A method and system for identifying dust raising of pumped storage power station based on deep tree ensemble

Abstract

The application provides a pumped storage power station dust raising identification method and system based on deep tree integration, and belongs to the technical field of dust raising identification. The technical problem to be solved is to provide an improved pumped storage power station dust raising identification method based on deep tree integration. The technical solution adopted to solve the above technical problem comprises the following steps: image acquisition preprocessing: obtaining the image of the pumped storage power station from a network camera with high-definition imaging function, and preprocessing the image; adopting multi-granularity scanning to obtain the dimension-reduced image feature information; training the deep tree integration algorithm by using the dimension-reduced image feature information until the accuracy of a certain layer no longer improves, and then stopping the training to obtain an identification model; inputting the preprocessed test sample, adopting multi-granularity scanning to obtain the dimension-reduced image feature information, using the trained identification model to identify the test image feature information, and obtaining an identification result; and the application is applied to dust raising identification.

Description

Technical Field

[0001] This invention provides a method and system for dust identification in pumped storage power stations based on deep tree ensemble, belonging to the field of dust identification technology. Background Technology

[0002] In pumped-storage power station environments, air quality monitoring is a fundamental task that requires special attention. Air quality in these environments is crucial not only for the health and safety of workers but also for the protection of power station equipment, which may be damaged by particulate or fibrous dust. Current dust identification methods primarily rely on deep neural networks, but these models have numerous training parameters, are complex, and are difficult and time-consuming to train and optimize. Therefore, this invention proposes a method and system for dust identification in pumped-storage power stations based on deep tree ensembles. Summary of the Invention

[0003] In order to overcome the shortcomings of the prior art, the technical problem to be solved by the present invention is to provide an improved method for dust identification in pumped storage power stations based on deep tree integration.

[0004] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows: a method for identifying dust pollution in pumped storage power stations based on deep tree ensemble, comprising the following steps: S1: Image acquisition and preprocessing: Images of the pumped storage power station are acquired from a network camera with high-definition imaging capabilities. The images are preprocessed, converted into binary grayscale images, and the samples are divided into training sample set and test sample set. S2: Employ multi-granularity scanning to acquire the dimension-reduced image feature information; S3: Use the image feature information after dimensionality reduction to train the deep tree ensemble algorithm until the accuracy of a certain layer no longer improves, then stop training to obtain the recognition model; S4: Input the preprocessed test sample, use multi-granularity scanning to obtain the dimensionality-reduced image feature information, use the trained recognition model to recognize the test image feature information, and obtain the recognition result.

[0005] The image preprocessing steps in step S1 are as follows: S1.1: Convert each frame of the image to a grayscale image; S1.2: Histogram specification of the grayscale image: S1.2.1: First, perform histogram equalization on the grayscale image to obtain the cumulative distribution T(s) for each pixel s; S1.2.2: Define the histogram according to the requirements and find the cumulative distribution G(Z); S1.2.3: For each T(s) (assuming its pixel value is ss), find the G(Z) value with the smallest difference from it (assuming the corresponding pixel value is zz), then after specification, transform ss into zz; S1.3: Use a low-pass filter to remove the image background; S1.4: Binarize the grayscale image after histogram specification and filtering; S1.5: Use a median filter to eliminate residual noise.

[0006] The preprocessed image is scaled and divided into square image blocks with a set side length of pixels. The extraction of the square images starts from the top left corner and scans the entire image, arranging them horizontally block by block, with a step size of one-tenth of the set pixels. Then, the images are labeled with two types of labels: those with dust and those without dust.

[0007] The multi-granularity scanning steps in step S2 are as follows: First, the preprocessed square image patch is used as input to GBDT and Xgboost to extract new features F. GBDT and F Xgboost Meanwhile, the square image patch is sampled through a sampling window to obtain feature subsamples. Then, GBDT and Xgboost are trained on each collected feature subsample, and a probability vector of length 2 is generated each time. After training, GBDT and Xgboost produce a predicted feature vector O of length 2 pixels. GBDT and O Xgboost , will F GBDT F Xgboost O GBDT O Xgboost These five feature vectors, along with the original feature X of the square image patch, are combined to obtain a new combined feature F. GBDT +F Xgboost +O GBDT +O Xgboost +X, the new features are reduced in dimensionality using PCA to obtain the dimensionality-reduced feature vector F. PCA Its dimension is V PCA .

