VOC leakage identification neural network model based on maximum attention mechanism
By constructing a VOC leakage recognition neural network model based on the maximum attention mechanism, the problem of accurate positioning and high computational complexity of VOC leakage detection in the prior art is solved, and rapid response and efficient identification are achieved in the industrial environment.
Patent Information
- Application Number
- CN202510362682.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-03-26
- Publication Date
- 2025-08-08
AI Technical Summary
The existing VOC leak detection methods are difficult to achieve large-area coverage and precise location of leakage sources in industrial environments, and the calculation complexity is high, which cannot meet the needs of rapid response and precise governance.
A VOC leakage recognition neural network model based on the maximum attention mechanism is designed. By building the maximum attention mechanism module, a multi-scale domain transformation method and a feature regression module, the computing efficiency is optimized and feature extraction capabilities are improved to achieve rapid response and precise positioning.
In industrial scenarios with resource-constrained, efficient and accurate VOC leak identification and positioning are achieved, with excellent real-time and robustness.
Smart Images
Figure FT_1 
Figure FT_2 
Figure SMS_1
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of volatile organic compound (VOC) leakage detection. A VOC leakage recognition neural network model based on the maximum attention mechanism is designed to efficiently and accurately identify and locate VOC leakage sources in industrial environments. Background Art
[0002] Volatile organic compounds (VOCs) are common pollutants in industrial production. Their leaks not only cause serious environmental pollution but can also cause safety incidents such as fires and explosions, endangering human health. Traditional VOC leak detection methods primarily rely on sensor arrays, detecting leaks by monitoring changes in gas concentration. However, these methods have numerous limitations. First, the sensors' limited detection range makes it difficult to cover large areas, and they are susceptible to environmental interference, resulting in low detection accuracy. Second, traditional methods cannot precisely locate the leak source and can only roughly determine the leak area, making it difficult to meet the requirements for rapid response and precise remediation. In recent years, with the development of artificial intelligence technology, deep learning-based leak detection methods have gradually gained attention. Deep learning models can automatically learn features from large amounts of data, improving detection accuracy. However, existing deep learning models still have some problems when handling VOC leak detection tasks. For example, the models' feature extraction of input data is insufficient, failing to effectively capture key information. Furthermore, the models' high computational complexity makes it difficult to respond quickly in real-time monitoring scenarios. To address these issues, this paper proposes a neural network model for VOC leak identification based on a maximum attention mechanism. By introducing an attention mechanism and multi-scale feature extraction, the model automatically focuses on key features in the data, improving the model's sensitivity to VOC leakage signals. Furthermore, the model's design optimizes computational efficiency, enabling rapid response while maintaining detection accuracy.
[0003] In summary, the present invention aims to provide an efficient, accurate, and adaptable neural network model for VOC leak identification to meet the stringent requirements for VOC leak detection in industrial environments and provide strong technical support for environmental protection and safe production. Summary of the Invention
[0004] This paper designs a neural network model for VOC leak identification based on a maximum attention mechanism. This model is implemented through three steps: constructing a maximum attention mechanism module, designing a multi-scale domain transformation method, and building a feature regression module. It can efficiently and accurately identify VOC leaks and assist in inferring leak sources, demonstrating excellent real-time performance, accuracy, and robustness in resource-constrained industrial scenarios.
[0005] The present invention is achieved through the following technical solution, including the following steps:
[0006] Step 1: Build the maximum attention mechanism module;
[0007] Step 2: Design a multi-scale domain transformation method;
[0008] Step 3: Build a feature regression module.
[0009] The creativity of the present invention is mainly reflected in:
[0010] (1) Compared with most Transformer-based models, the maximum attention mechanism designed in this paper can extract global information with a smaller model size and lower computational complexity, and is more suitable for detection and recognition tasks that have strict requirements on computational efficiency.
[0011] (2) Inspired by the perceptual characteristics of the brain with multi-scale adaptive domain transformation, the present invention designs a multi-scale domain transformation method based on this. By downsampling the input image to different sizes, performing domain transformation and then extracting features, the network can better capture the leakage characteristics of VOCs in the input image, thereby assisting in inferring the leakage source. BRIEF DESCRIPTION OF THE DRAWINGS
[0012] Figure 1 This is a flowchart of the sparse ViT module and maximum attention mechanism designed by the present invention.
[0013] Figure 2 This is a flow chart of the VOC leak recognition neural network model based on the maximum attention mechanism designed by the present invention. DETAILED DESCRIPTION
[0014] The following is a detailed description of an embodiment of the present invention. This embodiment is implemented based on the technical solution of the present invention, and provides a detailed implementation method and specific operation process. However, the protection scope of the present invention is not limited to the following embodiment.
