Garbage discharge recognition method and system applying deep learning model

Abstract

The application relates to the technical field of artificial intelligence, and provides a garbage emission identification method and system applying a deep learning model. The method comprises the following steps: firstly, collecting a flue gas image data set under different collection time periods and different garbage combustion conditions in a garbage treatment area; then, performing image enhancement processing to generate pretreated flue gas image data which retains spectral distribution characteristics and spatial texture details; then, calling a double-channel convolution attention network to perform feature mining, so as to obtain a flue gas spectral feature set containing spectral response intensity distribution in different wavelength intervals and spatial dimension feature response saliency distribution; finally, performing difference feature identification processing based on multi-scale residual connection to generate difference description features of emission substances, and inputting the difference description features and the flue gas spectral feature set into a classification network which is optimized through migration learning, so as to obtain heterogeneous emission substance identification labels indicating different types of pollutant components in garbage emission flue gas and relative content proportion relationships.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of artificial intelligence, and particularly relates to a garbage discharge identification method and system applying a deep learning model. BACKGROUND

[0002] In the garbage disposal process, in order to realize environmental monitoring and management, it is necessary to accurately identify the pollutant components and content proportions in the garbage discharge flue gas. Some existing technologies mainly rely on traditional chemical analysis methods, which determine the pollutant components and content by collecting flue gas samples and performing laboratory chemical analysis. Although this method can provide relatively accurate results, it has the problems of complex sampling process, long analysis period and high cost, and cannot monitor the dynamic changes of the garbage discharge flue gas in real time.

[0003] With the development of computer vision and machine learning technologies, some image recognition-based methods have been gradually applied to garbage discharge flue gas pollutant identification. These methods collect flue gas images, use image processing and machine learning algorithms for feature extraction and classification identification. However, the existing image recognition methods have some limitations: in the image collection stage, the influence of different collection time periods and garbage combustion conditions on the features of the flue gas images is not fully considered, resulting in a lack of comprehensiveness and representativeness of the collected data; in the feature extraction and classification identification process, the existing methods are difficult to effectively capture the spectral and spatial features in the flue gas images, and the identification accuracy and reliability of the pollutant components and content proportions are low, which cannot meet the needs of actual environmental monitoring and management. SUMMARY

[0004] The application provides a garbage discharge identification method and system applying a deep learning model.

[0005] In a first aspect, the application embodiment provides a garbage discharge identification method applying a deep learning model, which is applied to a garbage discharge identification system applying a deep learning model, and the method comprises:

[0006] Collecting a flue gas image data set of a garbage disposal area, wherein the flue gas image data set contains multiple groups of flue gas image units under different collection time periods, and each group of flue gas image units corresponds to different garbage combustion conditions;

[0007] Performing image enhancement processing on the flue gas image data set to generate preprocessed flue gas image data with spectral feature enhancement effect, wherein the preprocessed flue gas image data retains the spectral distribution characteristics and spatial texture details in the original flue gas image units;

[0008] The double-channel convolution attention network is called to perform a feature mining operation on the preprocessed flue gas image data, and through parallel processing of a spectral feature extraction branch and a spatial feature enhancement branch, a flue gas spectral feature set is obtained, which contains spectral response intensity distribution in different wavelength intervals and spatial dimension feature response significance distribution.

[0009] Based on multi-scale residual connection, difference feature identification processing is performed on the flue gas spectral feature set to generate emission substance difference description features corresponding to the garbage emission flue gas image. The flue gas spectral feature set and the emission substance difference description features are input into a classification network optimized by transfer learning, and through feature correlation modeling and classification probability prediction, heterogeneous emission substance identification labels are obtained, which are used to indicate different types of pollutant components contained in the garbage emission flue gas and the relative content proportion relationship of the pollutant components.

[0010] In a second aspect, the embodiments of the present application provide a garbage emission identification system applying a deep learning model, which includes a processor and a memory, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of the above method.

[0011] In a third aspect, the embodiments of the present application provide a computer readable storage medium, which includes a computer program, and when the computer program runs on a garbage emission identification system applying a deep learning model, the computer program is used to make the garbage emission identification system applying a deep learning model execute the steps of the above method.

[0012] By applying the embodiment of the present application, the accuracy and reliability of identifying heterogeneous emission substances in waste emission flue gas can be improved. First, a set of flue gas image data is comprehensively collected in different time periods and under different combustion conditions in the waste treatment area. The coverage of different collection periods and combustion conditions enables the data to fully reflect the dynamic changes and diversity of pollutant emissions during waste combustion. Then, the set of flue gas image data is subjected to image enhancement processing to generate preprocessed flue gas image data with enhanced spectral features, which not only retains the spectral distribution characteristics and spatial texture details of the original image but also strengthens the spectral features, enabling feature mining to more accurately capture key information in the flue gas. The enhancement of spectral features can highlight the unique performance of different pollutants in the spectrum, thereby enabling accurate differentiation and identification of pollutants. Then, a double-channel convolution attention network is called to mine features from the preprocessed flue gas image data. Through parallel processing of the spectral feature extraction branch and the spatial feature enhancement branch, the features of the flue gas image can be fully mined from both the spectral and spatial dimensions. The spectral feature extraction branch focuses on the spectral response intensity distribution in different wavelength intervals, enabling accurate identification of the spectral features of different pollutants; the spatial feature enhancement branch highlights the feature response significance distribution in the spatial dimension, which helps to determine the distribution of pollutants in space, and parallel processing of the two branches greatly improves the efficiency and comprehensiveness of feature mining. Finally, based on multi-scale residual connection, difference feature identification processing is performed on the set of flue gas spectral features to generate emission substance difference description features, which are input into a classification network optimized by transfer learning together with the set of flue gas spectral features. Multi-scale residual connection can effectively capture feature differences at different scales, making the emission substance difference description features better reflect the essential differences between pollutants. The classification network optimized by transfer learning combines the prior knowledge of the source domain data and the characteristics of the target domain data, and through feature correlation modeling and classification probability prediction, it can accurately obtain the identification labels of heterogeneous emission substances, clearly indicate the composition and relative content ratio of different types of pollutants in waste emission flue gas, and thus improve the accuracy and reliability of pollutant identification. BRIEF DESCRIPTION OF DRAWINGS

[0013] Figure 1 A flowchart of a waste emission identification method using a deep learning model provided by an embodiment of the present application.

[0014] Figure 2 A structural diagram of a waste emission identification system using a deep learning model provided by an embodiment of the present application. DETAILED DESCRIPTION

[0015] In order to make the purposes, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions of the present application will be described clearly and completely below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments described in the present application document, all other embodiments obtained by those of ordinary skill in the art without creative labor fall within the scope of protection of the present application technical solutions.

[0016] Referring to Figure 1 It is a garbage discharge identification method applying a deep learning model provided in the embodiments of the present application. The method can be applied to a garbage discharge identification system applying a deep learning model, and the specific process steps 110-140.

[0017] Step 110: Collecting a flue gas image data set of a garbage disposal area, the flue gas image data set containing multiple groups of flue gas image units under different collection time periods, and each group of flue gas image units corresponding to different garbage combustion conditions.

[0018] The embodiments of the present application focus on the garbage discharge identification scene and perform discharge identification on the garbage disposal area. It can be understood that, in order to comprehensively grasp the flue gas characteristics generated in the garbage combustion process, it is necessary to collect a flue gas image data set of the area. In the garbage disposal process, different collection time periods correspond to different garbage combustion stages, and the combustion conditions of each stage are different. For example, in the initial stage of garbage combustion, the garbage may be in the preheating and drying stage, and the combustion is not sufficient, so the flue gas generated at this time may contain more water vapor and incompletely combusted organic matter; as the combustion proceeds, the main combustion stage is entered, the combustion is relatively intense, and the composition and characteristics of the generated flue gas will change significantly; in the later stage of combustion, the garbage is gradually burned out, and the composition of the flue gas will change again.

[0019] In order to accurately capture the flue gas characteristics under different conditions, image collection needs to be performed at multiple different collection time periods. For example, a timed collection mode can be set to take flue gas images at certain time intervals. At the same time, the stability and accuracy of the collection equipment need to be ensured to obtain clear and reliable flue gas images. The collected flue gas image data set contains multiple groups of flue gas image units, and each group of image units corresponds to a set garbage combustion condition.

[0020] Step 120: Performing image enhancement processing on the flue gas image data set to generate preprocessed flue gas image data with spectral feature enhancement effect, the preprocessed flue gas image data retaining the spectral distribution characteristics and spatial texture details in the original flue gas image units.

[0021] Step 121: Performing non-uniform illumination correction processing on each group of flue gas image units in the flue gas image data set.

