Coal-fired power plant abnormal situation identification method based on multi-scale features

Abstract

The method for identifying abnormal conditions of coal-fired power plants based on multi-scale features belongs to the field of multi-scale features and abnormal condition identification. The method comprises the following steps: (1) constructing an image training set, (2) constructing a multi-scale feature abnormal condition identification network for coal-fired power plants, (3) training a deep network, and (4) detecting abnormal condition images of coal-fired power plants. This method not only reduces the missed detection of small targets or misjudgment of background interference, improves the identification accuracy of abnormal conditions, but also significantly improves the deployment adaptability and actual efficiency of the model on the edge platform.

Description

Technical Field

[0001] This invention relates to a method for identifying abnormal conditions in coal-fired power plants based on multi-scale features, and belongs to the field of multi-scale features and abnormal condition identification. Background Technology

[0002] In typical complex industrial environments like coal-fired power plants, plant monitoring not only needs to accurately identify abnormal behaviors such as personnel falling, unauthorized entry, and not wearing safety helmets, but also needs to effectively identify various equipment and environmental anomalies, including coal conveyor fires, coal and dust accumulation, ground water accumulation, cable tray damage, and insulation detachment. These multi-target, multi-scenario perception tasks place extremely high demands on the real-time performance and accuracy of the identification algorithms.

[0003] However, coal-fired power plants typically operate under harsh conditions such as high temperatures, high dust levels, and strong electromagnetic interference. Plant monitoring often relies on edge computing devices to run algorithms, which have extremely limited computing and storage resources. This makes traditional deep learning recognition models difficult to deploy directly due to their large number of parameters and high computational complexity, thus failing to meet real-time inspection needs. Furthermore, issues such as fluctuating lighting, occlusion interference, weather conditions, and diverse human behavior patterns in open environments further challenge traditional visual recognition methods. These methods often rely on handcrafted features and fixed rules, making them ill-suited to complex environmental changes and behavioral diversity. This results in insufficient robustness and generalization ability of the recognition models, failing to meet the accuracy and reliability requirements of real-world security scenarios.

[0004] Therefore, given the prominent contradiction between limited hardware resources at the edge and the need for perception in complex scenarios, how to improve the adaptability and recognition accuracy of algorithms in changing environments while ensuring real-time recognition has become a key research direction for promoting the practical application and maturity of intelligent inspection systems for power plants.

[0005] Extensive research has revealed that the YOLOv7-tiny model method, which is similar to this invention, is as follows:

[0006] While the YOLOv7-tiny model performs well in inference efficiency, its recognition performance remains significantly insufficient in open and complex industrial scenarios such as coal-fired power plants. Limited by its lightweight structure, the model has relatively weak feature representation capabilities, making it difficult to fully capture multi-scale and subtle target features in complex environments. Especially when facing real-world scenarios with drastic lighting changes, heavy dust interference, partial equipment occlusion, and various irregular structures, YOLOv7-tiny often exhibits problems such as insufficient feature extraction and loss of semantic information. This leads to a significant decrease in the accuracy of anomaly recognition and is prone to missing small targets or misjudging background interference, seriously affecting the reliability and practicality of intelligent inspection systems. Summary of the Invention

[0007] To improve the accuracy of anomaly identification and reduce the occurrence of missed detections of small targets or misjudgments of background interference, this invention proposes a multi-scale feature-based anomaly identification method for coal-fired power plants. This invention enhances the model's ability to represent and identify abnormal behavior features in complex scenarios by performing targeted network structure optimization. Simultaneously, the optimization process strictly controls the model parameter scale and computational complexity to ensure that the improved model maintains its lightweight characteristics, meeting the dual requirements of real-time performance and hardware resource adaptability for plant monitoring.