[0015] Example:
[0016] Step 1: Build the maximum attention mechanism module;
[0017] Considering that the attention mechanism in the Visual Transformer (ViT) has good global modeling capabilities, this paper designs a sparse ViT module, the core of which is to construct a new maximum attention mechanism module, such as Figure 1 As shown in Figure 2, the maximum attention mechanism module is implemented through the following four steps:
[0018] First, referring to the standard ViT, the input feature map is constructed into a flattened image block vector and category token through the two operations of "image block embedding" and "position embedding", and both are sent to the "layer normalization" network layer to obtain the preprocessed sequence A.
[0019] Secondly, pass A through a "linear mapping" network layer to obtain A's three feature vectors Q, K, and V:
[0020] Q=AW q
[0021] K=AW k
[0022] V=AW v
[0023] Among them, Q, K, and V are the query vector, key vector, and value vector of A respectively, and W q 、W k 、W v are the linear mapping weights corresponding to the above three vectors.
[0024] Then, the correlation between each image block sequence and other image block sequences is calculated to obtain the correlation matrix Matrix cc , then Matrix cc The Max(·) operation is used to keep only one sequence with the highest correlation with each image block sequence, and the index Idx of the sequence with the highest correlation is recorded. Then, the above K and V are pruned according to the index Idx to obtain the pruned key vector K Idx Sum value vector V Idx .
[0025] Finally, the Q vector of each image block sequence is compared with its pruned key vector K Idx , value vector V Idx Performing matrix operations, we can get the output of the maximum attention mechanism:
[0026]
[0027] Among them, “Softmax(·)” is the normalized exponential function, is the matrix transpose operation, S is the scale factor, and the output out of the maximum attention mechanism passes through a "layer normalization" network layer and two residual connections to obtain the output F of the sparse ViT module.
[0028] Step 2: Design a multi-scale domain transformation method;
[0029] Inspired by the brain's ability to perceive features using multi-scale adaptive domain transformation, the present invention designs a multi-scale domain transformation method based on this. Figure 2 shown.
[0030] First, the input image is transformed from the original RGB domain to the channel domain by a 1×1 convolution.
[0031] Then, the features transformed into the channel domain are subjected to Maxpooling operations with different step sizes l to obtain feature maps of different sizes. In the present invention, l = {0, 2, 4}, that is, a total of 3 sizes are obtained. The present invention designs a network path for the feature map of each size.
[0032] Finally, the feature maps of the three sizes are passed through the sparse ViT module designed in the first step to obtain three outputs F0, F2, and F4, which are merged through the concatenation operation to generate the feature vector F cat .
[0033] Step 3: Build feature regression module;
[0034] In order to obtain accurate recognition results, the present invention builds a feature regression module, such as Figure 2 As shown. This module consists of two fully connected layers, which transform the above feature vector F cat Mapping to the final recognition result.
Claims
1. Construct a maximum attention mechanism module; Considering that the attention mechanism in the Visual Transformer (ViT) has good global modeling capabilities, this paper designs a sparse ViT module, the core of which is to build a new maximum attention mechanism module. The maximum attention mechanism module is mainly implemented through the following four steps: First, referring to the standard ViT, the input feature map is used to construct a flattened image block vector and category token through the "image block embedding" and "position embedding" operations. Both are sent to the "layer normalization" network layer to obtain the preprocessed sequence A. Secondly, pass A through a "linear mapping" network layer to obtain A's three feature vectors Q, K, and V: Q=AW q K=AW k V=AW v in, Q, K, and V are the query vector, key vector, and value vector of A respectively. W q 、W k 、W v are the linear mapping weights corresponding to the above three vectors. Then, the correlation between each image block sequence and other image block sequences is calculated to obtain the correlation matrix Matrix cc , then Matrix cc The Max(·) operation is used to keep only one sequence with the highest correlation with each image block sequence, and the index Idx of the sequence with the highest correlation is recorded. Then, the above K and V are pruned according to the index Idx to obtain the pruned key vector K Idx Sum value vector V Idx . Finally, the Q vector of each image block sequence is compared with its pruned key vector K Idx , value vector V Idx Performing matrix operations, we can get the output of the maximum attention mechanism: Among them, "Softmax(·)" is the normalized exponential function, is the matrix transpose operation, S is the scale factor, and the output out of the maximum attention mechanism passes through a "layer normalization" network layer and two residual connections to obtain the output F of the sparse ViT module.
2. Design multi-scale domain transformation methods; Inspired by the brain's feature perception ability of multi-scale adaptive domain transformation, the present invention designs a multi-scale domain transformation method based on this: first, the input image is subjected to a 1×1 convolution to transform the image from the original RGB domain to the channel domain; then, the features transformed into the channel domain are subjected to Maxpooling operations with different step sizes l to obtain feature maps of different sizes. In the present invention, l = {0, 2, 4}, that is, a total of three sizes are obtained. The present invention designs a network path for the feature map of each size; finally, the feature maps of the above three sizes are respectively subjected to the sparse ViT module designed in the first step to obtain three outputs F0, F2, and F4, which are merged through the concatenation operation to generate a feature vector F cat .
3. Build a feature regression module; In order to obtain accurate recognition results, the present invention builds a feature regression module, which consists of two fully connected layers. cat Mapping to the final recognition result.