Operation site safety supervision method based on AI vision
By building a risk level tree and jointly training the safety supervision model and feature transformation model, the problem of insufficient risk feature aggregation and divergence capabilities in the existing methods is solved, and intelligent and efficient supervision of safety supervision on the operation site is realized.
Patent Information
- Application Number
- CN202510423859.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-04-07
- Publication Date
- 2025-07-18
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
The existing security supervision method based on computer vision cannot effectively integrate the complex and diverse risk information on the cooperative industry site, lacks full utilization of risk level relationships, cannot achieve the aggregation of similar risk characteristics and the divergence of dissimilar risk characteristics, and lacks effective loss function design, which limits the performance improvement of the model.
Build a risk level tree, jointly train the security supervision model and feature conversion model, use the security maximum and minimum loss function and the security four-element group loss function, and realize the aggregation of similar risk level feature vectors and divergence of different risk level feature vectors through parameter optimization of feature extraction layer, supervision layer and conversion layer.
It improves the accuracy and efficiency of safety supervision on the work site, realizes intelligent analysis and automated supervision of complex risk environments, reduces the subjectivity of manual supervision, and improves the level of safety management and work efficiency.
Smart Images
Figure CN120337018A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical fields of artificial intelligence and computer vision, and more specifically, the present invention relates to a method for safety supervision of a work site based on AI vision. Background Art
[0002] In modern industrial production, safety supervision of the work site is a key link to ensure the safety of personnel's lives and the smooth progress of production. Traditional safety supervision methods mainly rely on manual inspections and simple video monitoring systems. Manual inspections require a large amount of human input and are easily affected by subjective factors, making it difficult to achieve comprehensive coverage and real-time supervision. Although simple video monitoring systems can record the on-site situation, they lack intelligent analysis capabilities and cannot automatically identify and warn of potential safety risks. In recent years, with the development of artificial intelligence technology, safety supervision methods based on computer vision have gradually received attention. These methods can automatically detect information such as personnel behavior and equipment status in the work site through image recognition and analysis technology, thereby achieving a certain degree of safety supervision. However, most of the existing safety supervision methods based on computer vision can only identify simple safety violations and have insufficient comprehensive analysis capabilities for complex work site environments and various safety risk factors. In addition, the existing methods have limitations in feature extraction and risk classification and cannot make full use of the hierarchical relationship between risk levels to optimize the performance of the supervision model, resulting in the accuracy and reliability of the supervision results remaining to be improved.
[0003] In the process of implementing the embodiments of the present invention, the inventors found that there are at least the following problems or defects in the prior art: the existing safety supervision methods cannot effectively integrate the complex and diverse risk information in the work site and lack the full use of the hierarchical relationship of risk levels; in the process of feature extraction and transformation, it is impossible to aggregate similar risk features and diverge dissimilar risk features, resulting in insufficient discrimination ability of the model for different risk features; at the same time, the existing methods lack an effective loss function design in the training process and cannot fully guide the model to learn the distance relationship between risk probability transformation and feature vectors, thereby limiting the performance improvement of the safety supervision model. Summary of the Invention
[0004] The present invention provides a method for safety supervision of a work site based on AI vision, including:
[0005] Obtaining a work site data set and safety labels corresponding to each data in the data set, and constructing a risk level tree based on the risk level relationship of each safety label;
[0006] Based on the job site dataset, the safety labels corresponding to each data in the dataset, and the risk level tree, the safety supervision model and the feature transformation model are trained simultaneously to obtain a trained safety supervision model; wherein, the safety supervision model includes a feature extraction layer shared with the feature transformation model for extracting the feature information of the input data;
[0007] The safety supervision model further includes a supervision layer for performing safety supervision on the input data based on the number of nodes of the risk level tree; during training, the supervised risk probability value is transformed based on the risk level relationship of each safety label in the risk level tree;
[0008] The feature transformation model further includes a transformation layer for reorganizing the features of the input data to obtain a new feature vector; during training, based on the similarity information of the labels in the risk level tree, the new feature vectors with similar risk levels are aggregated with each other in the feature space, and the new feature vectors with dissimilar risk levels diverge from each other.
[0009] Further, the simultaneous training of the safety supervision model and the feature transformation model to obtain a trained safety supervision model includes: loading the job site dataset, the safety labels corresponding to each data, and the risk level tree, and simultaneously training the safety supervision model and the feature transformation model using the safety maximum-minimum loss function and the safety four-element group loss function, updating the parameters of the feature extraction layer, the supervision layer, and the transformation layer using gradient backpropagation, saving the parameters of the feature extraction layer, the supervision layer, and the transformation layer after training, and obtaining the safety supervision model based on the trained feature extraction layer and the supervision layer.
[0010] Further, the safety maximum-minimum loss function is expressed as:
[0011]
[0012] where B represents the number of data in the job site dataset; M represents the height of the data; N represents the width of the data; |V| represents the number of nodes in the risk level tree; y b,i,j,k represents the label value corresponding to the k-th risk level of the (i, j) risk point in the b-th data; represents the risk probability value after the risk maximum-minimum transformation of the label value corresponding to the k-th risk level of the (i, j) risk point in the b-th data.
[0013] Further, the transformed risk probability value of each risk point is obtained using the following formula:
[0014]
[0015] where A kDenote the set of parent nodes of risk level k in the risk level tree; D k Denote the set of child nodes of risk level k in the risk level tree; y b,i,j,l Denote the label value corresponding to the l-th risk level of the (i, j) risk point in the b-th data; Denote the transformed risk probability value corresponding to the k-th risk level of the (i, j) risk point in the b-th data.
[0016] Furthermore, the loss function of the security four-element group is expressed as:
[0017]
[0018] where DTT represents the set of effective security four-element group vectors; N b Denote the number of four-element groups in the set of effective security four-element group vectors; f s a Denote the first eigenvector in the s-th set of effective security four-element group vectors; f s p Denote the second eigenvector in the s-th set of effective security four-element group vectors; f s n Denote the third eigenvector in the s-th set of effective security four-element group vectors; f s q Denote the fourth eigenvector in the s-th set of effective security four-element group vectors; d() represents the vector distance function; m s Denote the security optimal margin of the s-th set of effective security four-element group vectors; Arbitrarily select four new eigenvectors output by the conversion layer and the corresponding security labels to form a set of four-element group vectors; Traverse the set of four-element group vectors to obtain a set of effective security four-element group vectors based on the risk level tree.