[0008] To achieve the aforementioned objectives and address the problems existing in the prior art, the technical solution adopted by this invention is: a method for identifying abnormal conditions in coal-fired power plants based on multi-scale features, characterized in that:

[0009] Step 1: Construct an image database of abnormal situations in coal-fired power plants, covering abnormal situations such as unauthorized crossing of dangerous areas, power plant personnel falling, and not wearing safety helmets. A total of 11,502 original images were collected and divided into training set: test set: validation set in a ratio of 8:1:1.

[0010] Step 2: Construct an anomaly identification network for coal-fired power plants based on multi-scale features. This network mainly includes a backbone network, a neck network, and a detection head, and specifically includes the following sub-steps:

[0011] (1) Input the abnormal situation image of the coal-fired power plant into the backbone network. The backbone network consists of four parts: four dual-branch feature extraction subnetworks DFES1, DFES2, DFES3 and DFES4, which are connected in sequence.

[0012] ① The DFES1 input is a 640×640×3 image of an anomaly in a coal-fired power plant, containing one RepVGG1_1, one RepVGG1_2, one dual-branch module Dufex1_1, and one convolutional module Conv1_1, all connected in series. The RepVGG structures are identical throughout this invention. The RepVGG1_1 network adopts a training-inference separation design. During the training phase, the RepVGG network structure includes four sets of convolutional units GConv1, GConv2, GConv3, and GConv4, connected in series. GConv1 contains one dual-branch module GC_Dufex1 and one ReLU function, also connected in series. GC_Dufex1 contains two parallel branches, each containing a 3×3 convolutional kernel with a stride of 2, and the features from the left and right branches are concatenated. GConv2, GConv3, and GConv4 have identical structures, each containing one three-branch module and one ReLU function, connected in series. The three-branch module comprises three parallel branches: the left branch contains a 3×3 convolutional kernel with a stride of 2; the middle branch contains a 1×1 convolutional kernel with a stride of 2; and the right branch contains the input of this module. The three branches concatenate features. During inference, the RepVGG network structure includes four sets of convolutional units: GConv5, GConv6, GConv7, and GConv8, connected sequentially. These four sets of convolutional units (GConv5, GConv6, GConv7, and GConv8) have identical structures, each including a 3×3 convolutional kernel with a stride of 2 and a ReLU function, connected sequentially.

[0013] Dufex1_1 contains two parallel branches. The left branch contains one convolutional module, Conv1_2; the right branch contains one convolutional module, Conv1_3, and two feature extraction modules, RepVGG1_3 and RepVGG1_4, which are connected sequentially. The left and right branches perform feature concatenation. Specifically, Conv1_2 has a kernel size of 1×1, a stride of 1, valid padding, and 32 output channels; Conv1_3 also has a kernel size of 1×1, a stride of 1, valid padding, and 32 output channels.

[0014] The parameters of the convolution module Conv1_1 are a kernel size of 1×1, a stride of 1, valid padding, and 64 output channels. The DFES1 module outputs a single feature IM1_1 with dimensions of 160×160×64.

[0015] ② The input to DFES2 is IM1_1, which contains one Maxpool1_1 module, one dual-branch module Dufex1_2, and one convolutional module Conv1_4, connected in series. Dufex1_2 contains two parallel branches: the left branch contains one convolutional module Conv1_5; the right branch contains one convolutional module Conv1_6 and two feature extraction modules RepVGG1_5 and RepVGG1_6, connected in series; the left and right branches perform feature concatenation. Maxpool1_1 has a 3×3 maxpooling kernel, a stride of 2, same padding, and 64 output channels; Conv1_5 has a 3×3 kernel, a stride of 1, same padding, and 64 output channels; Conv1_6 has a 3×3 kernel, a stride of 1, same padding, and 64 output channels.

[0016] The parameters for the convolutional module Conv1_4 are a kernel size of 1×1, a stride of 1, valid padding, and 128 output channels. The DFES2 module outputs a single feature IM1_2 with dimensions of 80×80×128.