[0022] In the waste disposal area, due to the complexity of lighting conditions, the collected flue gas images may have the problem of uneven illumination. Uneven illumination will affect the quality of the image and the feature extraction effect, so it is necessary to perform uneven illumination correction processing on each set of flue gas image units. A statistical analysis-based method can be used to achieve this goal. First, the image is partitioned, and the image is divided into multiple small regions. Then, the average brightness value of each small region is calculated to evaluate the illumination intensity of the region. For regions with too strong illumination, adjust by reducing the brightness value; for regions with too weak illumination, increase the brightness value. In specific operation, the adjustment amplitude can be determined according to the difference between the average brightness of each region and the average brightness of the entire image. In this way, the illumination distribution in the image is more uniform.

[0023] Step 122: Perform adaptive contrast adjustment operation on the corrected flue gas image unit by the regional histogram equalization algorithm to enhance the contrast of the weak spectral signal region in the image.

[0024] After the uneven illumination correction processing, in order to further enhance the features of the image, a regional histogram equalization algorithm can be used to perform adaptive contrast adjustment on the corrected flue gas image unit. The algorithm divides the image into multiple sub-regions, each sub-region is relatively independent for histogram equalization processing, so that the contrast can be adjusted adaptively according to the specific situation of each sub-region. For the weak spectral signal region in the image, the pixel value distribution of these regions may be concentrated, and through histogram equalization, these pixel values can be redistributed, so that the contrast of the region is enhanced. In actual operation, first determine the size and division method of the sub-region, then calculate the histogram of each sub-region. According to the shape and distribution of the histogram, the pixel values are remapped so that the pixel value distribution in the sub-region is more uniform. In this way, the features of the weak spectral signal region are highlighted, making it more clear and visible.

[0025] Step 123: Use a noise removal algorithm based on spatial domain filtering to perform noise suppression processing on the contrast-adjusted flue gas image unit to eliminate random noise introduced during image acquisition.

[0026] After contrast adjustment, there may still be some random noise in the image that interferes with feature extraction and analysis. Therefore, a noise removal algorithm based on spatial filtering is used to suppress noise in the image. Spatial filtering algorithm mainly operates on each pixel and its neighborhood pixels in the image. Common spatial filtering algorithms include mean filtering and median filtering. Mean filtering replaces the value of the current pixel with the average value of the neighborhood pixels, thereby smoothing the image and reducing the impact of noise. Median filtering sorts the values of the neighborhood pixels and takes the middle value as the value of the current pixel. This method has good effect on removing salt and pepper noise. In practical application, appropriate filtering algorithm and filtering window size are selected according to the type and intensity of noise. Through noise suppression processing, the image is more smooth, the interference of noise on subsequent processing is reduced, and the quality of the image is improved.

[0027] Step 124: Convert the flue gas image unit after noise suppression to HSV color space, and perform independent gain adjustment operation on hue channel and saturation channel to strengthen the distinguishability of flue gas spectral features in color space.

[0028] After noise suppression processing, the flue gas image unit is converted from RGB color space to HSV color space. HSV color space is more consistent with human perception of color, which divides color information into hue, saturation and brightness channels. In HSV color space, independent gain adjustment operation is performed on hue channel and saturation channel. For hue channel, adjusting the gain can change the type and distribution of colors, highlighting the differences in color between different types of pollutants. For example, some pollutants may have differentiated color characteristics, and by adjusting the gain of the hue channel, these colors can be made more vivid. For saturation channel, adjusting the gain can enhance the brightness of the color. In the flue gas of garbage discharge, different components of pollutants may have different saturation characteristics, and by adjusting the gain of the saturation channel, the distinguishability of these characteristics can be strengthened. In actual operation, appropriate gain adjustment parameters are determined according to the specific situation of the image and analysis requirements to achieve the purpose of strengthening the distinguishability of flue gas spectral features in color space.

[0029] Step 125: Convert the flue gas image unit after HSV color space adjustment back to RGB color space to obtain preprocessed flue gas image data with spectral feature enhancement effect, which retains the spectral distribution characteristics and spatial texture details in the original flue gas image unit.

[0030] After adjusting the HSV color space, the flue gas image unit is converted back to the RGB color space (the subsequent processing and analysis process is usually more suitable in the RGB color space). During the conversion process, the spectral distribution characteristics and spatial texture details in the original flue gas image unit need to be preserved. By converting the hue, saturation and brightness information in the HSV color space into the red, green and blue channel values in the RGB color space, the pre-processed flue gas image data with enhanced spectral characteristics is obtained. During the conversion process, attention should be paid to the accurate transmission and conversion of color information to avoid information loss or distortion. The final pre-processed flue gas image data not only retains the important features of the original image, but also enhances the spectral characteristics.

[0031] Step 130: calling a dual-channel convolution attention network to perform feature mining operations on the pre-processed flue gas image data, obtaining a flue gas spectral feature set through parallel processing of the spectral feature extraction branch and the spatial feature enhancement branch, the flue gas spectral feature set containing spectral response intensity distribution in different wavelength intervals and spatial dimension feature response significance distribution.

[0032] Step 131: synchronously inputting the pre-processed flue gas image data into the spectral feature extraction branch and the spatial feature enhancement branch of the dual-channel convolution attention network.

[0033] The dual-channel convolution attention network has two parallel branches, namely the spectral feature extraction branch and the spatial feature enhancement branch. The pre-processed flue gas image data is simultaneously input into the two branches for parallel feature mining, so as to fully utilize the spectral information and spatial information of the image and improve the efficiency and accuracy of feature extraction. In the garbage discharge identification scene, the spectral information can reflect the characteristics of different components in the flue gas, while the spatial information can reflect the distribution and morphological characteristics of the flue gas. Through parallel processing, the pre-processed flue gas image data can be analyzed from different angles at the same time, and more comprehensive and accurate feature information can be mined.

[0034] Step 132: in the spectral feature extraction branch, performing feature extraction on the spectral channel of the pre-processed flue gas image data through multi-level convolution operation to generate an initial spectral feature map.

[0035] Step 1321: inputting the pre-processed flue gas image data into the input layer of the spectral feature extraction branch, performing initial feature extraction on the pre-processed flue gas image data through a first convolution block to generate a first-level spectral feature map, the first convolution block containing a convolution layer, a batch normalization layer and an activation function layer connected in turn.

[0036] In the spectral feature extraction branch, the preprocessed flue gas image data is first input into the input layer. The first convolutional block is the starting stage of the entire feature extraction process, which is composed of convolutional layers, batch normalization layers and activation function layers connected in turn. The convolutional layers slide on the image through the convolution kernel, and perform convolution operations on the spectral channels of the preprocessed flue gas image data to extract local features of the image. The size and number of the convolution kernel will affect the type and number of the extracted features. The batch normalization layer normalizes the output of the convolutional layer, making the distribution of the data more stable and accelerating the convergence speed of the model. The activation function layer introduces a non-linear factor to enhance the expression ability of the model. Common activation functions include ReLU functions, etc. Through the processing of the first convolutional block, the first-level spectral feature map is generated, which contains the preliminary spectral features of the preprocessed flue gas image data.

[0037] Step 1322: input the first-level spectral feature map into the second convolutional block, expand the feature channel dimension by increasing the number of convolution kernels, and generate a second-level spectral feature map, the channel number of the second-level spectral feature map is an integer multiple of the channel number of the first-level spectral feature map.

[0038] Then, the first-level spectral feature map is input into the second convolutional block. In order to further enrich the feature information, the feature channel dimension is expanded by increasing the number of convolution kernels. The increase in the number of convolution kernels means that more different types of features can be extracted. The second convolutional block also contains convolutional layers, batch normalization layers and activation function layers. In the convolutional layer, the increased convolution kernel will perform convolution operations on the first-level spectral feature map to extract more feature information. The batch normalization layer and the activation function layer have the same function as in the first convolutional block, which are used for data normalization and introduction of nonlinearity, respectively. After processing, the second-level spectral feature map is generated, and the number of channels is an integer multiple of the number of channels of the first-level spectral feature map, thereby increasing the diversity and richness of the features.

[0039] Step 1323: input the second-level spectral feature map into the third convolutional block, reduce the spatial resolution of the feature map by convolution operation with a step of a preset value, generate a third-level spectral feature map, and the spatial size of the third-level spectral feature map is a preset proportion of the spatial size of the second-level spectral feature map.

[0040] Then, the second-level spectral feature map is input into a third convolutional block, which reduces the spatial resolution of the feature map by setting the stride of the convolution operation to a preset value. The setting of the stride can control the sliding interval of the convolution kernel on the image, thereby reducing the spatial size of the feature map. During the convolution operation, the convolution kernel slides on the second-level spectral feature map according to the preset stride, and the image is convolved. In this way, the spatial size of the generated third-level spectral feature map is a preset proportion of the spatial size of the second-level spectral feature map. The downsampling operation can reduce the computational amount of the feature map while retaining important feature information, so that the model pays more attention to the macro features and global information of the image.