[0019] Furthermore, the traversing the set of four-element group vectors to obtain a set of effective security four-element group vectors based on the risk level tree includes: The shortest path length between the label of the first eigenvector and the label of the second eigenvector in the risk level tree in the set of four-element group vectors is less than the shortest path length between the label of the first eigenvector and the label of the third eigenvector in the risk level tree, and less than the shortest path length between the label of the first eigenvector and the label of the fourth eigenvector in the risk level tree.
[0020] Furthermore, use the following formula to calculate the security optimal margin of the set of effective security four-element group vectors:
[0021]
[0022] where The label representing the first eigenvector in the s-th valid secure four-factor group vector set; The label representing the third eigenvector in the s-th valid secure four-factor group vector set; The label representing the second eigenvector in the s-th valid secure four-factor group vector set; The label representing the fourth eigenvector in the s-th valid secure four-factor group vector set; Ψ() represents the shortest path length between two labels in the risk level tree.
[0023] Furthermore, constructing the risk level tree based on the risk level relationships of each security label includes: for each security label, looking up its parent class in the risk database and searching layer by layer upwards until reaching the root node of the risk database to obtain the risk level relationship of each label;
[0024] Based on the risk level relationships of each label, construct the risk level tree T = {V, E}; where V represents the set of all nodes in the risk level tree, which are the label classes and their parent classes in the risk database; E represents the node relationships in the risk level tree.
[0025] Furthermore, in the security supervision model, the extracted feature information passes through the supervision layer, and based on the extracted feature information, risk classification is performed on each risk point in the data to obtain the security supervision result; where the feature information passes through a feature transformation module and performs feature transformation operations on the feature map to obtain the transformed feature information; where |V| represents the number of nodes in the risk level tree;
[0026] The transformed feature information passes through the Gaussian Error Linear Unit GELU activation function to perform non-linear transformation on each feature to obtain the non-linearly transformed feature information;
[0027] The non-linearly transformed feature information passes through a feature transformation module again and then through the sigmoid function to obtain the probability distribution of the risk for each risk point.
[0028] Furthermore, in the feature conversion model, the extracted feature information passes through the conversion layer to map the feature information in the data to a new space; where the feature information passes through a feature conversion kernel and performs feature conversion operations on the feature map to obtain the transformed feature information;
[0029] The transformed feature information passes through the Gaussian Error Linear Unit GELU activation function to perform non-linear transformation on each feature to obtain the non-linearly transformed feature information;
[0030] The non-linearly transformed feature information passes through a feature conversion kernel again and then through the sigmoid function to obtain the feature vector mapped to the new space.
[0031] The above embodiments of the present invention have at least the following beneficial effects: The present invention can effectively improve the accuracy and efficiency of safety supervision at the operation site. By constructing a risk level tree, various risk information at the operation site can be systematically integrated, enabling the safety supervision model to accurately classify and supervise different risk points based on the risk level relationship. At the same time, the feature transformation model can reorganize and optimize the features of the input data, causing the feature vectors of similar risk levels to aggregate in the feature space and the feature vectors of dissimilar risk levels to diverge from each other, thereby enhancing the model's ability to distinguish different risk features and further improving the accuracy of safety supervision. In addition, the combined use of the safety maximum-minimum loss function and the safety four-element group loss function can better guide the model training process, enabling the model to fully consider the transformation of risk probabilities and the distance relationship between feature vectors during the learning process, further enhancing the model's ability to identify and judge safety risks at the operation site and providing a more reliable basis for safety management at the operation site.
[0032] The present invention can also achieve the automation and intelligence of safety supervision at the operation site, reducing the subjectivity and limitations of manual supervision. Based on the trained safety supervision model, real-time data at the operation site can be quickly analyzed and processed, potential safety risks can be discovered in a timely manner, and corresponding warning information can be issued, providing timely safety guidance and decision-making support for on-site workers. This method can not only improve the safety management level at the operation site, reduce the probability of safety accidents, but also save labor costs and improve work efficiency. BRIEF DESCRIPTION OF THE DRAWINGS
[0033] By referring to the following detailed description with reference to the accompanying drawings, the above and other objects, features, and advantages of the exemplary embodiments of the present invention will become readily understood. In the drawings, several embodiments of the present invention are shown in an exemplary and non-limiting manner, wherein:
[0034] Figure 1 It is a schematic flowchart of a method for safety supervision of an operation site based on AI vision provided by an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS
[0035] The principles and spirit of the present invention will be described below with reference to several exemplary embodiments. It should be understood that these embodiments are provided only to enable those skilled in the art to better understand and implement the present invention, and do not limit the scope of the present invention in any way. On the contrary, these embodiments are provided to make the present invention more thorough and complete, and to be able to convey the scope of the present invention fully to those skilled in the art.
[0036] Those skilled in the art know that the embodiments of the present invention can be implemented as a system, device, equipment, method, or computer program product. Therefore, the present invention can be specifically implemented in the following forms, namely: completely hardware, completely software (including firmware, resident software, microcode, etc.), or a combination of hardware and software.
[0037] It should be noted that any number of elements in the drawings is for illustration rather than limitation, and any naming is only for distinction and does not have any limiting meaning.
[0038] The following refers to Figure 1 , Figure 1 a schematic flowchart of a method for safety supervision of a work site based on AI vision provided by an embodiment of the present invention. As Figure 1 shown, a method for safety supervision of a work site based on AI vision includes:
[0039] S1 Obtain a work site dataset and safety labels corresponding to each data in the dataset, and construct a risk level tree based on the risk level relationship of each safety label;
[0040] S2 Based on the work site dataset, safety labels corresponding to each data in the dataset, and the risk level tree, simultaneously train a safety supervision model and a feature transformation model to obtain a trained safety supervision model; wherein, the safety supervision model includes a feature extraction layer shared with the feature transformation model for extracting feature information of input data;
[0041] S3 The safety supervision model further includes a supervision layer for performing safety supervision on input data based on the number of nodes of the risk level tree; during training, transform the supervised risk probability value based on the risk level relationship of each safety label in the risk level tree;
[0042] S4 The feature transformation model further includes a transformation layer for reorganizing the features of input data to obtain a new feature vector; during training, based on the similarity information of the labels in the risk level tree, make the new feature vectors of similar risk levels aggregate with each other in the feature space, and the new feature vectors of dissimilar risk levels diverge from each other.