[0017] ③ The DFES3 input is IM1_2, which contains one Maxpool1_2, one dual-branch module Dufex1_3, and one convolutional module Conv1_7, connected in series. Dufex1_3 contains two parallel branches: the left branch contains one convolutional module Conv1_8; the right branch contains one convolutional module Conv1_9 and two feature extraction modules RepVGG1_7 and RepVGG1_8, connected in series; the left and right branches perform feature concatenation. Maxpool1_2 has a 3×3 maxpooling kernel, a stride of 2, same padding, and 128 output channels; Conv1_8 has a 3×3 kernel, a stride of 1, same padding, and 128 output channels; Conv1_9 has a 3×3 kernel, a stride of 1, same padding, and 128 output channels.

[0018] The parameters for the convolution module Conv1_7 are a kernel size of 1×1, a stride of 1, valid padding, and 256 output channels. The DFES3 module outputs a single feature IM1_3 with dimensions of 40×40×256.

[0019] ④ The DFES4 input is IM1_3, which contains one Maxpool1_3 module, one dual-branch module Dufex1_4, and one convolutional module Conv1_10, connected in series. Dufex1_4 contains two parallel branches: the left branch contains one convolutional module Conv1_11; the right branch contains one convolutional module Conv1_12 and two feature extraction modules RepVGG1_9 and RepVGG1_10, connected in series; the left and right branches perform feature concatenation. Maxpool1_3 has a 3×3 maxpooling kernel, a stride of 2, same padding, and 256 output channels; Conv1_11 has a 3×3 convolutional kernel, a stride of 1, same padding, and 256 output channels; Conv1_12 has a 3×3 convolutional kernel, a stride of 1, same padding, and 256 output channels.

[0020] The parameters for the convolution module Conv1_10 are a kernel size of 1×1, a stride of 1, valid padding, and 512 output channels. The DFES4 module outputs a single feature IM1_4 with dimensions of 20×20×512.

[0021] (2) The main channel feature IM1_4 output from the backbone network is input into the neck network. The neck network consists of 5 sets of convolutional processing units CP1, CP2, CP3, CP4 and CP5, which are connected in sequence.

[0022] ① CP1 consists of one main branch MB2_1 and one side branch LB2_1, with features from the main and side branches concatenated. MB2_1 takes IM1_4 as input and contains one SPP module, one convolutional module Conv2_1, and one upsample2_1 module, connected sequentially. LB2_1 takes IM1_3 as input and contains one convolutional module Conv2_2. The SPP module outputs a 20×20×256 feature IM2_1, which is the input to Conv2_1. Conv2_1 has a kernel size of 1×1, a stride of 1, valid padding, and 128 output channels. The upsample2_1 module outputs 128 channels. Conv2_2 has a kernel size of 1×1, a stride of 1, valid padding, and 128 output channels. The CP1 unit outputs a 40×40×256 feature IM2_2.

[0023] ② CP2 consists of one main branch MB2_2 and one side branch LB2_2, with features concatenated from the main and side branches. MB2_2's input is IM2_2, containing one RCS-OSA2_1 module, one convolutional module Conv2_3, and one Upsample2_2 module, connected serially. LB2_2's input is IM1_2, containing one convolutional module Conv2_4. The RCS-OSA modules in this invention all have the same structure. The RCS-OSA2_1 module contains one RepVGG2_1 network, one dual-branch module Dufex2_1, and one RepVGG2_2 network, connected serially. In Dufex2_1, the left branch is the output of RepVGG2_1; the right branch contains n / 2 RCS modules 2_1 and one dual-branch module Dufex2_2, connected serially. In Dufex2_2, the left branch is the output of n / 2 RCS modules 2_1; the right branch is the output of n / 2 RCS modules 2_2, which are connected in series. The output features of RepVGG2_1, n / 2 RCS modules 2_1, and n / 2 RCS modules 2_2 are channel-washed and then concatenated; where n is a positive integer and is the number of RCS modules in this stacked structure.

