to take countermeasures;

[0036] Record the sending time and receipt of security warning information, and track and record the subsequent handling of security warnings.

[0037] The technical solution of the present invention has at least the following advantages and beneficial effects:

[0038] Improve the accuracy of safety risk warnings: By introducing a large model based on image semantics and combining it with an image key change monitoring algorithm based on inter-frame differences, it is possible to automatically identify key change images containing potential risks from historical image data. Then, through positive and negative sample labeling and model training, a highly accurate industrial safety assessment model is constructed. Compared with traditional monitoring methods, it can more accurately capture and identify the precursors of safety incidents, significantly improving the accuracy of safety risk warnings.

[0039] Enhanced risk source locking and tracking capabilities: Leveraging the powerful analytical capabilities of large image semantic models, potential risk sources can be quickly locked and tracked in real-time image data, overcoming the shortcomings of target locking and tracking technology in traditional monitoring methods. This helps monitoring personnel locate risk points in a timely and accurate manner, thereby taking effective measures to intervene and prevent safety accidents.

[0040] Improve the intelligence level of safety risk management: Through automated processing and intelligent analysis, the workload of monitoring personnel is greatly reduced, and the efficiency and intelligence level of safety risk management are improved. At the same time, it can also automatically trigger safety warnings based on assessment and prediction results, providing a more scientific and efficient means for safety risk management in industrial parks.

[0041] Improve the overall safety and stability of industrial parks: enhance the governance capabilities of industrial parks in the field of safety risk management, reduce the incidence of safety accidents, thereby ensuring the stable progress of production activities and the safety of life and property of personnel, and providing strong guarantees for the sustainable development of industrial parks. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 Schematic diagram of the process of an industrial safety early warning method based on a large image semantic model according to an embodiment of the present invention;

[0043] Figure 2 Schematic diagram of the risk monitoring process of the industrial safety early warning method based on the image semantic large model according to an embodiment of the present invention;

[0044] Figure 3 A schematic diagram of the training process of the initial model for industrial safety assessment according to an embodiment of the present invention;

[0045] Figure 4 Schematic diagram of the verification process of the initial model for industrial safety assessment according to an embodiment of the present invention. DETAILED DESCRIPTION

[0046] The following is a specific implementation method with reference to the accompanying drawings.

[0047] Reference Figure 1 、 Figure 2 、 Figure 3 、 Figure 4 , an industrial safety early warning method based on image semantic large model, comprising the steps of:

[0048] Step 1: Obtain historical image data from the industrial site and pre-process the historical image dataset using an image key change monitoring algorithm based on inter-frame differences. This algorithm calculates the pixel differences between consecutive images and identifies key change images that contain potential risks.

[0049] In some embodiments, the specific process of obtaining historical image data of an industrial site, preprocessing the historical image data set using an image key change monitoring algorithm based on inter-frame differences, calculating the pixel differences between consecutive images, and identifying key change images containing potential risks is as follows:

[0050] Historical image data is collected from industrial field monitoring equipment to form a continuous image frame sequence; the difference between the corresponding pixels of each pair of adjacent image frames in the image frame sequence is calculated through the image key change monitoring algorithm based on inter-frame differences to obtain a pixel difference matrix; the pixel difference matrix is binarized according to a preset pixel difference threshold S, and the pixels with a difference greater than the pixel difference threshold S are marked as potential change points, and the remaining pixels are marked as background points; the binarized image is morphologically processed to eliminate small discrete pixels in the image and thicken the image contour; the number of white pixels in the morphologically processed image is counted. When the number of white pixels is greater than the preset threshold T, it is determined that there is a key change containing potential risks between the pair of adjacent image frames, and the pair of image frames is treated as a key change image for subsequent processing. The calculation formula for pixel difference is shown in the following formula (1):

[0051]

[0052] Where i and j represent the subscripts of two consecutive image frames, i.e., image pairs; I1 and I2 represent the i-th and j-th frames, respectively, and I1, I2∈R m×n , indicating that the image is an m×n pixel matrix; S represents the preset pixel difference threshold, which is used to determine whether the pixel point has changed significantly, S∈R; Represents the set of all points whose pixel difference is greater than the threshold S; countFilter is a function used to further process the set of pixels that meet the conditions, including thresholding, morphological processing, etc., and count the number of remaining white points after processing.