[0043] It should be noted that the present invention proposes a method for safety supervision of the operation site based on AI vision. The core lies in achieving efficient safety supervision of the operation site by constructing a risk level tree and simultaneously training a safety supervision model and a feature transformation model. The operation site dataset refers to the images, videos, or other relevant data collected from the operation site, which contain various information of the operation site, such as personnel behavior, equipment status, etc. The safety label is the safety risk category labeled in these data, such as not wearing a safety helmet, abnormal equipment operation, etc. The risk level tree is a hierarchical data structure used to represent the risk level relationship between different safety labels. For example, not wearing a safety helmet may belong to a higher-level risk category of insufficient personal protection. Through this structured representation, the safety risks of the operation site can be better understood and managed. At the same time, the joint training of the safety supervision model and the feature transformation model enables the model to learn more effective feature representations from the data and perform accurate safety supervision based on the risk level tree.
[0044] Specifically, the operation site dataset is collected through cameras or other sensors deployed in the actual operation site, and these data are used for model training after preprocessing. The safety label is obtained by labeling the risk points in the data according to the safety specifications and standards of the operation site. For example, in an image, a person not wearing a safety helmet is labeled. The construction of the risk level tree is based on the risk level relationship in the risk database, where each node represents a risk category, and the relationship between the parent node and the child node represents the high or low risk level. For example, insufficient personal protection is the parent node, and not wearing a safety helmet is its child node, indicating that not wearing a safety helmet is a specific manifestation of the higher-level risk of insufficient personal protection. The safety supervision model includes a feature extraction layer and a supervision layer. The feature extraction layer is used to extract useful feature information from the input data. For example, the edges, textures, etc. of an image are extracted through a convolutional neural network; the supervision layer classifies the extracted features based on the number of nodes in the risk level tree. The transformation layer of the feature transformation model is used to reorganize the features to generate new feature vectors, and these feature vectors are aggregated or diverged in the feature space according to the similarity of the risk level, thereby optimizing the performance of the model.
[0045] Preferably, the construction process of the risk level tree can be further refined as follows: First, extract the risk level information of each safety label from the risk database, and then construct a tree structure based on this information. For example, for the safety label of not wearing a safety helmet, the system will look up its parent category of insufficient personal protection in the risk database and continue to look up upward until reaching the root node of the safety risk at the job site. During the model training process, the feature extraction layer of the safety supervision model can use a pre-trained convolutional neural network (such as ResNet or VGG) as the basic architecture, with the input parameter being the image data of the job site. After multiple convolutional and pooling operations, the feature information of the image is extracted. The conversion layer of the feature conversion model can use techniques such as autoencoders or attention mechanisms, taking the output features of the feature extraction layer as input. By learning the similarity relationships between the features, the features are mapped to a new space, making the feature vectors of similar risk levels closer in the feature space and the feature vectors of dissimilar risk levels more dispersed. Such refined operation steps and parameter settings can further improve the performance and supervision effect of the model.
[0046] In some embodiments, the simultaneous training of the safety supervision model and the feature conversion model to obtain a trained safety supervision model includes: loading the job site dataset, the safety labels corresponding to each data, and the risk level tree, and simultaneously training the safety supervision model and the feature conversion model using the safety max-min loss function and the safety four-element group loss function. Update the parameters of the feature extraction layer, the supervision layer, and the conversion layer using gradient backpropagation. After the training is completed, save the parameters of the feature extraction layer, the supervision layer, and the conversion layer, and obtain a safety supervision model based on the trained feature extraction layer and supervision layer.
[0047] It should be noted that the process of simultaneously training the safety supervision model and the feature conversion model in the present invention is a key link to achieve efficient safety supervision. During the training process, the safety max-min loss function and the safety four-element group loss function are used. The design of these two loss functions aims to optimize the performance of the model so that it can better handle the safety supervision tasks at the job site. The safety max-min loss function transforms the risk probability values to ensure that the model can accurately identify and distinguish features of different risk levels. The safety four-element group loss function further enhances the model's ability to distinguish risk features by optimizing the distance relationship between the feature vectors. By updating the model parameters using gradient backpropagation, a trained safety supervision model is finally obtained, which can effectively supervise the job site data based on the feature extraction layer and the supervision layer.
[0048] Specifically, the safety maximum-minimum loss function and the safety four-factor group loss function are two key loss functions used to optimize model training in the present invention. The core of the safety maximum-minimum loss function lies in transforming the risk probability values, enabling the model to better handle the relationships between different risk levels. For example, for a specific risk point, the model needs to distinguish the features of low-risk levels from those of high-risk levels according to the hierarchical relationship in the risk level tree. The safety four-factor group loss function ensures that feature vectors of similar risk levels are close to each other in the feature space and those of dissimilar risk levels are far apart by optimizing the distance relationship between feature vectors. During the training process, the parameters of the feature extraction layer, supervision layer, and transformation layer of the model are updated through gradient backpropagation. The parameter update of the feature extraction layer is to extract more effective feature information, the parameter update of the supervision layer is to improve the accuracy of risk classification, and the parameter update of the transformation layer is to optimize the distribution of feature vectors. After the training is completed, these parameters are saved for subsequent use of the trained model for safety supervision.
[0049] Preferably, the construction and training processes of the safety supervision model and the feature transformation model can be further refined. The feature extraction layer of the safety supervision model can adopt a deep convolutional neural network (such as ResNet or VGG) as the basic architecture, with the input parameter being the image data of the job site. After multiple convolutional and pooling operations, the feature information of the image is extracted. The supervision layer classifies the extracted features based on the number of nodes in the risk level tree. The transformation layer of the feature transformation model can adopt techniques such as autoencoders or attention mechanisms, inputting the output features of the feature extraction layer, and mapping the features to a new space by learning the similarity relationships between features. During the training process, the safety maximum-minimum loss function ensures that the model can accurately identify the features of different risk levels by transforming the risk probability values. The safety four-factor group loss function further enhances the model's ability to distinguish risk features by optimizing the distance relationship between feature vectors. For example, for a job site image containing multiple risk points, the model first extracts the image features through the feature extraction layer, then conducts risk classification through the supervision layer, and at the same time optimizes the distribution of feature vectors through the transformation layer. Finally, through the joint optimization of the two loss functions, the model can more accurately identify and distinguish the features of different risk levels.