[0008] The steps for obtaining the recognition model in step S3 are as follows: The deep tree ensemble algorithm comprises a cascaded module consisting of multiple cascaded deep trees. The input of the first layer in the deep tree cascaded module is the feature vector F output by the multi-granularity scanning module. PCAAfter two GBDT and two Xgboost classification processes, four two-dimensional class vectors are obtained. Then, the four two-dimensional category vectors are compared with the feature vector output by the multi-granularity scanning module. PCA splicing produces a [V] PCA A 2×4 dimensional feature vector is used as the input to the second layer. And so on, the (N-1)th layer will generate [V] PCA A new feature vector of dimension +2×4 is used as the input to the Nth layer; Finally, the average value of the category vectors output by the Nth layer is calculated, and the category corresponding to the maximum value is selected as the final classification result of the dust map.

[0009] A dust identification system for pumped-storage power stations based on deep tree ensemble is characterized by comprising an image acquisition module, an image preprocessing module, a multi-granularity scanning module, and a deep tree cascade module. The image acquisition module acquires images of the pumped-storage power station through a high-resolution camera. The image preprocessing module performs grayscale, binarization, noise reduction, and scaling preprocessing on the acquired images. The multi-granularity scanning module performs dimensionality reduction processing on the preprocessed images using GBDT and XGBoost. The deep tree cascade module trains the system using the dimensionality-reduced feature information until the accuracy of a certain layer no longer improves, at which point training stops, resulting in a recognition model.

[0010] The image preprocessing module performs the following steps for image preprocessing: Convert each frame of the image to a grayscale image; Histogram specification for grayscale images: First, histogram equalization is performed on the grayscale image to obtain each pixel s and the cumulative distribution T(s); Determine the cumulative distribution G(Z) by standardizing the histogram as required; For each T(s) (assuming its pixel value is ss), find the G(Z) value with the smallest difference from it (assuming the corresponding pixel value is zz), then after specification, ss is transformed into zz; Use a low-pass filter to remove the image background; Binarize the grayscale image after histogram specification and filtering; Use a median filter to eliminate residual noise; The preprocessed image is scaled and divided into square image blocks with a set side length of pixels. The extraction of the square images starts from the top left corner and scans the entire image, arranging them horizontally block by block, with a step size of one-tenth of the set pixels. Then, the images are labeled with two types of labels: those with dust and those without dust.

[0011] The steps of the multi-granularity scanning module for dimensionality reduction of image feature information are as follows: First, the preprocessed square image patch is used as input to GBDT and Xgboost to extract new features F. GBDT and F Xgboost Meanwhile, the square image patch is sampled through a sampling window to obtain feature subsamples. Then, GBDT and Xgboost are trained on each collected feature subsample, and a probability vector of length 2 is generated each time. After training, GBDT and Xgboost produce a predicted feature vector O of length 2 pixels. GBDT and O Xgboost , will F GBDT F Xgboost O GBDT O Xgboost These five feature vectors, along with the original feature X of the square image patch, are combined to obtain a new combined feature F. GBDT +F Xgboost +O GBDT +O Xgboost +X, the new features are reduced in dimensionality using PCA to obtain the dimensionality-reduced feature vector F. PCA Its dimension is V PCA .

[0012] The structure of the deep tree cascade module includes: The input to the first layer of the deep tree cascade module is the feature vector F output by the multi-granularity scan module. PCA After two GBDT and two Xgboost classification processes, four two-dimensional class vectors are obtained; The four two-dimensional category vectors are compared with the feature vectors output by the multi-granularity scanning module. PCA splicing produces a [V] PCA A 2×4 dimensional feature vector is used as the input to the second layer. And so on, the (N-1)th layer will generate [V] PCA A new feature vector of dimension +2×4 is used as the input to the Nth layer; The average value of the category vectors output from the Nth layer is calculated, and the category corresponding to the maximum value of the average value is selected as the final classification result of the dust map.