[0024] The RCS-OSA2_1 module outputs a 40×40×128 feature IM2_3, which serves as the input to Conv2_3. Conv2_3 parameters include a 1×1 kernel size, a stride of 1, valid padding, and 64 output channels. The Upsample2_2 module outputs 64 channels; Conv2_4 parameters include a 1×1 kernel size, a stride of 1, valid padding, and 64 output channels. The CP2 unit outputs an 80×80×128 feature IM2_4.

[0025] ③ CP3 contains one main branch MB2_3 and one side branch LB2_3, with features from the main and side branches concatenated. The input to MB2_3 is IM2_4, which contains one RCS-OSA2_2 module and one convolutional module Conv2_5, connected sequentially; the output of LB2_3 is IM2_3. The RCS-OSA2_2 module outputs a feature IM2_5 with dimensions 80×80×64, which serves as the input to Conv2_5. The Conv2_5 parameters are a 3×3 kernel size, a stride of 2, same padding, and 128 output channels. The CP3 unit outputs a feature IM2_6 with dimensions 40×40×256.

[0026] ④ CP4 consists of one main branch MB2_4 and one side branch LB2_4, with features from the main and side branches concatenated. MB2_4 takes IM2_6 as input and contains one RCS-OSA2_3 module and one convolutional module Conv2_6, connected sequentially. LB2_4 is the output of IM2_1. The RCS-OSA2_3 module outputs a 40×40×128 feature IM2_7, which is the input to Conv2_6. Conv2_6 has a kernel size of 3×3, a stride of 2, same padding, and 256 output channels. The CP4 unit outputs a 20×20×512 feature IM2_8.

[0027] ⑤ The input to CP5 is IM2_8, which contains one RCS-OSA2_4 module, one Maxpool2_1 module, and one E_ELAN module, connected serially. The RCS-OSA2_4 module outputs a feature IM2_9 with dimensions 20×20×256, which is the input to Maxpool2_1. The parameters of Maxpool2_1 are a 3×3 max-pooling kernel, a stride of 2, same padding, and 256 output channels. The E_ELAN module also outputs 256 channels. The CP5 unit outputs a feature IM2_10 with dimensions 10×10×256.

[0028] (3) The five features output by the backbone network (IM1_1) and the neck network (IM2_5, IM2_7, IM2_9, and IM2_10) are input into the detection head network. The detection head network contains five TSCODE modules connected in parallel. In this invention, the TSCODE modules have the same structure. The inputs of the five TSCODE modules correspond to the five features mentioned above. Each TSCODE module contains a dual-branch structure, with the upper branch containing a semantic context coding unit (SCE) and the lower branch containing a detail-preserving coding unit (DPE), outputting in parallel. The SCE contains a main branch (MB3_1) and a side branch (LB3_1), and the features of the main and side branches are concatenated. The input of MB3_1 is the feature map of the current layer. It contains one Downsample3_1 module; the output of LB3_1 is the feature map of the next layer. The DPE consists of one main branch MB3_2, one side branch LB3_2, and one side branch LB3_3. The features from the main and side branches are concatenated. The output of MB3_2 is the feature map of the current layer. Side branch 3_2 input is the feature map of the next layer. It contains one Upsample3_1; the side branch 3_3 takes the feature map from the previous layer as its input. It contains one splicing module and one Down sample3_2 module, connected serially. The splicing module includes one Up sample3_2 module, specifically... After processing with Up sample3_2 and Perform feature splicing.

[0029] Step 3: Train the deep network. Use the training set built in Step 1 to train the deep network built in Step 2, and obtain the network parameters and model.

[0030] Step 4: Perform abnormal situation image recognition of coal-fired power plants. Use the abnormal situation recognition network of coal-fired power plants obtained in Step 3 to recognize the newly collected abnormal situation images of coal-fired power plants and obtain the abnormal situation recognition results.