[0050] In some embodiments, the safety maximum-minimum loss function is expressed as:
[0051]
[0052] where B represents the number of data in the job site dataset; M represents the height of the data; N represents the width of the data; |V| represents the number of nodes in the risk level tree; y b,i,j,kDenote the label value corresponding to the k-th risk level of the (i, j) risk point in the b-th data; Denote the risk probability value after the maximum-minimum risk transformation of the label value corresponding to the k-th risk level of the (i, j) risk point in the b-th data.
[0053] It should be noted that the safety maximum-minimum loss function mentioned in the present invention is an important part for optimizing the model training process. This loss function evaluates and adjusts the prediction results of the model by considering the risk points and corresponding risk levels of each data in the job site dataset. Its purpose is to enable the model to better learn and distinguish the differences between different risk levels by transforming the risk probability values, thereby improving the accuracy and reliability of the safety supervision model. In this loss function, the key parameters involved include the number of data in the job site dataset, the height and width of the data, and the number of nodes in the risk level tree, etc. These parameters jointly determine the calculation process and optimization objective of the loss function.
[0054] Specifically, the design of the safety maximum-minimum loss function is based on multiple key parameters of the job site dataset. Among them, the number of data refers to the total number of samples in the dataset, and these samples can be images, video frames or other forms of data, used to train the model to identify and classify safety risks. The height and width of the data represent the spatial dimensions of each data sample. For example, in image data, the height and width correspond to the vertical and horizontal pixel numbers of the image respectively. The number of nodes in the risk level tree refers to the total number of different risk levels included in the risk level tree, and these risk levels are organized in a tree structure, reflecting the hierarchical relationship between different safety labels. In the calculation process of the loss function, the risk points in each data sample will be evaluated and processed according to their corresponding risk level label values. Through the maximum-minimum risk transformation, the prediction results (i.e., risk probability values) of the model will be adjusted to better reflect the actual risk level relationship. This process ensures that the model can fully learn the differences between different risk levels during the training process, thereby improving the supervision effect.
[0055] Preferably, the calculation process of the safety maximum-minimum loss function can be further refined through the following steps. First, for each data sample in the job site dataset, the model outputs the risk probability values corresponding to each risk point. These probability values reflect the prediction confidence of the model for each risk point belonging to different risk levels. Then, through the risk maximum-minimum transformation, these probability values are adjusted to new risk probability values. This transformation process takes into account the hierarchical relationship in the risk level tree, enabling the model to better handle the relative differences between different risk levels. Specifically, the transformed risk probability values are adjusted according to the parent node and child node sets of each node in the risk level tree. For example, if a risk point belongs to a certain high-risk level, and the parent node set of this level contains lower-risk levels, then the transformed probability value will increase accordingly to highlight the importance of the high-risk level. In this way, the safety maximum-minimum loss function can effectively guide the model to learn more accurate risk classification capabilities, thereby improving the reliability of job site safety supervision in practical applications.
[0056] In some embodiments, the following formula is used to obtain the transformed risk probability values of each risk point:
[0057]
[0058] where A k represents the parent node set of risk level k in the risk level tree; D k represents the child node set of risk level k in the risk level tree; y b,i,j,l represents the label value corresponding to the l-th risk level of the (i, j) risk point in the b-th data; represents the transformed risk probability value corresponding to the k-th risk level of the (i, j) risk point in the b-th data.
[0059] It should be noted that the transformation process of the risk probability values mentioned in the present invention is an important part in the training of the safety supervision model. Through this transformation, the model can better handle and understand the relative relationship between different risk levels, thereby improving the ability to identify and classify job site safety risks. This transformation process is based on the structure of the risk level tree, taking into account the relative position of each risk level in its parent node and child node sets, enabling the model to more accurately reflect the hierarchical relationship between risk levels. This transformation method not only optimizes the output result of the model, but also enhances the adaptability of the model to complex risk scenarios.
[0060] Specifically, the transformation process of the risk probability value involves the parent node set and the child node set of the risk level tree. The parent node set refers to the set of all superior nodes of a certain risk level in the risk level tree, and these parent nodes represent higher-level risk categories. The child node set refers to the set of all inferior nodes of a certain risk level, and these child nodes represent more specific risk manifestation forms. During the transformation process, the model adjusts the corresponding probability value according to the positions of each risk level in its parent node and child node sets. For example, if the parent node set of a risk level contains multiple high-risk levels, the transformed probability value of this risk level may be increased accordingly to reflect its higher risk level. On the contrary, if the child node set of a risk level contains multiple low-risk levels, its transformed probability value may be decreased accordingly. Through this transformation method based on the risk level tree structure, the model can more accurately evaluate the actual risk level of each risk point.
[0061] Preferably, the transformation process of the risk probability value can be further refined through the following steps. First, the model determines the parent node set and the child node set of each risk level according to the structure of the risk level tree. Then, for each risk point, the model calculates its relative position weights in the parent node set and the child node set. These weights can be obtained through statistical methods or model learning, and are used to reflect the importance of each risk level in its context. Next, the model adjusts the risk probability value according to these weights. For example, for a risk point, the high-risk weight in its parent node set is relatively large, and the low-risk weight in its child node set is relatively small. Then, the model adjusts the probability value of this risk point to a value closer to the high risk through a weighting formula. This weighted transformation method not only considers the structural information of the risk level tree but also can be dynamically adjusted according to actual data, thereby improving the accuracy and robustness of the model for risk assessment.
[0062] In some embodiments, the loss function of the secure four-factor group is expressed as:
[0063]
[0064] where DTT represents the set of valid secure four-factor group vectors; N b represents the number of four-factor groups in the set of valid secure four-factor group vectors; f s a represents the first feature vector in the s-th set of valid secure four-factor group vectors; f s p represents the second feature vector in the s-th set of valid secure four-factor group vectors; f s n represents the third feature vector in the s-th set of valid secure four-factor group vectors; f sq represents the fourth eigenvector in the s-th effective secure four-factor group vector set; d() represents the vector distance function; m s represents the secure optimal margin of the s-th effective secure four-factor group vector set; randomly select four new eigenvectors output by the conversion layer and their corresponding security labels to form a four-factor group vector set; traverse the four-factor group vector set and obtain the effective secure four-factor group vector set based on the risk level tree.