[0013] The advantages of this invention compared to existing technologies are as follows: The new method for detecting and analyzing dust from pumped storage power stations using a deep tree ensemble model provided by this invention relies on a deep tree ensemble classification algorithm in the field of pattern recognition. This algorithm can not only automatically select features and discover some features that are difficult to discover manually, but also generate representation vectors of high-order features for training. Moreover, compared with deep neural networks, the proposed method is easier to optimize, has a faster training and testing speed, and better generalization performance. This method helps to design effective countermeasures to reduce the harmful effects of dust. Attached Figure Description

[0014] The present invention will be further described below with reference to the accompanying drawings: Figure 1 This is a flowchart of the method of the present invention; Figure 2 This is a schematic diagram of the image acquisition module of the present invention; Figure 3 This is a flowchart of the image preprocessing process of the present invention; Figure 4 This is a schematic diagram of the structure of the multi-granularity scanning module of the present invention; Figure 5 This is a schematic diagram of the structure of the deep tree cascade module of the present invention. Detailed Implementation

[0015] like Figure 1-5 As shown, the dust identification method for pumped storage power stations based on deep tree ensemble provided by this invention mainly includes: 1) Obtain images of the pumped storage power station from a network camera with high-definition imaging capabilities, preprocess the images to convert them into binary grayscale images, and divide the samples into training sample sets and test sample sets according to the needs of the algorithm.

[0016] 2) Multi-granularity scanning is used to obtain the image feature information after dimensionality reduction.

[0017] 3) Use the reduced feature information to train the deep tree ensemble algorithm until the accuracy of a certain layer no longer improves, then stop training to obtain the recognition model.

[0018] 5) Input the test sample, use multi-granularity scanning to obtain the dimensionality-reduced feature information, use the trained recognition model to recognize the feature information of the test image, and obtain the recognition result.

[0019] The structure of the image acquisition module of the present invention is as follows: Figure 2As shown, a webcam with high-definition imaging capability is used. To prevent dust from settling on the sensor surface and to avoid dust accumulation and overlapping, the webcam is placed at a 45-degree angle to the ground. Additionally, to ensure stable image acquisition under any lighting conditions, a light-emitting diode (LED) is placed vertically to the webcam to enhance image contrast. The webcam connects to a PC via a USB interface. The PC can acquire the entire data stream or single frames, and the acquired images are saved on the PC in uncompressed JPEG format.

[0020] The image preprocessing module of this invention preprocesses the images acquired by the acquisition module, improving the focus and contrast of the images and eliminating image noise. The preprocessing steps are as follows: Figure 3 As shown. Specifically includes: Step 1: To reduce the computational burden, convert each frame of the image to a grayscale image.

[0021] Step 2: Histogram specification is applied to the grayscale image to enhance contrast and make details more clearly visible. The specific steps are as follows: (1) First, perform histogram equalization on the original image to obtain the cumulative distribution T(s) for each pixel s; (2) Define the histogram according to the required specifications and find the cumulative distribution G(Z); (4) For each T(s) (assuming its pixel value is ss), find the G(z) value with the smallest difference in G(Z) (assuming the corresponding pixel value is zz), then after standardization, transform ss into zz.

[0022] Step 3: Use a low-pass filter to remove the background from the image. The cutoff frequency of the low-pass filter is obtained through Fourier transform amplitude spectrum analysis of the image.

[0023] Step 4: In order to eliminate some of the noise generated during the histogram specification process, the grayscale image after histogram specification and filtering is binarized.

[0024] Step 5: Use a median filter to eliminate residual noise.

[0025] The output of the preprocessing stage is a binary image, where dust is indicated by 0 (black) and non-dust is indicated by 1 (white).