[0041] Step 1324: input the third-level spectral feature map into a fourth convolutional block, capture multi-scale spectral information by combining convolution operations of different receptive fields, and generate a fourth-level spectral feature map.

[0042] The third-level spectral feature map is then input into a fourth convolutional block, which captures multi-scale spectral information by combining convolution operations of different receptive fields. Convolution kernels of different receptive fields can focus on different scale features of the image, and convolution kernels of large receptive fields can capture global features, and convolution kernels of small receptive fields can capture local detail features. In actual operation, a plurality of convolution kernels of different sizes are used to perform convolution operations on the third-level spectral feature map in parallel. Each convolution kernel extracts different scale feature information, which is then fused. In this way, a fourth-level spectral feature map is generated, which contains multi-scale spectral information and can more comprehensively reflect the spectral features of the flue gas.

[0043] Step 1325: take the fourth-level spectral feature map as an initial spectral feature map, and the channel dimension of the initial spectral feature map contains spectral response characteristics of different wavelength intervals.

[0044] Finally, the fourth-level spectral feature map is taken as an initial spectral feature map, and the channel dimension of the initial spectral feature map contains spectral response characteristics of different wavelength intervals. In the garbage discharge identification scene, different pollutants may have different spectral response intensities in different wavelength intervals. Through multi-level convolution operation, the initial spectral feature map extracted from the preprocessed flue gas image data can accurately reflect the spectral response characteristics of these different wavelength intervals.

[0045] Step 133: in the spatial feature enhancement branch, a spatial attention mechanism is used to assign weights to the spatial dimension of the preprocessed flue gas image data, and a spatial attention weight map is generated.

[0046] Step 1331: perform global maximum pooling and global average pooling operations on the channel dimension of the preprocessed flue gas image data to generate two single-channel feature maps.

[0047] In the spatial feature enhancement branch, first, the preprocessed flue gas image data is subjected to global maximum pooling and global average pooling operations in the channel dimension. The global maximum pooling operation is to find the maximum value on each channel and combine these maximum values into a single-channel feature map; the global average pooling operation is to calculate the average value on each channel and combine these average values into another single-channel feature map. Through these two pooling operations, the channel information of the preprocessed flue gas image data can be compressed from different angles to generate two single-channel feature maps, which respectively reflect different features of the image. The feature map obtained by global maximum pooling highlights the maximum value information in the image, while the feature map obtained by global average pooling reflects the average information of the image.

[0048] Step 1332: performing a channel splicing operation on the two single-channel feature maps to generate a double-channel feature map.

[0049] Then, the generated two single-channel feature maps are subjected to a channel splicing operation. Channel splicing is to merge two single-channel feature maps in the channel dimension to generate a double-channel feature map, so that different feature information obtained by global maximum pooling and global average pooling can be integrated, so that the maximum value information and average information of the image can be considered at the same time. In the splicing process, it is necessary to ensure that the spatial dimensions of the two single-channel feature maps are consistent to ensure the correctness and effectiveness of splicing.

[0050] Step 1333: performing a convolution operation on the double-channel feature map to compress the channel number to a single channel to generate an initial spatial attention weight map.

[0051] Then, the double-channel feature map is subjected to a convolution operation. By sliding the convolution kernel on the double-channel feature map, convolution operation is performed to compress the channel number to a single channel. The convolution operation can extract the feature information in the double-channel feature map and integrate it into a single channel. After the convolution operation, an initial spatial attention weight map is generated, which contains spatial attention information of the image. Each pixel point corresponds to a weight value, and these weight values reflect the importance of the pixel point in the spatial dimension.

[0052] Step 1334: applying an activation function to the initial spatial attention weight map to normalize the weight values to a preset interval.

[0053] In order to make the weight values of the initial spatial attention weight map comparable and interpretable, an activation function is applied to normalize the weight values to a pre-set interval. Common activation functions such as the Sigmoid function can map the weight values to an interval between 0 and 1. Through the action of the activation function, the range of weight values is more reasonable, facilitating subsequent processing and analysis. The normalized weight values can more accurately reflect the spatial importance of different regions in the image.

[0054] Step 1335: Match the normalized spatial attention weight map with the spatial dimensions of the original pre-processed flue gas image data, with each spatial position corresponding to a unique weight value.

[0055] Then, the normalized spatial attention weight map is matched with the spatial dimensions of the original pre-processed flue gas image data. Ensure that each spatial position in the spatial attention weight map corresponds to a pixel point in the original pre-processed flue gas image data, and each spatial position corresponds to a unique weight value. Thus, the spatial dimensions of the pre-processed flue gas image data can be weighted according to these weight values, highlighting the features of important regions.

[0056] Step 1336: Adjust the distribution of the spatial attention weight map through an iterative optimization process, so that the weight values are positively correlated with the spatial saliency of the flue gas region.

[0057] In order to make the spatial attention weight map more accurately reflect the spatial saliency of the flue gas region, the distribution of the spatial attention weight map is adjusted through an iterative optimization process. In the iterative optimization process, a loss function is used to measure the difference between the spatial attention weight map and the actual saliency of the flue gas region. According to the calculation result of the loss function, the weight values of the spatial attention weight map are adjusted through an optimization algorithm. After several iterations, the weight values are positively correlated with the spatial saliency of the flue gas region, i.e. the higher the spatial saliency of the flue gas region, the greater the corresponding weight value.

[0058] Step 1337: The spatial attention weight map after iterative optimization is taken as the spatial attention weight map.

[0059] Finally, the spatial attention weight map after iterative optimization is taken as the final spatial attention weight map, which can accurately reflect the importance distribution in the spatial dimensions of the pre-processed flue gas image data.

[0060] Step 134: Multiply the initial spectral feature map and the spatial attention weight map element by element to obtain a spatially weighted spectral feature map.

[0061] The generated initial spectral feature map is element-wise multiplied with the spatial attention weight map. Element-wise multiplication is to multiply the elements at corresponding positions in the initial spectral feature map and the spatial attention weight map. Through this operation, the weight values in the spatial attention weight map are applied to the initial spectral feature map to obtain a spatially weighted spectral feature map. In the garbage discharge recognition scene, the weight values in the spatial attention weight map reflect the spatial importance of different regions in the image. Through element-wise multiplication, the features of important regions in the initial spectral feature map are enhanced, and the features of unimportant regions are suppressed. Thus, the spatially weighted spectral feature map highlights the features of the smoke region more, improving the quality and accuracy of the features.

[0062] Step 135: Perform multi-scale feature fusion processing on the spatially weighted spectral feature map to generate a multi-scale spectral feature map by parallel convolution operations of different sizes of convolution kernels on the feature map.

[0063] The multi-scale feature fusion processing is performed on the spatially weighted spectral feature map. Different sizes of convolution kernels are used to perform parallel convolution operations on the spatially weighted spectral feature map. Different sizes of convolution kernels can capture different scale feature information. Large size convolution kernels can focus on global features of the image, and small size convolution kernels can focus on local detailed features of the image. In the parallel convolution operation process, multiple convolution kernels of different sizes slide on the spatially weighted spectral feature map simultaneously to perform convolution operations. Each convolution kernel extracts different scale feature information, which is then fused. In this way, a multi-scale spectral feature map is generated, which contains spectral feature information of different scales and can more comprehensively reflect the features of the smoke.

[0064] Step 136: Channel dimensionally splice the multi-scale spectral feature map to obtain a smoke spectral feature set containing response characteristics of different wavelength intervals.

[0065] Finally, the generated multi-scale spectral feature map is channel dimensionally spliced. Channel dimensionally splicing is to merge multiple multi-scale spectral feature maps in the channel dimension to obtain a smoke spectral feature set containing response characteristics of different wavelength intervals. In the splicing process, the spatial dimensions of each multi-scale spectral feature map need to be consistent to ensure the correctness and effectiveness of the splicing. Through channel dimensionally splicing, spectral feature information of different scales is integrated together to form a more comprehensive and richer smoke spectral feature set, which contains spectral response intensity distribution of different wavelength intervals and spatial dimension feature response significance distribution.

[0066] Step 140: Based on the multi-scale residual connection, difference feature identification processing is performed on the set of flue gas spectrum features to generate a difference description feature corresponding to the waste discharge flue gas image. The flue gas spectrum feature set and the difference description feature of the discharge substance are input into the classification network optimized by transfer learning. Through feature correlation modeling and classification probability prediction, the heterogeneous discharge substance recognition label is obtained. The heterogeneous discharge substance recognition label is used to indicate the different types of pollutant components contained in the waste discharge flue gas and the relative content proportion relationship of the pollutant components.