[0065] It should be noted that the secure four-factor group loss function mentioned in the present invention is an important tool for optimizing the training processes of the feature conversion model and the security supervision model. By considering the distance relationship between eigenvectors, this loss function ensures that the model can better distinguish eigenvectors of different risk levels. Specifically, it uses the hierarchical relationship in the risk level tree to screen out effective eigenvector combinations, and by optimizing the distances between these combinations, eigenvectors of similar risk levels are made closer in the feature space, while eigenvectors of dissimilar risk levels are made farther apart. This method can significantly improve the model's ability to distinguish risk features, thereby enhancing the accuracy of security supervision.
[0066] Specifically, the core of the secure four-factor group loss function lies in the construction and optimization of the effective secure four-factor group vector set. The effective secure four-factor group vector set consists of four eigenvectors and their corresponding security labels, and these four eigenvectors respectively represent features of different risk levels. When constructing the four-factor group vector set, the shortest path length in the risk level tree needs to be considered, that is, the hierarchical distance between any two labels in the tree. For example, the shortest path length between the labels of the first eigenvector and the second eigenvector in the risk level tree should be less than the shortest path length between the label of the first eigenvector and the labels of the other two eigenvectors (the third and the fourth). This means that the first and the second eigenvectors are more similar in terms of risk level, while being more different from the other two eigenvectors. In this way, the model can better learn the hierarchical relationship between risk levels and effectively distinguish features of different risk levels in the feature space. In addition, the calculation of the secure optimal margin is also based on the shortest path length in the risk level tree, further optimizing the distance relationship between eigenvectors.
[0067] Preferably, the implementation process of the four-factor group loss function for security can be refined through the following steps. First, randomly select four feature vectors and their corresponding security labels from the feature vectors output by the conversion layer of the feature conversion model to form a four-factor group vector set. Then, based on the risk level tree, calculate the shortest path length between the labels of these four feature vectors, and filter out the valid four-factor group vector set that meets the conditions. For example, a threshold can be set, and only when the shortest path length between the labels of the first feature vector and the second feature vector is less than this threshold, are they considered similar. Next, calculate the security optimal margin of the valid four-factor group vector set, which reflects the ideal distance between feature vectors in the feature space. Finally, by optimizing this margin, adjust the parameters of the model so that feature vectors with similar risk levels are more clustered in the feature space, while feature vectors with dissimilar risk levels are more dispersed. In practical applications, the performance of the model can be further improved by adjusting the weight parameters of the loss function to balance the distance optimization objectives between different feature vectors.
[0068] In some embodiments, traversing the four-factor group vector set and obtaining a valid security four-factor group vector set based on the risk level tree includes: the shortest path length between the label of the first feature vector and the label of the second feature vector in the four-factor group vector set in the risk level tree is less than the shortest path length between the label of the first feature vector and the label of the third feature vector in the risk level tree, and less than the shortest path length between the label of the first feature vector and the label of the fourth feature vector in the risk level tree.
[0069] It should be noted that the screening condition of the four-factor group vector set mentioned in the present invention is determined based on the shortest path length of the risk level tree. This condition ensures that in the feature space, feature vectors with similar risk levels can be correctly identified and distinguished. Specifically, by comparing the shortest path lengths between the corresponding labels of different feature vectors, feature vector combinations with similar or dissimilar relationships in the risk level tree can be filtered out. This method can effectively improve the ability of the model to distinguish risk features, enabling the model to better learn the hierarchical relationship between risk levels during the training process, thereby improving the accuracy and reliability of security supervision.
[0070] Specifically, the screening criteria for the four-element group vector set involve the key concept of the shortest path length in the risk level tree. The shortest path length refers to the length of the shortest path connecting any two labels in the risk level tree through the hierarchical structure of the tree. For example, if two labels are on the same branch and adjacent in the tree, their shortest path length is shorter, indicating that they are relatively similar in risk level; while if two labels are on different branches, their shortest path length is longer, indicating that they are quite different in risk level. When screening the four-element group vector set, it is necessary to compare the shortest path lengths between the labels of the first feature vector and the second, third, and fourth feature vectors. Only when the shortest path length between the labels of the first feature vector and the second feature vector is less than its shortest path lengths with the labels of the third and fourth feature vectors, the four-element group vector set is considered valid. This condition ensures that in the feature space, feature vectors with similar risk levels can be correctly clustered together, while feature vectors with dissimilar risk levels can be effectively separated.
[0071] Preferably, the process of screening the four-element group vector set can be further refined through the following steps. First, randomly select four feature vectors and their corresponding security labels from the feature vectors output by the conversion layer of the feature conversion model to form a four-element group vector set. Then, based on the risk level tree, calculate the shortest path lengths between the labels of these four feature vectors. For example, the breadth-first search algorithm can be used to calculate the shortest path length between any two labels. Next, according to the above conditions, compare the shortest path lengths between the label of the first feature vector and the labels of the second, third, and fourth feature vectors, and screen out the valid four-element group vector sets that meet the conditions. If a four-element group vector set does not meet the conditions, it will be discarded and a new set of feature vectors will be selected for calculation. In this way, it can be ensured that the screened four-element group vector set has a clear similarity or dissimilarity relationship in the risk level tree, thereby providing effective input for the calculation of the subsequent security four-element group loss function. This process not only optimizes the training process of the model but also improves the model's ability to distinguish risk features.
[0072] In some embodiments, the following formula is used to calculate the security optimal margin of the valid security four-element group vector set:
[0073]
[0074] Wherein, represents the label of the first feature vector in the s-th valid security four-element group vector set; represents the label of the third feature vector in the s-th valid security four-element group vector set; represents the label of the second feature vector in the s-th valid security four-element group vector set; The label representing the fourth eigenvector in the s-th effective set of safety four-factor group vectors; Ψ() represents the shortest path length between two labels in the risk level tree.
[0075] It should be noted that the calculation method of the safety optimal margin mentioned in the present invention is implemented based on the shortest path length of the labels in the risk level tree. The core of this method is to provide an optimization objective for the safety four-factor group loss function by quantifying the hierarchical relationship between the labels. The calculation result of the safety optimal margin reflects the ideal distance relationship between the eigenvectors of different risk levels in the feature space, thereby guiding the optimization of the distribution of eigenvectors during the model training process. In this way, the model can better learn the differences between risk levels and improve the ability to identify and classify the safety risks at the operation site.