[0053] For each pixel (i, j), calculate and The difference between them (usually the absolute value difference or Euclidean distance); for each pixel, if its difference is greater than the threshold S, it is considered that the pixel has changed significantly; the set of pixels that meet the conditions (i.e., the difference is greater than S) is passed to the countFilter function for processing, which thresholds these pixels (usually set to 255, i.e., white), and then performs morphological processing (such as corrosion, expansion, filling small holes, connecting disconnected objects, etc.) to eliminate small discrete pixels and thicken the image outline; after processing, the number of remaining white pixels in the image is counted.

[0054] When the number of white pixels returned by the countFilter function is greater than the preset threshold T, it is considered that there is a key change containing potential risks between the pair of adjacent image frames, and the pair of image frames are treated as key change images for subsequent processing.

[0055] Step 2: Label the images with potential risks in the key change images as positive samples, and label the images without potential risks in the key change images as negative samples to obtain a labeled dataset.

[0056] In some embodiments, the images with potential risks in the key change images are annotated as positive samples, and the images without potential risks in the key change images are annotated as negative samples. The specific process of obtaining the annotated dataset is as follows:

[0057] According to industrial production safety standards, a safety scenario prompt word dataset is constructed for the potential risks in various industrial production scenarios. According to the industrial production scenarios in the key change images, corresponding safety scenario prompt words are selected. The images with potential risks in the key change images are positively labeled with safety scenario prompt words, and the images without potential risks in the key change images are negatively labeled to obtain the labeled dataset.

[0058] Based on government-issued industrial safety standards and industry best practices, safety regulations and requirements for various industrial production scenarios were compiled. Based on these safety regulations and requirements, a dataset of safety scenario prompt words was constructed. These prompt words should accurately describe the potential risks in each scenario and serve as a reference for subsequent annotation. For each key change image, corresponding prompt words were selected from the safety scenario prompt word dataset based on the industrial production scenario, such as hard hat scene, person intrusion scene, standard operating procedure scene, and phone call scene. These prompt words will be used to guide subsequent annotation work, ensuring accuracy and consistency. For each key change image, the selected safety scenario prompt word and the actual content of the image are used to determine whether it presents a potential risk. If the image does contain a potential risk corresponding to the prompt word (such as a person not wearing a hard hat, improper operation, etc.), it is labeled as a positive example. If the image does not contain any potential risk, or the risk is not related to the prompt word, it is labeled as a negative example. During the annotation process, the accuracy and objectivity of the annotation results should be ensured to avoid subjective assumptions and misjudgments. All labeled critical change images are classified into positive and negative samples to form a labeled dataset. This dataset will be used to train the subsequent industrial safety assessment model, enabling it to accurately identify potential risks in industrial production scenarios. After the labeling work is completed, quality control and review should be conducted to ensure the accuracy and consistency of the labeled results. Domain experts or experienced safety management personnel can review the labeled results to promptly identify and correct any errors or omissions. The labeling process can be carried out manually, semi-automatically, or automatically.

[0059] Table 1 below is a comparison table based on government safety standards for the chemical industry:

[0060] Table 1 Comparison example of industrial production safety standards

[0061]

[0062] Create safety scenario prompt words based on the standards in Table 1. Table 2 below shows an example:

[0063] Table 2 Example of safety scenario prompt words

[0064]

[0065] Step 3: Build an initial industrial safety assessment model based on the image semantic model, train the industrial safety assessment model through the labeled data set, and obtain the industrial safety assessment model.