[0031] The beneficial effects of this invention are: a method for identifying abnormal conditions in coal-fired power plants based on multi-scale features, comprising the following steps: (1) constructing an image training set, (2) constructing a multi-scale feature network for identifying abnormal conditions in coal-fired power plants, (3) training a deep network, and (4) performing image recognition of abnormal conditions in coal-fired power plants. Compared with existing technologies, the advantages of this invention are as follows: A RepVGG module with a superior structure is introduced into the backbone network. This module adopts a multi-branch structure during training and a single-branch structure during inference. During training, multi-branch convolution enhances feature representation capabilities, effectively improving the problems of insufficient feature extraction and semantic information loss in the original model. During inference, it is fused into a single convolution, maintaining high efficiency, thereby improving the model's ability to perceive subtle targets and complex textures without sacrificing speed. An RCS-OSA module is added to the neck structure. This module achieves effective fusion and information reorganization of multi-scale features, enhancing the network's adaptability to targets of different sizes and irregular structures in coal-fired power plants, and helping to reduce missed detections and misjudgments caused by changes in target scale or dust interference. In the detection head, a TSCODE decoupling structure is introduced. By decoupling classification and regression tasks at the feature level, the feature conflict between the two tasks in occluded scenarios is alleviated, improving detection stability and accuracy under conditions such as partial occlusion of equipment and sudden changes in illumination. In summary, these improvements collectively enhance the model's feature extraction, multi-scale adaptation, and occlusion resistance capabilities in the complex and open environment of coal-fired power plants, thereby significantly improving recognition accuracy and system reliability while maintaining high inference efficiency suitable for edge deployment. Attached Figure Description

[0032] Figure 1 These are some of the original images collected.

[0033] Figure 2 This is a schematic diagram of a network for identifying abnormal situations in coal-fired power plants based on multi-scale features.

[0034] Figure 3This is a diagram of the multi-branch residual structure during RepVGG training.

[0035] Figure 4 This is a diagram of the VGG-style single-branch structure during RepVGG inference.

[0036] Figure 5 This is the RCS-OSA structure diagram.

[0037] Figure 6 This is the SCE structure diagram.

[0038] Figure 7 This is a structural diagram of DPE.

[0039] Figure 8 These are accuracy test charts for eight categories using different models.

[0040] Figure 9 These are diagrams showing the effectiveness of different models in identifying abnormal situations in coal-fired power plants. Detailed Implementation

[0041] The present invention will be further described below with reference to the accompanying drawings. The present invention will be specifically described below through embodiments. It should be noted that the following embodiments are only for further illustration of the present invention and are not limiting thereto, unless otherwise stated.

[0042] The specific embodiments of the present invention are described in detail below with reference to the technical solutions:

[0043] Step 1: To ensure the safety and equipment reliability of coal-fired power plants, an image database of abnormal situations in coal-fired power plants was constructed, targeting typical safety hazards and anomalies in key areas of the plant. This database covers anomalies such as unauthorized crossings of dangerous areas, falls by power plant personnel, and failure to wear safety helmets. The image samples cover different shooting angles and include various lighting conditions, both day and night. A total of 11,502 original images were collected. Some of these original images are shown below. Figure 1 As shown, the original images are divided into training set: test set: validation set in a ratio of 8:1:1.

[0044] Step 2: Construct an anomaly identification network for coal-fired power plants based on multi-scale features, with the following structure: Figure 2 As shown, the network mainly consists of a backbone network, a neck network, and a detection head, and specifically includes the following sub-steps:

[0045] (1) Input the abnormal situation image of the coal-fired power plant into the backbone network. The backbone network consists of four parts: four dual-branch feature extraction subnetworks DFES1, DFES2, DFES3 and DFES4, which are connected in sequence.

[0046] ① The DFES1 input is a 640×640×3 image of an anomaly in a coal-fired power plant, containing one RepVGG1_1, one RepVGG1_2, one dual-branch module Dufex1_1, and one convolutional module Conv1_1, connected serially. The RepVGG structures are identical throughout this invention. The RepVGG1_1 network adopts a training-inference separation design, and the relevant design structure is as follows: Figure 3 , Figure 4 As shown.