[0076] Specifically, the calculation of the safety optimal margin involves the key concept of the shortest path length of the labels in the risk level tree. The shortest path length refers to the length of the shortest path connecting any two labels through the hierarchical structure of the tree in the risk level tree. For example, if two labels are in the same branch and adjacent in the tree, their shortest path length is shorter, indicating that they are relatively similar in risk level; while if two labels are in different branches, their shortest path length is longer, indicating that they have a greater difference in risk level. When calculating the safety optimal margin, it is necessary to consider the shortest path lengths between the labels of the first, second, third, and fourth eigenvectors in the four-factor group vector set. Specifically, the safety optimal margin can be determined by comparing the shortest path lengths between the labels of the first eigenvector and the second eigenvector, as well as the shortest path lengths between the first eigenvector and the labels of the third and fourth eigenvectors. This margin value reflects the ideal distance that should be maintained between the eigenvectors of similar risk levels and the minimum distance that should be maintained between the eigenvectors of dissimilar risk levels in the feature space.
[0077] Preferably, the calculation process of the safety optimal interval can be further refined through the following steps. First, obtain the labels of each feature vector in the four-element group vector set from the risk level tree. Then, calculate the shortest path length between these labels. For example, the breadth-first search algorithm can be used to achieve this. Next, determine the safety optimal interval according to the calculated shortest path length. For example, a benchmark value can be set such that the difference between the shortest path length between the labels of the first feature vector and the second feature vector and the benchmark value, plus the difference between the shortest path length between the labels of the first feature vector and the third and fourth feature vectors and the benchmark value, is used as the calculation result of the safety optimal interval. This interval value can be used as an important parameter in the loss function to guide the model to optimize the distance relationship between feature vectors during the training process. For example, during model training, the model parameters can be adjusted to make the actual distance between feature vectors as close as possible to the calculated safety optimal interval, thereby improving the model's ability to distinguish risk features.
[0078] In some embodiments, constructing the risk level tree based on the risk level relationships of each safety label includes: for each safety label, look up its parent class in the risk database and search layer by layer upward until reaching the root node of the risk database to obtain the risk level relationship of each label;
[0079] Based on the risk level relationships of each label, construct the risk level tree T = {V, E}; where V represents the set of all nodes of the risk level tree, which are the label classes and their parent classes in the risk database; E represents the node relationships of the risk level tree.
[0080] It should be noted that constructing the risk level tree based on the risk level relationships of each safety label in the present invention is an important basis for the efficient operation of the safety supervision model. The risk level tree is a hierarchical data structure used to represent the risk level relationships between different safety labels. Through this structure, the hierarchy and association of each safety label in terms of risk level can be clearly displayed, thereby providing richer semantic information for the model and enabling it to better understand and process the safety risks at the job site. The process of constructing the risk level tree involves the analysis and arrangement of the risk level relationships of each safety label, and this process needs to be completed in combination with the information in the risk database to ensure that the risk level tree can accurately reflect the actual safety risk system.
[0081] Specifically, the construction process of the risk level tree involves several key concepts. First, a safety label refers to the safety risk categories labeled in the data at the operation site, such as not wearing a safety helmet and abnormal equipment operation. The risk level relationship refers to the hierarchical relationship between these safety labels. For example, not wearing a safety helmet may belong to a higher-level risk category of insufficient personal protection. The risk database is a database that stores various safety labels and their risk level relationships, providing the basic data for the construction of the risk level tree. When constructing the risk level tree, for each safety label, it is necessary to find its parent category, that is, the higher-level risk category, in the risk database and search upward layer by layer until the root node of the risk database, that is, the highest-level risk category, is reached. For example, for the safety label of not wearing a safety helmet, its parent category may be insufficient personal protection, and the parent category of insufficient personal protection may be the safety risk at the operation site. In this way, a complete risk level tree can be constructed, where each node represents a risk category, and the relationship between nodes represents the level of risk.
[0082] Preferably, the construction process of the risk level tree can be further refined through the following steps. First, extract all safety labels and their corresponding parent category information from the risk database. Then, based on this information, construct an initial risk level tree, where each safety label serves as a node, and the relationship between nodes is determined according to the parent category information. Next, optimize the risk level tree, such as by merging nodes with the same parent category or adjusting the hierarchical relationship of nodes to better reflect the actual risk level system. In the model, the risk level tree can be stored in the form of a graph, where each node stores the information of the safety label, and the edges represent the hierarchical relationship between nodes. In the safety supervision model, the risk level tree can be an important part of the model, providing prior knowledge of the risk level relationship for the model. For example, in the supervision layer of the model, according to the structure of the risk level tree, the features of different risk levels can be weighted, thereby increasing the attention of the model to the features of high risk levels.
[0083] In some embodiments, in the safety supervision model, the extracted feature information passes through the supervision layer, and based on the extracted feature information, risk classification is performed on each risk point in the data to obtain a safety supervision result; wherein, the feature information undergoes a feature transformation operation on the feature map through a feature transformation module to obtain the transformed feature information; wherein, |V| represents the number of nodes in the risk level tree;
[0084] The transformed feature information passes through the Gaussian Error Linear Unit (GELU) activation function to perform a non-linear transformation on each feature, obtaining the non-linearly transformed feature information;
[0085] The feature information after the non-linear transformation passes through a feature transformation module again, and then passes through the sigmoid function to obtain the probability distribution of the risk for each risk point.
[0086] It should be noted that the safety supervision model and the feature transformation model mentioned in the present invention adopt a series of operations such as feature transformation and non-linear activation when processing the data at the job site to achieve accurate classification of risk points and effective mapping of feature vectors. The core of this process lies in optimizing the extracted feature information through the feature transformation module and the activation function, thereby enhancing the model's ability to identify complex risk features. The role of the feature transformation module is to further process the feature information output by the feature extraction layer to make it more suitable for subsequent risk classification and feature mapping operations. The non-linear activation function introduces non-linear factors into the model, enabling the model to better capture the complex relationships between features, thereby improving the performance of the model.
[0087] Specifically, the feature transformation module and the non-linear activation function play important roles in the safety supervision model and the feature transformation model. The role of the feature transformation module is to further process the feature information output by the feature extraction layer, for example, through convolution operations, pooling operations or other linear transformations, to convert the feature information into a form more suitable for subsequent processing. The non-linear activation function such as the Gaussian Error Linear Unit (GELU) is used to perform non-linear transformation on each feature, enabling the model to capture the complex relationships between features. The GELU activation function is a smooth non-linear function that combines linear and non-linear characteristics, which can effectively avoid the problem of gradient disappearance while maintaining the non-linear expression ability of the model. In the feature transformation model, the role of the feature transformation kernel is to map the feature information in the data to a new space for better feature recombination and optimization. This process, through the combination of non-linear transformation and the activation function, makes the feature vectors with similar risk levels close to each other in the feature space, while the feature vectors with dissimilar risk levels move away from each other.