[0066] In some embodiments, an initial industrial safety assessment model is constructed based on a large image semantic model, and the industrial safety assessment model is trained using a labeled dataset. The specific process of obtaining the industrial safety assessment model is as follows:

[0067] Using the image semantics big model as the base model and combining it with expertise in the field of industrial production safety, we fine-tune the base model to obtain an initial model for industrial safety assessment. Open-source image semantic big models, such as ViT (Vision Transformer), can be used. Fine-tuning the base model involves adding specific layers or modules related to industrial production safety and adjusting the model's parameters to better suit the specific tasks of industrial production safety. The fine-tuning process can be performed through the following steps:

[0068] Data preparation: Collect image data related to industrial production safety and perform appropriate preprocessing, such as cropping, scaling, and normalization;

[0069] Model structure adjustment: Adjust the model structure according to task requirements, such as adding specific convolutional layers, attention mechanisms, or fully connected layers;

[0070] Loss function design: Design a suitable loss function to measure the performance of the model on industrial production safety tasks;

[0071] Training strategy: Determine key parameters such as learning rate, batch size, and number of iterations during training.

[0072] The labeled dataset is divided into a training set and a validation set. The training set is used to train the initial industrial safety assessment model, while the validation set is used to evaluate the model's performance and adjust its parameters during training. The split ratio can be determined based on the specific task and dataset size.

[0073] After several rounds of iterative training, the performance of the initial industrial safety assessment model on the validation set reaches the preset accuracy threshold, and a trained industrial safety assessment model is obtained.

[0074] In some embodiments, during the training process of the initial model for industrial safety assessment, in order to reduce storage and computing costs while maintaining the performance of the model in vertical fields, the Lora fine-tuning algorithm can be used to simulate the gradient matrix in training by introducing low-rank decomposition and adjust the number of parameters that need to be updated. During the training process, the model needs to be verified regularly using a validation set to evaluate the generalization ability of the model. The verification process can include calculating indicators such as the accuracy, recall rate, F1 score of the model, and observing whether the performance of the model on the validation set is stable. Based on the verification results, the model parameters, learning rate and other hyperparameters can be adjusted, and iterative training can be performed. The iterative training process can be repeated many times until the performance of the model on the validation set reaches a preset accuracy threshold.

[0075] In some embodiments, in order to reduce the storage and computing requirements of the model, the AWQ (Adaptive Weighted Quantization) quantization algorithm is used to quantize the industrial safety assessment model. The specific process is as follows:

[0076] Key weights in the industrial safety assessment model are located through activation distribution to reduce quantization errors; equivalent transformations are used to amplify the channels of key weights; floating-point numbers (float32) in the industrial safety assessment model are converted to integers (Int4); and after quantization, the quantized industrial safety assessment model is verified. Since the large image semantic model still uses the mainstream transformer structure, by mapping image features to the vector space where semantic features reside, and then combining the two and inputting them into the transformer's hidden layer, the input vector changes from the plain text token_id of the LLM to the image embedding plus the text token_id. Subsequent calculations based on the transformer structure are exactly the same as those of the pure language model (LLM). Therefore, the mathematical principles of the AWQ algorithm can be maintained and the framework can be expanded to be compatible with mainstream large image semantic models.

[0077] In some embodiments, the LMdeploy model acceleration framework is used to accelerate the inference of the quantized industrial safety assessment model. During the inference acceleration process, the industrial safety assessment model is optimized through persistent batch processing, blocking KV cache, dynamic segmentation and fusion, tensor parallel computing, and high-performance CUDA kernels, and the performance of the optimized industrial safety assessment model is tested. The average inference speed is 2.4 times faster than that of conventional FP16 inference, and the request throughput is increased by about 1.8 times. Table 3 below shows the performance of the industrial safety assessment model on the A100 graphics card after using the LMdeploy model acceleration framework:

[0078] Table 3 Industrial safety assessment model acceleration optimization parameters

[0079]

[0080] Among them, model indicates the model type used; batch indicates the batch size, that is, the maximum batch size that can be processed during inference; tp indicates the number of tensor parallel GPUs, that is, the total number of GPU cards participating in tensor parallel computing; num_prompts indicates the number of prompts, which is equal to the total number of requests received; PRS indicates the number of requests per second; FTL indicates the first token delay, that is, the response delay time of the first token; throughput indicates the throughput, that is, the number of tokens that can be output per second.