[0067] Step 141: A multi-scale residual connection network is constructed, which includes a plurality of residual blocks of different scales. Each residual block includes a convolution layer, a batch normalization layer, an activation function layer, and a skip connection connected in sequence.

[0068] In the waste discharge identification scene, in order to more accurately identify the difference features of the discharge substance, a multi-scale residual connection network is constructed, which includes a plurality of residual blocks of different scales. Each residual block is composed of a convolution layer, a batch normalization layer, an activation function layer, and a skip connection connected in sequence. The convolution layer is used to extract features by sliding the convolution kernel on the input feature map and performing convolution operation on the features to extract different feature information. The batch normalization layer normalizes the output of the convolution layer to make the data distribution more stable and speed up the convergence speed of the model. The activation function layer introduces a nonlinear factor to enhance the expression ability of the model. The skip connection directly adds the input feature map to the output of the convolution layer, which can alleviate the gradient vanishing problem and enable the network to learn more complex features. Different scale residual blocks can capture different scale feature information. Large-scale residual blocks focus on global features, and small-scale residual blocks focus on local detail features.

[0069] Step 142: The set of flue gas spectrum features is input into the input layer of the multi-scale residual connection network, and the set of flue gas spectrum features is processed by different scale residual blocks for feature conversion and enhancement to generate a set of multi-scale residual features.

[0070] The previously obtained set of flue gas spectrum features is input into the input layer of the multi-scale residual connection network. In the network, different scale residual blocks process the set of flue gas spectrum features in sequence. Each residual block converts and enhances the input features through the combination of convolution layer, batch normalization layer, activation function layer, and skip connection. Large-scale residual blocks extract and enhance global features of the set of flue gas spectrum features, and small-scale residual blocks focus on local feature processing. After processing by multiple residual blocks of different scales, a set of multi-scale residual features is generated, which contains feature information at different scales and can more comprehensively reflect the features of the waste discharge flue gas.

[0071] Step 143: Perform difference feature identification processing on the multi-scale residual feature set, and generate emission substance difference description features by calculating the difference degrees between different residual block output features, wherein the difference degrees are measured by the cosine similarity between feature vectors.

[0072] The difference feature identification processing is performed on the generated multi-scale residual feature set. In order to calculate the difference degrees between different residual block output features, the cosine similarity between feature vectors is used for measurement. First, the feature vectors output by different residual blocks are extracted from the multi-scale residual feature set. Then, the cosine similarity between these feature vectors is calculated. The cosine similarity can measure the cosine value of the included angle between two vectors. The smaller the included angle, the greater the cosine similarity, indicating that the two vectors are more similar. The greater the included angle, the smaller the cosine similarity, indicating that the two vectors are more different. By calculating the cosine similarity between different residual block output feature vectors, the difference degrees between them are obtained. According to these difference degrees, the emission substance difference description features are generated, which can reflect the difference between different emission substances in the waste emission flue gas.

[0073] Step 144: Feature splicing is performed on the flue gas spectrum feature set and the emission substance difference description features to generate a fusion feature vector.

[0074] The flue gas spectrum feature set and the emission substance difference description features are spliced. Feature splicing is to merge two features in the channel dimension to generate a fusion feature vector. In the garbage emission identification scene, the flue gas spectrum feature set contains the spectral response characteristics of different wavelength intervals and the spatial dimension feature response significance distribution, while the emission substance difference description features reflect the differences between different emission substances. Through feature splicing, these two different types of feature information are integrated to form a more comprehensive and rich fusion feature vector, which contains more feature information and can improve the accuracy and reliability of the classification network.

[0075] Step 145: Input the fusion feature vector into the classification network optimized by transfer learning, wherein the initial weight parameters of the classification network are obtained by pre-training on the source domain dataset.

[0076] Step 1451: Select a source domain image dataset containing multiple industrial emission flue gas scenes, wherein the source domain image dataset contains flue gas image samples with different pollutant components and different concentration levels, and each sample is labeled with pollutant component type and content ratio label.

[0077] In order to optimize the transfer learning of the classification network, first, a source domain image dataset is selected, which contains multiple industrial emission flue gas scenes, covering flue gas image samples of different pollutant components and different concentration levels. Each sample is labeled with pollutant component type and content ratio label. In the actual selection process, the diversity and representativeness of the source domain image dataset need to be ensured, including various pollutants and different concentration conditions, so that the classification network can learn more extensive feature information during pre-training, improving the generalization ability of the model.

[0078] Step 1452: Construct a basic classification network, which includes a feature extraction subnetwork, a feature fusion subnetwork, and a classification output subnetwork connected in sequence, wherein the feature extraction subnetwork adopts a deep convolutional neural network structure.

[0079] Then, a basic classification network is constructed, which is composed of a feature extraction subnetwork, a feature fusion subnetwork, and a classification output subnetwork connected in sequence. The feature extraction subnetwork adopts a deep convolutional neural network structure, which extracts features from the input image data through multiple convolutional layers and pooling layers. The convolutional layer can extract local features of the image, and the pooling layer can reduce the spatial size of the feature map and reduce the computational complexity. The feature fusion subnetwork fuses the features output by the feature extraction subnetwork, integrating feature information at different levels. The classification output subnetwork performs classification prediction based on the fused features and outputs the classification probability of different pollutant components.

[0080] Step 1453: Pre-train the basic classification network using the source domain image dataset, set the preset training rounds, batch size, and initial learning rate, and calculate the prediction loss value using the cross-entropy loss function.

[0081] The selected source domain image dataset is used to pre-train the basic classification network. In the pre-training process, the preset training rounds, batch size, and initial learning rate are set. The training rounds represent the number of times the model is trained on the entire source domain image dataset, and the batch size is the number of samples used for each training. The initial learning rate controls the step size of the model parameter update. The prediction loss value of the model is calculated by the cross-entropy loss function. The cross-entropy loss function can measure the difference between the model's prediction and the true label, and the smaller the loss value, the more accurate the model's prediction. During training, the model's parameters are constantly adjusted to gradually reduce the loss value.

[0082] Step 1454: Update the network weight parameters based on the backpropagation algorithm and gradient descent optimizer, and after each training round, evaluate the model performance using the source domain validation set, and record the classification accuracy and loss value of the model on the source domain validation set.

[0083] The network weight parameters are updated based on the backpropagation algorithm and gradient descent optimizer. The backpropagation algorithm is used to calculate the gradient of the loss function with respect to the model parameters, and the gradient descent optimizer updates the model parameters based on the gradient information, so that the loss value gradually decreases. After each training round, the performance of the model is evaluated using the source domain validation set. The source domain validation set is a portion of data divided from the source domain image dataset, used to verify the generalization ability of the model. The classification accuracy and loss value of the model on the source domain validation set are recorded, and the training effect and performance of the model are judged by observing the changes of these indicators.

[0084] Step 1455: When the classification accuracy of the model on the source domain validation set no longer improves for a preset number of consecutive times, stop pre-training and save the current network weight parameters as initial weight parameters.

[0085] When the classification accuracy of the model on the source domain validation set no longer improves for a preset number of consecutive times, it means that the model has reached a good state, and further training may cause overfitting. At this time, the pre-training is stopped, and the current network weight parameters are saved as initial weight parameters, which contain the feature information learned by the model on the source domain dataset.

[0086] Step 1456: Select the waste treatment area flue gas image dataset as the target domain dataset, and perform data augmentation processing on the target domain dataset to generate an expanded target domain dataset.

[0087] The waste treatment area flue gas image dataset is selected as the target domain dataset. In order to increase the diversity and quantity of the target domain dataset, data augmentation processing is performed on the target domain dataset. Data augmentation processing can generate new image samples by rotating, flipping, scaling, etc. Through data augmentation processing, an expanded target domain dataset is generated, which contains more samples and richer feature information, which can improve the performance of the model on the target domain.

[0088] Step 1457: Load the initial weight parameters to the base classification network, freeze the first network layer weight parameters of the feature extraction subnet, and unfreeze the second network layer weight parameters of the feature extraction subnet, as well as all network layer weight parameters of the feature fusion subnet and the classification output subnet.

[0089] The saved initial weight parameters are loaded to the base classification network. Then, the first network layer weight parameters of the feature extraction subnet are frozen. Freezing weight parameters means that these parameters will not be updated during fine-tuning training. Unfreeze all network layer weight parameters of the second network layer of the feature extraction subnet, as well as the feature fusion subnet and the classification output subnet, so that these parameters can be updated during fine-tuning training. In this way, the feature information learned on the source domain dataset can be utilized, and the model can be appropriately adjusted on the target domain dataset to adapt to the features of the target domain.