[0088] Preferably, the operation steps of feature transformation and non-linear activation can be further refined as follows. In the security supervision model, the feature information extracted by the feature extraction layer first passes through a feature transformation module, which can adopt the structure of a multi-layer perceptron (MLP) or a convolutional neural network (CNN) to further process the feature information. The processed feature information undergoes a non-linear transformation through the GELU activation function to enhance the model's ability to capture complex relationships between features. Then, the feature information after the non-linear transformation passes through a feature transformation module again, and finally, the risk probability distribution of each risk point is obtained through the sigmoid function. In the feature transformation model, the feature information extracted by the feature extraction layer passes through a transformation layer, which can adopt technologies such as autoencoders or attention mechanisms to map the feature information to a new space. The mapped feature information also undergoes a non-linear transformation through the GELU activation function, and then is optimized through the feature transformation kernel again. Finally, the feature vector mapped to the new space is obtained through the sigmoid function. This process not only optimizes the representation of feature information but also improves the model's ability to distinguish different risk features.
[0089] In some embodiments, in the feature transformation model, the extracted feature information passes through a transformation layer to map the feature information in the data to a new space; wherein, the feature information passes through a feature transformation kernel to perform a feature transformation operation on the feature map to obtain the transformed feature information;
[0090] The transformed feature information passes through the Gaussian error linear unit GELU activation function to perform a non-linear transformation on each feature to obtain the feature information after the non-linear transformation;
[0091] The feature information after the non-linear transformation passes through a feature transformation kernel again and then through the sigmoid function to obtain the feature vector mapped to the new space.
[0092] It should be noted that the feature transformation model mentioned in the present invention processes the feature information through the feature transformation kernel and the non-linear activation function to achieve the effective mapping and optimization of the feature vector. The core of this process is to map the feature information to a new space through the feature transformation kernel, and at the same time use the non-linear activation function to perform a non-linear transformation on the features, thereby enhancing the model's expression ability and discrimination ability for features. Finally, the feature vector mapped to the new space is obtained through the sigmoid function. These feature vectors can better reflect the differences between different risk levels in the feature space and provide a more effective feature representation for subsequent security supervision.
[0093] Specifically, the feature transformation kernel and the non-linear activation function play a key role in the feature transformation model. The feature transformation kernel is a function or module for feature mapping, which can transform the feature information in the original feature space to a new feature space, making the features more separable in the new space. For example, the feature transformation kernel can be a multi-layer perceptron (MLP) or a convolutional kernel for linear or non-linear transformation of features. Non-linear activation functions such as the Gaussian Error Linear Unit (GELU) are used to introduce non-linearity, enabling the model to capture complex relationships between features. The GELU activation function is a smooth non-linear function that combines linear and non-linear characteristics, effectively avoiding the vanishing gradient problem while maintaining the non-linear representation ability of the model. In the feature transformation model, the feature transformation kernel first performs a mapping operation on the feature information extracted by the feature extraction layer, and then performs a non-linear transformation through the GELU activation function to obtain the non-linearly transformed feature information. These feature information are further optimized through the feature transformation kernel, and finally the feature vectors mapped to the new space are obtained through the sigmoid function.
[0094] Preferably, the construction of the feature transformation model and the feature processing process can be further refined through the following steps. First, the feature information extracted by the feature extraction layer is input into the feature transformation kernel. The feature transformation kernel can adopt the structure of a multi-layer perceptron (MLP), which contains multiple fully connected layers and activation functions for layer-by-layer transformation of features. For example, an MLP with two layers can be set, and the number of neurons in each layer can be adjusted according to the complexity of the actual data. The output of the first layer is non-linearly transformed through the GELU activation function to enhance the expression ability of the features. Then, the non-linearly transformed feature information is input into the feature transformation kernel again for further optimization. This process can be optimized by adjusting the parameters of the MLP or introducing regularization techniques. Finally, the feature information is mapped to the interval [0, 1] through the sigmoid function to obtain the feature vectors mapped to the new space. These feature vectors can better reflect the differences between different risk levels in the feature space, providing a more effective feature representation for the subsequent safety supervision model.
[0095] The above-mentioned embodiments of the present invention have the following beneficial effects: This method can improve the intelligent level of safety supervision at the operation site. By constructing a risk level tree and jointly training a safety supervision model and a feature transformation model, the hierarchical relationship between tags can be fully utilized to optimize the risk identification effect. Among them, the shared design of the feature extraction layer can reduce the consumption of computing resources. The supervision mechanism based on the nodes of the risk level tree in the supervision layer can enhance the accuracy of risk classification. The feature recombination function of the transformation layer can aggregate similar risk features in space and separate different features, thereby improving the discriminant ability of the model. Through the joint optimization of the safety maximum-minimum loss function and the safety four-factor group loss function, the prediction accuracy of risk probability and the quality of feature space distribution can be improved simultaneously, enabling the model to have stronger generalization performance.
[0096] This method can also optimize the end-to-end risk supervision process, forming a complete closed-loop from feature extraction to risk classification. Through the probability transformation of the parent node and child node sets in the risk level tree, the correlation between risk levels can be more accurately reflected. By dynamically calculating the four-factor group loss of the safety optimal margin, the sample spacing in the feature space can be adaptively adjusted. The combination of the GELU activation function and the sigmoid function in the feature transformation model can enhance the non-linear feature expression ability, and the shortest path calculation mechanism of the risk level tree can ensure that the feature space mapping conforms to the actual risk association logic. These technologies work together to finally construct a safety supervision system with high precision, strong robustness and interpretability.
[0097] Furthermore, the storage medium of the embodiment of the present application stores program instructions that can implement all the above methods. Among them, the program instructions can be stored in the above storage medium in the form of a software product, including several instructions to enable a computer device (which can be a personal computer, a server, or a network device, etc.) or a processor to execute all or part of the steps of the methods described in various embodiments of the present application. The aforementioned storage medium includes: various media that can store program codes such as USB flash drives, mobile hard disks, read-only memories (ROM, Read-Only Memory), random access memories (RAM, Random Access Memory), magnetic disks or optical discs, or terminal devices such as computers, servers, mobile phones, and tablets.