[0081] To evaluate the performance of the industrial safety assessment model in industrial settings, we developed a test set specifically for helmet wearing. This test set was set in the refining furnace operating area of a steel plant. According to safety regulations, workers are required to wear helmets. Table 4 below shows the performance of the fine-tuned industrial safety assessment model on this test set:

[0082] Table 4. Characterization of industrial safety assessment model on the specially designed test set

[0083]

[0084]

[0085] Images labeled "No One" indicate that the industrial scene is in standby mode, with no personnel present. This demonstrates that the large-scale model can accurately identify the working status of industrial scenes. The industrial safety assessment model achieved 90% accuracy on this test set.

[0086] Step 4: Acquire real-time image data from the industrial site and input it into the industrial safety assessment model to obtain assessment prediction results. Based on these results, issue appropriate safety warnings. Capture image data in real time from surveillance cameras at the industrial site. Ensure that the clarity and quality of the image data meet the model input requirements. Store and transmit the captured image data in an appropriate format (e.g., JPEG, PNG, etc.). Perform necessary preprocessing on the captured real-time image data, such as resizing and normalizing it, to ensure that it is consistent with the image data format used during training. Apply an inter-frame difference image key change detection algorithm to filter out static or slightly changing images to reduce the model's computational burden.

[0087] In some embodiments, a false positive rate threshold is set for the evaluation and prediction results, and risk categories of the industrial safety assessment model with a false positive rate greater than the threshold in the evaluation and prediction are screened and stored; the screened false positive risk categories are presented to domain experts for verification; based on the feedback from domain experts, the false positive image data is independently collected, and the false positive image data is added to the subsequent iterative training process of the industrial safety assessment model.

[0088] In some embodiments, the specific process of issuing corresponding security warnings based on the evaluation and prediction results is as follows:

[0089] Analyze the assessment and prediction results of the industrial safety assessment model to identify potential safety risks;

[0090] Generate corresponding security warning information based on the identified potential security risks, combined with the preset security risk levels and response measures library;

[0091] Send security warning information to relevant security managers or systems in real time through preset communication channels to take countermeasures;

[0092] Record the sending time and receipt of security warning information, and track and record the subsequent handling of security warnings.

Claims

1. An industrial safety early warning method based on a large image semantic model, characterized in that: Including steps: Acquire historical image data from industrial sites and pre-process the historical image dataset using an image key change monitoring algorithm based on inter-frame differences. This algorithm calculates the pixel differences between consecutive images and identifies key change images that may contain potential risks. Positive samples are labeled for images with potential risks in key change images, and negative samples are labeled for images without potential risks in key change images to obtain a labeled dataset; Build an initial industrial safety assessment model based on the image semantic model, and train the industrial safety assessment model through the labeled data set to obtain the industrial safety assessment model; Acquire real-time image data of industrial sites and input the real-time image data into the industrial safety assessment model to obtain assessment prediction results, and issue corresponding safety warnings based on the assessment prediction results.

2. The industrial safety early warning method based on the image semantic large model according to claim 1 is characterized in that: The specific process of obtaining historical image data of an industrial site, preprocessing the historical image data set based on an image key change monitoring algorithm based on inter-frame differences, calculating the pixel differences between consecutive images, and identifying key change images containing potential risks is as follows: Collect historical image data from industrial field monitoring equipment to form a continuous sequence of image frames; By using the image key change detection algorithm of inter-frame difference, the difference between the corresponding pixels of each pair of adjacent image frames in the image frame sequence is calculated to obtain the pixel difference matrix; According to the preset pixel difference threshold S, the pixel difference matrix is binarized, and the pixels with a difference greater than the pixel difference threshold S are marked as potential change points, and the remaining pixels are marked as background points; Perform morphological processing on the binarized image to eliminate small discrete pixels in the image and thicken the image outline; The number of white pixels in the image after morphological processing is counted. When the number of white pixels is greater than the preset threshold T, it is determined that there is a key change containing potential risks between the pair of adjacent image frames, and the pair of image frames is treated as the key change image for subsequent processing.