[0090] Step 1458: Fine-tuning the unfrozen base classification network using the extended target domain dataset, using an initial learning rate, and adjusting the training process by monitoring the performance indicators of the target domain validation set.

[0091] Fine-tuning the unfrozen base classification network using the extended target domain dataset. During the fine-tuning process, an initial learning rate is used. At the same time, the performance indicators of the target domain validation set are monitored to adjust the training process. The target domain validation set is a part of data divided from the extended target domain dataset, used to verify the performance of the model on the target domain. By observing the changes of indicators such as classification accuracy and loss value of the target domain validation set, the parameters and strategies of the training are adjusted to ensure the continuous improvement of the performance of the model on the target domain.

[0092] Step 1459: During the fine-tuning process, when the loss value of the target domain validation set rises continuously for a preset number of times, reduce the learning rate and continue training; when the classification accuracy of the target domain validation set reaches a preset accuracy, stop fine-tuning, save the corresponding network weight parameters, and obtain the classification network optimized by transfer learning.

[0093] During the fine-tuning process, if the loss value of the target domain validation set rises continuously for a preset number of times, it indicates that the model may have overfitting or excessive learning rate problems. At this time, reduce the learning rate and continue training to reduce the step size of model parameter update, so that the model learns more stably. When the classification accuracy of the target domain validation set reaches a preset accuracy, it indicates that the model has achieved good performance on the target domain, and the fine-tuning is stopped. Save the corresponding network weight parameters. After the above transfer learning optimization process, the classification network optimized by transfer learning is obtained, which can better distinguish different pollutant components in waste discharge flue gas.

[0094] Step 146: In the classification network, the feature association modeling layer models the association relationship between the fusion feature vectors to generate a pollutant component association matrix.

[0095] Step 1461: Input the fusion feature vector into the feature association modeling layer, and map the fusion feature vector to a feature space of a first preset dimension through a first fully connected layer to generate a first feature vector.

[0096] In the classification network, the feature correlation modeling layer starts processing the fusion feature vector. First, the fusion feature vector is input into the first fully connected layer. The fully connected layer maps the input fusion feature vector to a feature space of a first preset dimension, generating a first feature vector. Each neuron in the fully connected layer is connected to each element of the input vector, and through weighted summation and activation function, the input vector is converted into a new feature vector. The selection of the first preset dimension is determined according to the specific task and data characteristics to ensure that effective feature information can be extracted.

[0097] Step 1462: input the first feature vector into the second fully connected layer, map it to a feature space of a second preset dimension, generate a second feature vector, and the second preset dimension corresponds to the preset number of pollutant component categories.

[0098] Then, the first feature vector is input into the second fully connected layer, which maps the first feature vector to a feature space of a second preset dimension, generating a second feature vector. The second preset dimension corresponds to the preset number of pollutant component categories, so that the dimension of the feature vector can be matched with the number of pollutant component categories, so as to model the correlation between the pollutant components. Through the mapping of the second fully connected layer, the feature information is further extracted and converted, so that the feature vector can better reflect the relationship between the pollutant components.

[0099] Step 1463: transpose the second feature vector to generate a transposed vector of the second feature vector.

[0100] The generated second feature vector is transposed. The transposition operation exchanges the rows and columns of the second feature vector to generate a transposed vector of the second feature vector. In subsequent calculations, the transposed vector will be used for matrix multiplication with the second feature vector to generate an association matrix. The purpose of the transposition operation is to meet the operation rules of matrix multiplication, to ensure the correctness and effectiveness of the calculation.

[0101] Step 1464: calculate the matrix product of the second feature vector and the transposed vector to generate an initial association matrix, and the number of rows and columns of the initial association matrix is the second preset dimension.

[0102] The matrix product of the second feature vector and the transposed vector is calculated. Matrix multiplication is to multiply the corresponding elements of each row of the second feature vector and each column of the transposed vector and sum them up to get a new matrix, which is the initial association matrix. The number of rows and columns of the initial association matrix is the second preset dimension, which corresponds to the preset number of pollutant component categories, and each element in the matrix represents the correlation degree between different pollutant components.

[0103] Step 1465: Perform row standardization on the initial correlation matrix, making the sum of each row element in the matrix equal to 1, to generate a standardized correlation matrix.

[0104] In order to make the elements in the initial correlation matrix comparable and interpretable, row standardization is performed. Row standardization is to divide each row element of the initial correlation matrix by the sum of the elements in that row, so that the sum of each row element in the matrix is 1. Through row standardization, a standardized correlation matrix is generated, and the elements in the matrix represent the relative probability of the occurrence of other pollutant components in the presence of a certain pollutant component.

[0105] Step 1466: Construct a pollutant component prior knowledge matrix, which is constructed based on known pollutant coexistence relationships, and the matrix elements represent the probability of the simultaneous occurrence of two pollutant components.

[0106] A pollutant component prior knowledge matrix is constructed, which is constructed based on known pollutant coexistence relationships, and the matrix elements represent the probability of the simultaneous occurrence of two pollutant components. In the actual construction process, relevant prior knowledge and data need to be collected and sorted out, and the coexistence of different pollutants is analyzed. According to this information, the value of each element in the matrix is determined. The pollutant component prior knowledge matrix can provide additional information and constraints for subsequent correlation modeling, making the results of the model more consistent with the actual situation.

[0107] Step 1467: Perform element-by-element multiplication of the standardized correlation matrix and the pollutant component prior knowledge matrix to generate a weighted correlation matrix that integrates data-driven feature correlation information and prior knowledge.

[0108] Perform element-by-element multiplication of the standardized correlation matrix and the pollutant component prior knowledge matrix. Element-by-element multiplication is to multiply the elements at corresponding positions in two matrices to generate a weighted correlation matrix. Through this operation, data-driven feature correlation information and prior knowledge are integrated. The weighted correlation matrix considers both the correlation relationships learned from the fused feature vector and the known pollutant coexistence relationships, making the modeling of correlation relationships more accurate and reliable.

[0109] Step 1468: Perform nonlinear transformation on the weighted correlation matrix to enhance the expression of strong correlation relationships through an activation function layer to generate a pollutant component correlation matrix, where the element in the ith row and jth column of the matrix represents the correlation strength between the ith pollutant component and the jth pollutant component.

[0110] Finally, a nonlinear transformation is performed on the weighted correlation matrix. The weighted correlation matrix is processed through an activation function layer to enhance the expression of strong correlation. The activation function can perform nonlinear mapping on the elements in the weighted correlation matrix, highlighting the elements of strong correlation and suppressing the elements of weak correlation. After processing by the activation function layer, a pollutant component correlation matrix is generated, and the element in the ith row and jth column of the matrix represents the correlation strength between the ith pollutant component and the jth pollutant component.

[0111] Step 147: Based on the pollutant component correlation matrix, the existence probability and relative content ratio of different pollutant components are calculated by a classification probability prediction layer.

[0112] Based on the generated pollutant component correlation matrix, the classification probability prediction layer starts to calculate the existence probability and relative content ratio of different pollutant components. The classification probability prediction layer can use methods such as softmax activation function to convert the correlation strength information in the pollutant component correlation matrix into probability values. By calculating and comparing the probability of each pollutant component, the existence probability of different pollutant components is obtained. At the same time, according to these existence probabilities, the relative content ratio of different pollutant components is calculated. In the calculation process, it is necessary to ensure that the sum of the probabilities of all pollutant components is 1, and the sum of the relative content ratios is also 1, so as to accurately reflect the relative content of different pollutant components in the waste emission flue gas.

[0113] Step 148: Generating a heterogeneous emission substance identification label according to the existence probability and relative content ratio, the identification label containing the type of pollutant component and its corresponding relative content ratio relationship.

[0114] According to the calculated existence probability and relative content ratio of different pollutant components, a heterogeneous emission substance identification label is generated, which contains the type of pollutant component and its corresponding relative content ratio relationship. In the generation process, the type of pollutant component and the corresponding relative content ratio are sorted and recorded to form a clear label, which can intuitively reflect the different types of pollutant components contained in the waste emission flue gas and their relative content ratio.

[0115] As other scalable embodiments, after obtaining the heterogeneous emission substance identification label, it further includes:

[0116] Step 151: Associating and mapping the heterogeneous emission substance identification label with the corresponding preprocessed flue gas image data and collection time period information to obtain a labeled flue gas image unit containing pollutant component type, relative content ratio, spatial texture details, and collection timestamp.