[0047] During training, the RepVGG network structure includes four sets of convolutional units: GConv1, GConv2, GConv3, and GConv4, connected sequentially. GConv1 contains a two-branch module GC_Dufex1 and a ReLU function, also connected sequentially. GC_Dufex1 contains two parallel branches, each containing a 3×3 convolutional kernel with a stride of 2, and the features from the left and right branches are concatenated. GConv2, GConv3, and GConv4 have the same structure, each containing a three-branch module and a ReLU function, connected sequentially. The three-branch module contains three parallel branches: the left branch contains a 3×3 convolutional kernel with a stride of 2; the middle branch contains a 1×1 convolutional kernel with a stride of 2; and the right branch contains the input of this module. The features from the left, middle, and right branches are concatenated. During inference, the RepVGG network structure includes four sets of convolutional units: GConv5, GConv6, GConv7, and GConv8, connected sequentially. The four convolutional units GConv5, GConv6, GConv7, and GConv8 have the same structure. Each convolutional unit includes a 3×3 convolutional kernel with a stride of 2 and a ReLU function, which are connected in series.

[0048] Dufex1_1 contains two parallel branches. The left branch contains one convolutional module, Conv1_2; the right branch contains one convolutional module, Conv1_3, and two feature extraction modules, RepVGG1_3 and RepVGG1_4, which are connected sequentially. The left and right branches perform feature concatenation. Specifically, Conv1_2 has a kernel size of 1×1, a stride of 1, valid padding, and 32 output channels; Conv1_3 also has a kernel size of 1×1, a stride of 1, valid padding, and 32 output channels.

[0049] The parameters of the convolution module Conv1_1 are a kernel size of 1×1, a stride of 1, valid padding, and 64 output channels. The DFES1 module outputs a single feature IM1_1 with dimensions of 160×160×64.

[0050] ② The input to DFES2 is IM1_1, which contains one Maxpool1_1 module, one dual-branch module Dufex1_2, and one convolutional module Conv1_4, connected in series. Dufex1_2 contains two parallel branches: the left branch contains one convolutional module Conv1_5; the right branch contains one convolutional module Conv1_6 and two feature extraction modules RepVGG1_5 and RepVGG1_6, connected in series; the left and right branches perform feature concatenation. Maxpool1_1 has a 3×3 maxpooling kernel, a stride of 2, same padding, and 64 output channels; Conv1_5 has a 3×3 kernel, a stride of 1, same padding, and 64 output channels; Conv1_6 has a 3×3 kernel, a stride of 1, same padding, and 64 output channels.

[0051] The parameters for the convolutional module Conv1_4 are a kernel size of 1×1, a stride of 1, valid padding, and 128 output channels. The DFES2 module outputs a single feature IM1_2 with dimensions of 80×80×128.

[0052] ③ The DFES3 input is IM1_2, which contains one Maxpool1_2, one dual-branch module Dufex1_3, and one convolutional module Conv1_7, connected in series. Dufex1_3 contains two parallel branches: the left branch contains one convolutional module Conv1_8; the right branch contains one convolutional module Conv1_9 and two feature extraction modules RepVGG1_7 and RepVGG1_8, connected in series; the left and right branches perform feature concatenation. Maxpool1_2 has a 3×3 maxpooling kernel, a stride of 2, same padding, and 128 output channels; Conv1_8 has a 3×3 kernel, a stride of 1, same padding, and 128 output channels; Conv1_9 has a 3×3 kernel, a stride of 1, same padding, and 128 output channels.

[0053] The parameters for the convolution module Conv1_7 are a kernel size of 1×1, a stride of 1, valid padding, and 256 output channels. The DFES3 module outputs a single feature IM1_3 with dimensions of 40×40×256.