3. The industrial safety early warning method based on image semantic big model according to claim 1 is characterized in that: The specific process of labeling the images with potential risks in the key change images and labeling the images without potential risks in the key change images as negative samples to obtain the labeling dataset is as follows: Based on industrial production safety standards, we build a safety scenario prompt word dataset targeting potential risks in various industrial production scenarios. Select the corresponding safety scenario prompt words based on the industrial production scenes in the key change images; The images with potential risks in the key change images are positively labeled using safety scenario prompt words, and the images without potential risks in the key change images are negatively labeled to obtain the labeled dataset.

4. The industrial safety early warning method based on image semantic big model according to claim 1 is characterized in that: The specific process of constructing the initial industrial safety assessment model based on the image semantic model and training the industrial safety assessment model through the labeled data set to obtain the industrial safety assessment model is as follows: The image semantic model is used as the basic model. Combined with the professional knowledge in the field of industrial production safety, the basic model is fine-tuned to obtain the initial model for industrial safety assessment. The labeled dataset is divided into a training set and a validation set, and the training set is input into the initial model of industrial safety assessment for training; After several rounds of iterative training, the performance of the initial industrial safety assessment model on the validation set reaches the preset accuracy threshold, and a trained industrial safety assessment model is obtained.

5. The industrial safety early warning method based on image semantic big model according to claim 4 is characterized in that: During the training process of the industrial safety assessment initial model, the Lora fine-tuning algorithm is adopted to simulate the gradient matrix in training by introducing low-rank decomposition and adjust the number of parameters that need to be updated.

6. The industrial safety early warning method based on image semantic big model according to claim 1 is characterized in that: The AWQ quantification algorithm is used to quantify the industrial safety assessment model.

7. The industrial safety early warning method based on image semantic big model according to claim 6 is characterized in that: The specific process of using the AWQ quantification algorithm to quantify the industrial safety assessment model is as follows: The key weights in the industrial safety assessment model are located by activating the distribution; the channels of the key weights are amplified by using the equivalent transformation method; the floating-point data type in the industrial safety assessment model is converted into the integer data type; after the quantization processing is completed, the quantized industrial safety assessment model is verified.

8. The industrial safety early warning method based on image semantic big model according to claim 6 is characterized in that: The LMdeploy model acceleration framework is used to accelerate the inference of the quantified industrial safety assessment model. During the inference acceleration process, the industrial safety assessment model is optimized through persistent batch processing, blocking KV cache, dynamic segmentation and fusion, tensor parallel computing, and high-performance CUDA kernel, and the performance of the optimized industrial safety assessment model is tested.

9. The industrial safety early warning method based on image semantic big model according to claim 1 is characterized in that: Set a false positive rate threshold for the assessment and prediction results, filter and store risk categories with a false positive rate greater than the threshold in the assessment and prediction results of the industrial safety assessment model; and present the filtered false positive risk categories to domain experts for verification. Based on the feedback from domain experts, the misjudged image data is collected independently and included in the subsequent iterative training of the industrial safety assessment model.

10. The industrial safety early warning method based on image semantic big model according to claim 1, characterized in that: The specific process of making corresponding safety warnings based on the evaluation and prediction results is as follows: Analyze the assessment and prediction results of the industrial safety assessment model to identify potential safety risks; Generate corresponding security warning information based on the identified potential security risks, combined with the preset security risk levels and response measures library; Send security warning information to relevant security managers or systems in real time through preset communication channels to take countermeasures; Record the sending time and receipt of security warning information, and track and record the subsequent handling of security warnings.

Images