[0117] After obtaining the heterogeneous emission substance identification label, it is associated and mapped with the corresponding pre-processed flue gas image data and collection time period information. Through this association and mapping, the pollutant component type, relative content ratio, and spatial texture details of the pre-processed flue gas image data, and the collection time stamp are combined to obtain a labeled flue gas image unit. In the actual association process, a data structure can be established to store these information together, so that each labeled flue gas image unit contains rich information, not only the component and content information of the pollutant, but also the spatial characteristics of the image and the collection time information.

[0118] Step 152: performing normalization processing on the relative content ratio of the pollutant components in the labeled flue gas image unit, mapping the content ratio values of different pollutant components to a preset color intensity interval, generating a multi-channel color mapping matrix, and the number of channels of the color mapping matrix is consistent with the number of pollutant component types in the heterogeneous emission substance identification label.

[0119] The relative content ratio of the pollutant components in the labeled flue gas image unit is normalized. The normalization processing is to map the content ratio values of different pollutant components to a preset color intensity interval. First, determine the range of the preset color intensity interval, then according to the maximum and minimum values of the relative content ratio of the pollutant components, linearly map the content ratio value of each pollutant component to fall within the preset color intensity interval. According to the normalized content ratio value, a multi-channel color mapping matrix is generated. The number of channels of the color mapping matrix is consistent with the number of pollutant component types in the heterogeneous emission substance identification label, and each channel corresponds to a pollutant component. The elements in the matrix represent the color intensity value of the pollutant component at the corresponding position. Through this color mapping, the distribution of different pollutant components can be intuitively displayed.

[0120] Step 153: aligning the multi-channel color mapping matrix with the spatial dimension of the pre-processed flue gas image data, superimposing the color mapping matrix to the corresponding spatial region of the pre-processed flue gas image data through a pixel-by-pixel weighted fusion algorithm, generating a pollutant spatial distribution enhanced image, and the color intensity in the pollutant spatial distribution enhanced image is positively correlated with the relative content ratio of the pollutant.

[0121] The generated multi-channel color mapping matrix is aligned with the spatial dimensions of the preprocessed flue gas image data. Ensure that the spatial dimensions of the color mapping matrix and the preprocessed flue gas image data are consistent, and each pixel position corresponds to the same spatial position. Then, the color mapping matrix is superimposed on the corresponding spatial region of the preprocessed flue gas image data through a pixel-by-pixel weighted fusion algorithm. In the weighted fusion process, according to the color intensity value in the color mapping matrix and the pixel value of the preprocessed flue gas image data, the fusion is carried out according to a certain weight. Finally, a pollutant spatial distribution enhancement image is generated, in which the color intensity is positively correlated with the relative content proportion of the pollutant. The stronger the color, the higher the relative content proportion of the pollutant in that area, so that the spatial distribution of the pollutant can be observed directly.

[0122] Step 154: Based on the collection period information, the pollutant spatial distribution enhancement images of different collection periods are time-sequentially sorted to generate a pollutant content change sequence, which contains the evolution process of the spatial distribution of the pollutant component in the continuous period.

[0123] Based on the collection period information, the pollutant spatial distribution enhancement images of different collection periods are time-sequentially sorted. According to the order of collection time, these images are arranged to generate a pollutant content change sequence, which contains the evolution process of the spatial distribution of the pollutant component in the continuous period. By observing this sequence, the change of the content and distribution of the pollutant in the waste discharge flue gas over time can be understood. For example, it can be seen how the content of some pollutants gradually increases or decreases, and how the spatial distribution of the pollutant changes.

[0124] Step 155: Integrate the dynamic change sequence with the pollutant component type information in the heterogeneous emission substance identification label to generate an interactive monitoring visualization query interface containing pollutant type annotation, content proportion value and dynamic change trend.

[0125] Finally, the generated pollutant content change sequence is integrated with the pollutant component type information in the heterogeneous emission substance identification label. The pollutant type annotation, content proportion value and dynamic change trend are combined to generate an interactive monitoring visualization query interface. In this interface, the information of the pollutant can be displayed intuitively through graphs, charts and other means. Users can query the content and distribution of the pollutant at different times through interactive operation, and observe the dynamic change trend of the pollutant. The above interface can provide a convenient tool for waste discharge monitoring and management.

[0126] As another scalable embodiment, after obtaining the heterogeneous emission substance identification label, it further includes:

[0127] Step 161: Extract the pollutant composition entity set and the corresponding relative content ratio attribute from the heterogeneous emission substance identification tag, and extract the working condition characteristic entity set from each group of smoke image unit corresponding to the garbage combustion working condition. The working condition characteristic entity set includes combustion temperature, oxygen supply rate, and material composition type.

[0128] After obtaining the heterogeneous emission substance identification tag, the pollutant composition entity set and the corresponding relative content ratio attribute are extracted from the tag. At the same time, the working condition characteristic entity set is extracted from each group of smoke image unit corresponding to the garbage combustion working condition. The working condition characteristic entity set includes combustion temperature, oxygen supply rate, and material composition type, etc. In the extraction process, it is necessary to ensure accurate acquisition of these information and organize them into appropriate data structures.

[0129] Step 162: Construct an initial knowledge graph triple set containing pollutant composition-content ratio-working condition characteristics and pollutant composition-coexistence frequency-pollutant composition. The coexistence frequency is obtained by counting the number of times of the appearance of the pollutant composition pair at different collection time periods.

[0130] Based on the extracted pollutant composition entity set, relative content ratio attribute and working condition characteristic entity set, an initial knowledge graph triple set is constructed, which contains pollutant composition-content ratio-working condition characteristics, pollutant composition-coexistence frequency-pollutant composition and other triple relationship. In the construction process, the relationship between the elements of each triple is determined. The coexistence frequency is obtained by counting the number of times of the appearance of the pollutant composition pair at different collection time periods. For example, for two pollutant compositions, the number of times they appear simultaneously in all collection time periods is counted, and then divided by the total collection time period to obtain their coexistence frequency. The coexistence frequency can reflect the degree of association between different pollutant compositions.

[0131] Step 163: Call the emission substance difference description feature output by the multi-scale residual connection network to optimize the relationship weight of the initial knowledge graph triple set, use the difference degree information in the emission substance difference description feature as the adjustment basis of the triple relationship strength, and generate a weighted knowledge graph triple set.

[0132] The emission substance difference description features output by the multi-scale residual connection network are used to optimize the relationship weights of the initial knowledge graph triple set. The difference degree information in the emission substance difference description features is used as the basis for adjusting the relationship strength of the triple. The difference degree information can reflect the differences between different emission substances. According to this difference degree information, the relationship weights in the initial knowledge graph triple set are adjusted. For example, if the difference degree between two pollutant components is large, the relationship weight between them can be appropriately reduced; if the difference degree is small, the relationship weight can be appropriately increased. Through the above optimization process, a weighted knowledge graph triple set is generated, so that the relationships in the knowledge graph more accurately reflect the actual situation.

[0133] Step 164: Construct a directed and weighted knowledge graph based on the weighted knowledge graph triple set. In the directed and weighted knowledge graph, the nodes are pollutant component entities and working condition feature entities, the directed edges are the relationship types in the triples, and the edge weights are the optimized relationship strength values.

[0134] A directed and weighted knowledge graph is constructed based on the weighted knowledge graph triple set. In this knowledge graph, the nodes are pollutant component entities and working condition feature entities, the directed edges are the relationship types in the triples, and the edge weights are the optimized relationship strength values. In the construction process, according to the triple relationships in the weighted knowledge graph triple set, the connection mode and weight of the nodes and edges are determined. The directed and weighted knowledge graph can intuitively show the relationships between pollutant components, content proportions, and working condition features, and can be used to mine the internal mechanisms of garbage combustion processes and pollutant emissions.

[0135] Step 165: Perform a community discovery algorithm on the directed and weighted knowledge graph to identify pollutant component communities with a coexistence frequency reaching a preset frequency and corresponding working condition features, and generate a pollutant-working condition association rule set.

[0136] A community discovery algorithm is performed on the constructed directed and weighted knowledge graph. The community discovery algorithm can divide the nodes in the knowledge graph into different communities, and the nodes within each community have a strong association relationship. By performing the community discovery algorithm, pollutant component communities with a coexistence frequency reaching a preset frequency and corresponding working condition features are identified. In the identification process, according to the coexistence frequency and edge weight information, it is determined which nodes belong to the same community. According to the identification result, a pollutant-working condition association rule set is generated, which contains the association rules between different pollutant component communities and corresponding working condition features. These rules can provide important guidance for the optimization of garbage combustion processes and the control of pollutant emissions.