[0054] ④ The DFES4 input is IM1_3, which contains one Maxpool1_3 module, one dual-branch module Dufex1_4, and one convolutional module Conv1_10, connected in series. Dufex1_4 contains two parallel branches: the left branch contains one convolutional module Conv1_11; the right branch contains one convolutional module Conv1_12 and two feature extraction modules RepVGG1_9 and RepVGG1_10, connected in series; the left and right branches perform feature concatenation. Maxpool1_3 has a 3×3 maxpooling kernel, a stride of 2, same padding, and 256 output channels; Conv1_11 has a 3×3 convolutional kernel, a stride of 1, same padding, and 256 output channels; Conv1_12 has a 3×3 convolutional kernel, a stride of 1, same padding, and 256 output channels.

[0055] The parameters for the convolution module Conv1_10 are a kernel size of 1×1, a stride of 1, valid padding, and 512 output channels. The DFES4 module outputs a single feature IM1_4 with dimensions of 20×20×512.

[0056] (2) The main channel feature IM1_4 output from the backbone network is input into the neck network. The neck network consists of 5 sets of convolutional processing units CP1, CP2, CP3, CP4 and CP5, which are connected in sequence.

[0057] ① CP1 consists of one main branch MB2_1 and one side branch LB2_1, with features from the main and side branches concatenated. MB2_1 takes IM1_4 as input and contains one SPP module, one convolutional module Conv2_1, and one upsample2_1 module, connected sequentially. LB2_1 takes IM1_3 as input and contains one convolutional module Conv2_2. The SPP module outputs a 20×20×256 feature IM2_1, which is the input to Conv2_1. Conv2_1 has a kernel size of 1×1, a stride of 1, valid padding, and 128 output channels. The upsample2_1 module outputs 128 channels. Conv2_2 has a kernel size of 1×1, a stride of 1, valid padding, and 128 output channels. The CP1 unit outputs a 40×40×256 feature IM2_2.

[0058] ② CP2 consists of one main branch MB2_2 and one side branch LB2_2, with features concatenated from the main and side branches. MB2_2's input is IM2_3, containing one RCS-OSA2_1 module, one convolutional module Conv2_3, and one upsample2_2 module, connected in sequence. LB2_2's input is IM1_2, containing one convolutional module Conv2_4. The RCS-OSA modules in this invention all have the same structure, as shown below. Figure 5 As shown. The RCS-OSA2_1 module contains one RepVGG2_1 network, one dual-branch module Dufex2_1, and one RepVGG2_2 network, which are connected in series. In Dufex2_1, the left branch is the output of RepVGG2_1; the right branch contains n / 2 RCS modules 2_1 and one dual-branch module Dufex2_2, which are connected in series. In Dufex2_2, the left branch is the output of n / 2 RCS modules 2_1; the right branch is n / 2 RCS modules 2_2, which are connected in series. The output features of RepVGG2_1, n / 2 RCS modules 2_1, and n / 2 RCS modules 2_2 are channel-washed and then concatenated; where n is a positive integer, which is the number of RCS modules in this stacked structure. Its value can be adjusted according to the feature extraction accuracy requirements in different application scenarios. In this invention, the value of n ranges from [2, 8].

[0059] The RCS-OSA2_1 module outputs a 40×40×128 feature IM2_3, which serves as the input to Conv2_3. Conv2_3 parameters include a 1×1 kernel size, a stride of 1, valid padding, and 64 output channels. The Upsample2_2 module outputs 64 channels; Conv2_4 parameters include a 1×1 kernel size, a stride of 1, valid padding, and 64 output channels. The CP2 unit outputs an 80×80×128 feature IM2_4.

[0060] ③ CP3 contains one main branch MB2_3 and one side branch LB2_3, with features from the main and side branches concatenated. MB2_3 takes IM2_6 as input and contains one RCS-OSA2_2 module and one convolutional module Conv2_5, connected sequentially. LB2_3 outputs IM2_3. The RCS-OSA2_2 module outputs a feature IM2_5 with dimensions 80×80×64, which serves as the input to Conv2_5. Conv2_5 parameters are a 3×3 kernel size, a stride of 2, same padding, and 128 output channels. The CP3 unit outputs a feature IM2_6 with dimensions 40×40×256.