[0098] The above description is only some preferred embodiments of the present invention and an explanation of the technical principles applied. Those skilled in the art should understand that the scope of the invention involved in the embodiments of the present invention is not limited to the technical solutions formed by the specific combination of the above technical features. At the same time, it should also cover other technical solutions formed by any combination of the above technical features or their equivalent features without departing from the above inventive concept. For example, the technical solutions formed by mutually replacing the above features with the technical features (but not limited to) having similar functions disclosed in the embodiments of the present invention.
Claims
1. A method for safety supervision of the operation site based on AI vision, characterized in that, Including: Obtain a dataset of the operation site and the safety labels corresponding to each data in the dataset, and construct a risk level tree based on the risk level relationships of the safety labels; Based on the operation site dataset, the safety labels corresponding to each data in the dataset, and the risk level tree, simultaneously train the safety supervision model and the feature transformation model to obtain a trained safety supervision model; wherein, the safety supervision model includes a feature extraction layer shared with the feature transformation model for extracting feature information of the input data; The safety supervision model further includes a supervision layer for performing safety supervision on the input data based on the number of nodes of the risk level tree; during training, transform the supervised risk probability values based on the risk level relationships of the safety labels in the risk level tree; The feature transformation model further includes a transformation layer for reorganizing the features of the input data to obtain a new feature vector; during training, based on the similarity information of the labels in the risk level tree, make the new feature vectors of similar risk levels aggregate with each other in the feature space, and the new feature vectors of dissimilar risk levels diverge from each other.
2. The method according to claim 1, characterized in that, The simultaneous training of the safety supervision model and the feature transformation model to obtain a trained safety supervision model includes: loading the operation site dataset, the safety labels corresponding to each data, and the risk level tree, simultaneously training the safety supervision model and the feature transformation model using the safety max-min loss function and the safety four-element group loss function, updating the parameters of the feature extraction layer, the supervision layer, and the transformation layer using gradient backpropagation, saving the parameters of the feature extraction layer, the supervision layer, and the transformation layer after training, and obtaining the safety supervision model based on the trained feature extraction layer and supervision layer.
3. The method according to claim 2, wherein The safety max-min loss function is expressed as: Where B represents the number of data in the job site dataset; M represents the height of the data; N represents the width of the data; |V| represents the number of nodes in the risk level tree; y b,i,j,k represents the label value corresponding to the k-th risk level of the (i, j) risk point in the b-th data; represents the risk probability value after the risk maximum-minimum transformation of the label value corresponding to the k-th risk level of the (i, j) risk point in the b-th data.
4. The method according to claim 3, wherein Use the following formula to obtain the transformed risk probability value of each risk point: Among them, A k represents the set of parent nodes of the risk level k in the risk level tree; D k represents the set of child nodes of the risk level k in the risk level tree; y b,i,j,l represents the label value corresponding to the l-th risk level of the (i, j) risk point in the b-th data; represents the transformed risk probability value corresponding to the k-th risk level of the (i, j) risk point in the b-th data.
5. The method according to claim 2, wherein The safety four-element group loss function is expressed as: Among them, DTT represents the set of effective secure four-element group vectors; N b represents the number of four-element groups in the set of effective secure four-element group vectors; represents the first eigenvector in the s-th set of effective secure four-element group vectors; represents the second eigenvector in the s-th set of effective secure four-element group vectors; represents the third eigenvector in the s-th set of effective secure four-element group vectors; represents the fourth eigenvector in the s-th set of effective secure four-element group vectors; d() represents the vector distance function; m s represents the secure optimal margin of the s-th set of effective secure four-element group vectors; Optionally, select four new feature vectors output by the transformation layer and the corresponding safety labels to form a four-element group vector set; Traverse the four-element group vector set to obtain an effective safety four-element group vector set based on the risk level tree.
6. The method according to claim 5, wherein The traversing the four-element group vector set to obtain an effective safety four-element group vector set based on the risk level tree includes: the shortest path length between the label of the first feature vector and the label of the second feature vector in the four-element group vector set in the risk level tree is less than the shortest path length between the label of the first feature vector and the label of the third feature vector in the risk level tree, and less than the shortest path length between the label of the first feature vector and the label of the fourth feature vector in the risk level tree.
7. The method according to claim 6, wherein Use the following formula to calculate the safety optimal margin of the effective safety four-element group vector set: Among them, represents the label of the first eigenvector in the s-th effective secure four-factor group vector set; represents the label of the third eigenvector in the s-th effective secure four-factor group vector set; represents the label of the second eigenvector in the s-th effective secure four-factor group vector set; represents the label of the fourth eigenvector in the s-th effective secure four-factor group vector set; Ψ() represents the shortest path length between two labels in the risk level tree.
8. The method according to any one of claims 1-7, characterized in that, The constructing the risk level tree based on the risk level relationships of the safety labels includes: for each safety label, search for its parent class in the risk database, and search layer by layer upward until reaching the root node of the risk database to obtain the risk level relationship of each label; Based on the risk level relationship of each tag, a risk level tree T = {V, E} is constructed; where V represents the set of each node of the risk level tree, which is the tag class and its parent class in the risk database; E represents the node relationship of the risk level tree.
9. The method according to claim 8, wherein In the safety supervision model, the extracted feature information passes through the supervision layer, and risk classification is performed on each risk point in the data based on the extracted feature information to obtain a safety supervision result; where the feature information passes through a feature transformation module to perform a feature transformation operation on the feature map to obtain the transformed feature information; where |V| represents the number of nodes of the risk level tree. The transformed feature information passes through the Gaussian Error Linear Unit (GELU) activation function to perform a non-linear transformation on each feature to obtain the non-linearly transformed feature information. The non-linearly transformed feature information passes through a feature transformation module again, and then through the sigmoid function to obtain the probability distribution of the risk of each risk point.
10. The method according to claim 8, wherein In the feature conversion model, the extracted feature information passes through the conversion layer to map the feature information in the data to a new space; where the feature information passes through a feature conversion kernel to perform a feature conversion operation on the feature map to obtain the transformed feature information. The transformed feature information passes through the Gaussian Error Linear Unit (GELU) activation function to perform a non-linear transformation on each feature to obtain the non-linearly transformed feature information. The non-linearly transformed feature information passes through a feature conversion kernel again, and then through the sigmoid function to obtain the feature vector mapped to the new space.