[0137] By applying the embodiment of the present application, the accuracy and reliability of identifying heterogeneous emission substances in waste emission flue gas can be improved. First, a set of flue gas image data is comprehensively collected in different time periods and under different combustion conditions in the waste treatment area. The coverage of different collection periods and combustion conditions enables the data to fully reflect the dynamic changes and diversity of pollutant emissions during waste combustion. Then, the set of flue gas image data is subjected to image enhancement processing to generate preprocessed flue gas image data with enhanced spectral features, which not only retains the spectral distribution characteristics and spatial texture details of the original image but also strengthens the spectral features, enabling feature mining to more accurately capture key information in the flue gas. The enhancement of spectral features can highlight the unique performance of different pollutants in the spectrum, thereby enabling accurate differentiation and identification of pollutants. Then, a double-channel convolution attention network is called to mine features from the preprocessed flue gas image data. Through parallel processing of the spectral feature extraction branch and the spatial feature enhancement branch, the features of the flue gas image can be fully mined from both the spectral and spatial dimensions. The spectral feature extraction branch focuses on the spectral response intensity distribution in different wavelength intervals, enabling accurate identification of the spectral features of different pollutants; the spatial feature enhancement branch highlights the feature response significance distribution in the spatial dimension, which helps to determine the distribution of pollutants in space, and parallel processing of the two branches greatly improves the efficiency and comprehensiveness of feature mining. Finally, a multi-scale residual connection is used to perform difference feature identification processing on the set of flue gas spectral features to generate difference description features of emission substances, which are input into a classification network optimized by transfer learning together with the set of flue gas spectral features. The multi-scale residual connection can effectively capture feature differences at different scales, making the difference description features of emission substances better reflect the essential differences between pollutants. The classification network optimized by transfer learning combines the prior knowledge of the source domain data and the characteristics of the target domain data, and through feature correlation modeling and classification probability prediction, it can accurately obtain the identification labels of heterogeneous emission substances, clearly indicate the composition and relative content ratio of different types of pollutants in the waste emission flue gas, and thus improve the accuracy and reliability of pollutant identification.

[0138] Based on the same inventive concept, the embodiment of the present application also provides a waste emission identification system applying a deep learning model. Referring to Figure 2 Fig. 1 shows a possible structure of a waste emission identification system applying a deep learning model provided in the embodiment of the present application, Figure 2 In the embodiment, the waste emission identification system applying a deep learning model 200 includes a processor 210 and a memory 220. The memory 220 stores a computer program executable by the processor 210, and the processor 210 can execute the steps of the waste emission identification method applying a deep learning model by executing the instructions stored in the memory 220.

[0139] Based on the same inventive concept, the embodiments of the present application provide a computer readable storage medium comprising a computer program, when the computer program is run on a garbage discharge identification system applying a deep learning model, the computer program is configured to cause the garbage discharge identification system applying the deep learning model to perform the steps of the garbage discharge identification method applying the deep learning model described above. In some possible implementation manners, each aspect of the garbage discharge identification method applying the deep learning model provided by the present application can also be implemented in the form of a program product, which comprises a computer program, when the program product is run on a garbage discharge identification system applying a deep learning model, the computer program is configured to cause the garbage discharge identification system applying the deep learning model to perform the steps of the garbage discharge identification method applying the deep learning model described above, for example, the garbage discharge identification system applying the deep learning model can perform the steps as shown in Figure 1 .

[0140] The above description is merely preferred exemplary embodiments of the present application, but not intended to limit the implementation of the present application. Those skilled in the art can easily make corresponding modifications or changes according to the main concept and spirit of the present application.

Claims

1. A method of garbage discharge recognition using a deep learning model, the method comprising: The method comprises: Collecting a flue gas image data set of a waste treatment area, the flue gas image data set containing multiple groups of flue gas image units under different collection time periods, each group of flue gas image units corresponding to different waste combustion conditions; Performing image enhancement processing on the flue gas image data set to generate preprocessed flue gas image data with spectral feature enhancement effect, the preprocessed flue gas image data retaining spectral distribution characteristics and spatial texture details in the original flue gas image units; Calling a dual-channel convolution attention network to perform feature mining operations on the preprocessed flue gas image data, obtaining a flue gas spectral feature set through parallel processing of a spectral feature extraction branch and a spatial feature strengthening branch, the flue gas spectral feature set containing spectral response intensity distribution in different wavelength intervals and feature response significance distribution in spatial dimensions; Performing difference feature recognition processing on the flue gas spectral feature set based on multi-scale residual connection to generate emission substance difference description features corresponding to waste emission flue gas, inputting the flue gas spectral feature set and the emission substance difference description features into a classification network optimized through transfer learning, and obtaining heterogeneous emission substance recognition labels through feature correlation modeling and classification probability prediction, the heterogeneous emission substance recognition labels being used to indicate different types of pollutant components contained in waste emission flue gas and relative content proportion relationships of the pollutant components: A multi-scale residual connection network is constructed, the multi-scale residual connection network containing multiple residual blocks of different scales, each residual block containing convolution layers, batch normalization layers, activation function layers and skip connections connected in sequence; The flue gas spectral feature set is input into an input layer of the multi-scale residual connection network, and the flue gas spectral feature set is processed through feature conversion and enhancement by residual blocks of different scales to generate a multi-scale residual feature set; Difference feature recognition processing is performed on the multi-scale residual feature set to generate emission substance difference description features by calculating the difference degrees between the output features of different residual blocks, the difference degrees being measured by cosine similarity between feature vectors; The flue gas spectral feature set and the emission substance difference description features are spliced to generate a fusion feature vector; The fusion feature vector is input into a classification network optimized through transfer learning, initial weight parameters of the classification network being obtained through pre-training on a source domain data set; In the classification network, the fusion feature vector is modeled for pollutant component correlation relationship by a feature correlation modeling layer to generate a pollutant component correlation matrix; Based on the pollutant component correlation matrix, the existence probability and relative content proportion of different pollutant components are calculated by a classification probability prediction layer; A heterogeneous emission substance recognition label is generated according to the existence probability and relative content proportion, the recognition label containing pollutant component types and corresponding relative content proportion relationships.

2. The method of claim 1, wherein, The image enhancement processing on the flue gas image data set to generate preprocessed flue gas image data with spectral feature enhancement effect comprises: Performing illumination non-uniformity correction processing on each group of flue gas image units in the flue gas image data set; An adaptive contrast adjustment operation is performed on the corrected flue gas image unit by a sub-region histogram equalization algorithm to enhance the contrast of the weak spectral signal region in the image; A noise removal algorithm based on spatial domain filtering is used to perform noise suppression processing on the flue gas image unit after contrast adjustment to eliminate random noise introduced in the image acquisition process; The flue gas image unit after noise suppression is converted to the HSV color space, and independent gain adjustment operations are performed on the hue channel and the saturation channel to strengthen the distinguishability of the flue gas spectral features in the color space; The flue gas image unit after HSV color space adjustment is converted back to the RGB color space to obtain preprocessed flue gas image data with spectral feature enhancement effect, and the size of the preprocessed flue gas image data remains consistent with that of the original flue gas image unit.

3. The method of claim 1, wherein, The double-channel convolution attention network is called to perform feature mining operations on the preprocessed flue gas image data, and through parallel processing of the spectral feature extraction branch and the spatial feature enhancement branch, a flue gas spectral feature set is obtained, including: The preprocessed flue gas image data is input into the spectral feature extraction branch and the spatial feature enhancement branch of the double-channel convolution attention network; In the spectral feature extraction branch, multi-level convolution operations are performed on the spectral channels of the preprocessed flue gas image data to generate initial spectral feature maps; In the spatial feature enhancement branch, a spatial attention mechanism is used to allocate weights to the spatial dimensions of the preprocessed flue gas image data to generate a spatial attention weight map; The initial spectral feature map and the spatial attention weight map are multiplied element by element to obtain a spatially weighted spectral feature map; Multi-scale feature fusion processing is performed on the spatially weighted spectral feature map, and parallel convolution operations are performed on the feature map using convolution kernels of different sizes to generate multi-scale spectral feature maps; The multi-scale spectral feature maps are concatenated in the channel dimension to obtain a flue gas spectral feature set containing response characteristics in different wavelength intervals.

4. The method of claim 3, wherein, In the spectral feature extraction branch, multi-level convolution operations are performed on the spectral channels of the preprocessed flue gas image data to generate initial spectral feature maps, including: The preprocessed flue gas image data is input into the input layer of the spectral feature extraction branch, and the first convolution block is used to perform initial feature extraction on the preprocessed flue gas image data to generate a first-level spectral feature map, the first convolution block including a convolution layer, a batch normalization layer, and an activation function layer connected in sequence; The first-level spectral feature map is input into the second convolution block, and the feature channel dimension is expanded by increasing the number of convolution kernels to generate a second-level spectral feature map, the number of channels of the second-level spectral feature map being an integer multiple of the number of channels of the first-level spectral feature map; The second-level spectral feature map is input into the third convolution block, and the spatial resolution of the feature map is reduced by convolution operation with a step size of a preset value to generate a third-level spectral feature map, the spatial size of the third-level spectral feature map being a preset proportion of the spatial size of the second-level spectral feature map; Input the third-level spectral feature map into a fourth convolutional block, capture multi-scale spectral information through convolution operations of different receptive fields, and generate a fourth-level spectral feature map; Take the fourth-level spectral feature map as an initial spectral feature map, and the channel dimension of the initial spectral feature map contains spectral response characteristics of different wavelength intervals.