[0026] After image preprocessing, training samples were determined. The preprocessed image was scaled to 500×500. To obtain more training and test sets, the image was divided into 100×100 square image blocks with sides of 100 pixels. These blocks were extracted by scanning the entire image from the top left corner, arranging them horizontally block by block with a step size of 10 pixels. Image labels were manually added. There were two types of labels: with dust and without dust. The labeled samples were then divided into training and test sets.

[0027] Multi-granularity scanning of training samples achieves dimensionality reduction of feature information. The multi-granularity scanning module, such as... Figure 4 As shown: The multi-granularity scanning module first uses the preprocessed 100×100 dimensional image as input to GBDT and Xgboost to generate new features. and Simultaneously, the 100×100 dimensional image is sampled through a 10×10 sampling window to obtain 100 feature subsamples. GBDT and Xgboost then train on each of these subsamples, generating a probability vector of length 2 for each training iteration. After training, GBDT and Xgboost will produce a feature vector O of length 100×2. GBDT and O Xgboost F GBDT F Xgboost O GBDT O Xgboost These five feature vectors, along with the original feature X of the 100×100 image, are combined to obtain new combined features. The new feature is then subjected to PCA for dimensionality reduction to obtain the dimensionality-reduced feature vector. Its dimension is Among them, F GBDT For features extracted using GBDT, F Xgboost To use the features extracted by Xgboost, O GBDT O is a feature composed of prediction results generated using GBDT. Xgboost These are the features that comprise the prediction results generated using Xgboost.

[0028] The deep tree ensemble algorithm is trained using the dimensionality-reduced feature information, and the structure diagram of the deep tree cascade module is shown below. Figure 5 As shown, in the cascaded module, except for the first layer which uses the feature vector F output by the multi-granularity scanning module, PCA In addition to serving as input, each subsequent layer combines the feature vector output from the previous layer with the feature vector F output from the multi-granularity scanning module. PCAThe concatenation is used as the input to this layer. The cascade module first takes the feature vector F_PCA output from the multi-granularity scanning module as input, and after two GBDT and two Xgboost classification processes, obtains four two-dimensional class vectors; then, these four two-dimensional class vectors are concatenated with the feature vector F output from the multi-granularity scanning structure. PCA splicing produces a [V] PCA A feature vector of dimension [+2×4] is generated, which serves as the input to the second layer; and so on, the (N-1)th layer will produce [V] dimensional feature vectors. PCA A new feature vector of dimension +2×4 is used as the input to the Nth layer; finally, the average value of the category vectors output by the Nth layer is calculated, and the category corresponding to the maximum value is selected as the final classification result of the dust map.

[0029] Regarding the specific structure of this invention, it should be noted that the connection relationships between the various components and modules used in this invention are definite and achievable. Except as specifically described in the embodiments, their specific connection relationships can bring about corresponding technical effects and solve the technical problems proposed by this invention without relying on the execution of corresponding software programs. Unless otherwise specifically described, the models and connection methods of the components, modules, and specific parts appearing in this invention are all prior art such as published patents, published journal articles, or common knowledge that can be obtained by those skilled in the art before the application date, and need not be elaborated. This makes the technical solution provided in this case clear, complete, and achievable, and can reproduce or obtain corresponding physical products based on this technical means.

[0030] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for identifying dust pollution in pumped-storage power stations based on deep tree ensemble, characterized in that: Includes the following steps: S1: Image acquisition and preprocessing: Images of the pumped storage power station are acquired from a network camera with high-definition imaging capabilities. The images are preprocessed, converted into binary grayscale images, and the samples are divided into training sample set and test sample set. S2: Employ multi-granularity scanning to acquire the dimension-reduced image feature information; S3: Use the image feature information after dimensionality reduction to train the deep tree ensemble algorithm until the accuracy of a certain layer no longer improves, then stop training to obtain the recognition model; S4: Input the preprocessed test sample, use multi-granularity scanning to obtain the dimensionality-reduced image feature information, use the trained recognition model to recognize the test image feature information, and obtain the recognition result; The image preprocessing steps in step S1 are as follows: S1.1: Convert each frame of the image to a grayscale image; S1.2: Histogram specification of the grayscale image: S1.2.1: First, perform histogram equalization on the grayscale image to obtain the cumulative distribution T(s) for each pixel s; S1.2.2: Define the histogram according to the requirements and find the cumulative distribution G(Z); S1.2.3: For each T(s), its pixel value is ss. Find the G(Z) value with the smallest difference from it in G(Z), and the corresponding pixel value is zz. Then, after standardization, ss is transformed into zz. S1.3: Use a low-pass filter to remove the image background; S1.4: Binarize the grayscale image after histogram specification and filtering; S1.5: Use a median filter to eliminate residual noise; The preprocessed image is scaled and divided into square image blocks with a set side length of pixels. The extraction of the square images starts from the top left corner and scans the entire image, arranging them horizontally block by block, with a step size of one-tenth of the set pixels. Then, the images are labeled with two types of labels: those with dust and those without dust. The multi-granularity scanning steps in step S2 are as follows: First, the preprocessed square image patch is used as input to GBDT and Xgboost to extract new features F. GBDT and F Xgboost Meanwhile, the square image patch is sampled through a sampling window to obtain feature subsamples. Then, GBDT and Xgboost are trained on each collected feature subsample, and a probability vector of length 2 is generated each time. After training, GBDT and Xgboost produce a predicted feature vector O of length 2 pixels. GBDT and O Xgboost , will F GBDT F Xgboost O GBDT O Xgboost These five feature vectors, along with the original feature X of the square image patch, are combined to obtain a new combined feature F. GBDT +F Xgboost +O GBDT +O Xgboost +X, the new features are reduced in dimensionality using PCA to obtain the dimensionality-reduced feature vector F. PCA Its dimension is V PCA .

2. The method for dust identification in pumped storage power stations based on deep tree ensemble as described in claim 1, characterized in that: The steps for obtaining the recognition model in step S3 are as follows: The deep tree ensemble algorithm comprises a cascaded module consisting of multiple cascaded deep trees. The input of the first layer in the deep tree cascaded module is the feature vector F output by the multi-granularity scanning module. PCA After two GBDT and two Xgboost classification processes, four two-dimensional class vectors are obtained; Then, the four two-dimensional category vectors are compared with the feature vector F output by the multi-granularity scanning module. PCA splicing produces a [V] PCA A 2×4 dimensional feature vector is used as the input to the second layer. And so on, the (N-1)th layer will generate [V] PCA A new feature vector of dimension +2×4 is used as the input to the Nth layer; Finally, the average value of the category vectors output by the Nth layer is calculated, and the category corresponding to the maximum value is selected as the final classification result of the dust map.

3. A dust identification system for pumped storage power stations based on deep tree ensemble, characterized in that: To implement the dust identification method for pumped storage power stations based on deep tree ensemble as described in claim 1, the system includes an image acquisition module, an image preprocessing module, a multi-granularity scanning module, and a deep tree cascade module. The image acquisition module acquires images of the pumped storage power station through a high-resolution camera. The image preprocessing module performs grayscale, binarization, noise reduction, and scaling preprocessing on the acquired images. The multi-granularity scanning module performs dimensionality reduction processing on the preprocessed images using GBDT and XGBoost for image feature information. The deep tree cascade module trains the system using the dimensionality-reduced feature information until the accuracy of a certain layer no longer improves, at which point training stops, resulting in an identification model.

4. The dust identification system for pumped storage power stations based on deep tree ensemble as described in claim 3, characterized in that: The structure of the deep tree cascade module includes: The input to the first layer of the deep tree cascade module is the feature vector F output by the multi-granularity scan module. PCA After two GBDT and two Xgboost classification processes, four two-dimensional class vectors are obtained; Combine the four two-dimensional category vectors with the feature vector F output by the multi-granularity scanning module. PCA splicing produces a [V] PCA A 2×4 dimensional feature vector is used as the input to the second layer. And so on, the (N-1)th layer will generate [V] PCA A new feature vector of dimension +2×4 is used as the input to the Nth layer; The average value of the category vectors output from the Nth layer is calculated, and the category corresponding to the maximum value of the average value is selected as the final classification result of the dust map.

Images