5. The method of claim 3, wherein, In the spatial feature enhancement branch, a spatial attention mechanism is used to allocate weights to the spatial dimension of the preprocessed flue gas image data to generate a spatial attention weight map, including: Performing global maximum pooling and global average pooling operations on the channel dimension of the preprocessed flue gas image data to generate two single-channel feature maps; Performing channel concatenation operation on the two single-channel feature maps to generate a double-channel feature map; Performing convolution operation on the double-channel feature map to compress the channel number to a single channel to generate an initial spatial attention weight map; Applying an activation function to the initial spatial attention weight map to normalize the weight values to a preset interval; Matching the normalized spatial attention weight map with the spatial size of the original preprocessed flue gas image data, each spatial position corresponds to a unique weight value; Adjust the distribution of the spatial attention weight map through an iterative optimization process to make the weight value positively correlated with the spatial saliency of the flue gas region; The spatial attention weight map obtained through the iterative optimization process is taken as the spatial attention weight map.

6. The method of claim 1, wherein, The construction of the multi-scale residual connection network includes: Determining the basic architecture of the multi-scale residual connection network, which includes an input layer, at least three residual block groups of different scales, a feature fusion layer, and an output layer, each residual block group is composed of a plurality of residual blocks of the same type connected in series; Constructing a first-scale residual block group, the residual blocks in the first-scale residual block group use a smaller size of convolution kernel to capture local subtle difference features in the flue gas spectral feature set, and the skip connection of each residual block uses an identity mapping method to directly connect the input and output; Constructing a second-scale residual block group, the residual blocks in the second-scale residual block group use a medium size of convolution kernel, and a convolution layer of a set size is introduced in the skip connection to adjust the channel dimension, so that the channel number of the input feature map and the output feature map remains the same; Constructing a third-scale residual block group, the residual blocks in the third-scale residual block group use a larger size of convolution kernel, and a channel attention mechanism module is added before the convolution operation to enhance the expression ability of key spectral features by allocating weights to feature channels; Connecting the first-scale residual block group, the second-scale residual block group and the third-scale residual block group in order from high to low according to the feature map resolution, and performing down-sampling processing on the adjacent residual block groups through a convolution layer with a preset step; Setting a lateral connection channel at the output end of each residual block group to deliver the output feature maps of the residual block groups of different scales to the feature fusion layer, and the feature fusion layer performs weighted aggregation processing on feature maps of different scales through an adaptive weight fusion algorithm.

7. The method of claim 1, wherein, The difference feature recognition processing on the multi-scale residual feature set generates an emission substance difference description feature by calculating the difference between the output features of different residual blocks, including: extracting feature maps output by each residual block group from the multi-scale residual feature set as a first feature map, a second feature map and a third feature map, the first feature map corresponding to the first scale residual block group output, the second feature map corresponding to the second scale residual block group output, and the third feature map corresponding to the third scale residual block group output; performing feature alignment processing on the first feature map and the second feature map, adjusting the second feature map to the same spatial size as the first feature map through an upsampling operation to generate an aligned second feature map; performing channel dimension splicing on the first feature map and the aligned second feature map, and learning the fusion and difference relationship between the first feature map and the aligned second feature map through a convolution operation to generate a first fusion difference feature map, the first fusion difference feature map containing the correlation and difference information of local features and middle-level features at different scales; performing feature alignment processing on the second feature map and the third feature map, adjusting the third feature map to the same spatial size as the second feature map through an upsampling operation to generate an aligned third feature map; performing channel dimension splicing on the second feature map and the aligned third feature map, and learning the fusion and difference relationship between the two through a convolution operation to generate a second fusion difference feature map, the second fusion difference feature map containing the correlation and difference information of middle-level features and global features at different scales; performing channel dimension splicing on the first fusion difference feature map and the second fusion difference feature map to generate an initial fusion difference feature map, the channel number of the initial fusion difference feature map being the sum of the channel numbers of the first fusion difference feature map and the second fusion difference feature map; performing feature dimension reduction processing on the initial difference feature map, compressing the channel number to a preset dimension through a convolution layer of a set size to generate a dimension-reduced difference feature map; applying a spatial attention mechanism to the dimension-reduced difference feature map to generate a spatial difference weight map, regions with higher weight values in the spatial difference weight map corresponding to regions with significant emission substance differences; performing element-by-element multiplication operation on the dimension-reduced difference feature map and the spatial difference weight map to obtain a weighted difference feature map; performing global average pooling operation on the weighted difference feature map to generate a one-dimensional feature vector as an emission substance difference description feature.

8. The method of claim 1, wherein, The fusion feature vector is input into a classification network optimized by transfer learning, and the initial weight parameters of the classification network are obtained by pre-training on a source domain dataset, including: selecting a source domain image dataset containing multiple industrial emission flue gas scenes, the source domain image dataset containing flue gas image samples of different pollutant compositions and different concentration levels, each sample being labeled with pollutant composition type and content ratio label; constructing a basic classification network, the basic classification network including a feature extraction subnetwork, a feature fusion subnetwork and a classification output subnetwork connected in turn, wherein the feature extraction subnetwork adopts a deep convolutional neural network structure; pre-training the basic classification network using the source domain image dataset, setting a preset training round, batch size and initial learning rate, and calculating the prediction loss value through a cross-entropy loss function; The network weight parameters are updated based on a back propagation algorithm and a gradient descent optimizer, and the model performance is evaluated using the source domain validation set after each training round, and the classification accuracy and loss value of the model on the source domain validation set are recorded; When the classification accuracy of the model on the source domain validation set does not improve for a preset number of times in succession, the pre-training is stopped, and the current network weight parameters are saved as initial weight parameters; Selecting a flue gas image data set of a waste treatment area as a target domain data set, and performing data enhancement processing on the target domain data set to generate an expanded target domain data set; The initial weight parameters are loaded into the basic classification network, the first network layer weight parameters of the feature extraction subnet are frozen, and the second network layer weight parameters of the feature extraction subnet and all network layer weight parameters of the feature fusion subnet and the classification output subnet are unfrozen; The expanded target domain data set is used to fine-tune the unfrozen basic classification network, an initial learning rate is adopted, and the performance indicators of the target domain validation set are monitored to adjust the training process; During the fine-tuning training process, when the loss value of the target domain validation set increases for a preset number of times in succession, the learning rate is reduced for continuous training; When the classification accuracy of the target domain validation set reaches a preset accuracy, the fine-tuning training is stopped, the corresponding network weight parameters are saved, and a classification network optimized through migration learning is obtained; In the classification network, the feature association modeling layer models the association relationship between the fusion feature vectors and the pollutant components to generate a pollutant component association matrix, including: The fusion feature vector is input into the feature association modeling layer, the fusion feature vector is mapped to a feature space of a first preset dimension through a first full connection layer to generate a first feature vector; The first feature vector is input into a second full connection layer and mapped to a feature space of a second preset dimension to generate a second feature vector, and the second preset dimension corresponds to a preset number of pollutant component categories; The second feature vector is transposed to generate a transposed vector of the second feature vector; The matrix product of the second feature vector and the transposed vector is calculated to generate an initial association matrix, and the number of rows and the number of columns of the initial association matrix are both the second preset dimension; The initial association matrix is subjected to row standardization processing so that the sum of the elements in each row of the matrix is 1 to generate a standardized association matrix; A pollutant component priori knowledge matrix is constructed, the pollutant component priori knowledge matrix is constructed based on known pollutant coexistence relationships, and the matrix elements represent the probability of the simultaneous occurrence of two pollutant components; The standardized association matrix and the pollutant component priori knowledge matrix are multiplied element by element to generate a weighted association matrix, and the weighted association matrix fuses feature association information and prior knowledge driven by data; The weighted association matrix is subjected to nonlinear transformation, the expression of the strong association relationship is enhanced through an activation function layer, and a pollutant component association matrix is generated, and the element in the ith row and jth column of the pollutant component association matrix represents the association strength between the ith pollutant component and the jth pollutant component. 9.A garbage discharge recognition system applying a deep learning model, characterized by, It comprises a processor and a memory, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of the method of any one of claims 1-8. It comprises a processor and a memory, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of the method of any one of claims 1-8.

Images