[0061] ④ CP4 consists of one main branch MB2_4 and one side branch LB2_4, with features from the main and side branches concatenated. MB2_4 takes IM2_6 as input and contains one RCS-OSA2_3 module and one convolutional module Conv2_6, connected sequentially. LB2_4 takes IM2_1 as input. The RCS-OSA2_3 module outputs a 40×40×128 feature IM2_7, which is the input to Conv2_6. Conv2_6 has a kernel size of 3×3, a stride of 2, same padding, and 256 output channels. The CP4 unit outputs a 20×20×512 feature IM2_8.

[0062] ⑤ The input to CP5 is IM2_8, which contains one RCS-OSA2_4 module, one Maxpool2_1 module, and one E_ELAN module, connected serially. The RCS-OSA2_4 module outputs a feature IM2_9 with dimensions 20×20×256, which is the input to Maxpool2_1. The parameters of Maxpool2_1 are a 3×3 max-pooling kernel, a stride of 2, same padding, and 256 output channels. The E_ELAN module also outputs 256 channels. The CP5 unit outputs a feature IM2_10 with dimensions 10×10×256.

[0063] (3) Input the five features IM1_1 output from the backbone network and IM2_5, IM2_7, IM2_9, and IM2_10 output from the neck network into the detection head network. The detection head network contains five sets of TSCODE modules connected in parallel. In this invention, the TSCODE modules all have the same structure. The inputs of the five sets of TSCODE modules correspond to the five features mentioned above. Each set of TSCODE modules contains a dual-branch structure, with the upper branch containing a semantic context coding unit (SCE), as shown in the figure. Figure 6 As shown; the lower branch contains one detail-preserving coding unit (DPE), with the structure as follows: Figure 7 As shown; parallel output. The SCE consists of one main branch MB3_1 and one side branch LB3_1, with the main and side branch features concatenated. The input to MB3_1 is the feature map of the current layer. It contains one Downsample3_1 module; the output of LB3_1 is the feature map of the next layer. The DPE consists of one main branch MB3_2, one side branch LB3_2, and one side branch LB3_3. The features from the main and side branches are concatenated. The output of MB3_2 is the feature map of the current layer. LB3_2 input is the feature map of the next layer. It contains one Up sample 3_1; the input of LB3_3 is the feature map of the previous layer. It contains one splicing module and one Down sample3_2 module, connected serially. The splicing module includes one Up sample3_2 module, specifically... After processing with Up sample3_2 and Perform feature splicing.

[0064] Step 3: Train the deep network. Use the training set built in Step 1 to train the deep network built in Step 2, and obtain the network parameters and model.

[0065] Step 4: Perform abnormal situation image recognition of coal-fired power plants. Use the abnormal situation recognition network of coal-fired power plants obtained in Step 3 to recognize the newly collected abnormal situation images of coal-fired power plants and obtain the abnormal situation recognition results.

[0066] While ensuring recognition accuracy, the model's parameter count and computational overhead have been significantly reduced, improving its deployment adaptability and practical efficiency on edge platforms, and providing strong technical support for real-time intelligent perception of factory monitoring in complex environments.

[0067] To further verify the effectiveness of the proposed method, the detection model proposed in this invention was compared with mainstream object detection models on a self-built dataset. The results are shown in Table 1, where P is the accuracy, R is the recall, mAP50 is the average accuracy, Param is the number of model parameters, and FLOPs is the computational cost of the model.

[0068] Table 1. Relevant indicators under the comparative experiment of different models

[0069]

[0070] As shown in Table 1, the model of this invention outperforms mainstream models such as Faster R-CNN, YOLOv3-tiny, YOLOv5n, YOLOv5s, YOLOv6, YOLOv8n, and YOLOv10n in all three metrics: P, R, and mAP50. Compared to Faster... Despite a slightly higher parameter count, R-CNN significantly improves accuracy, with a 42.6% increase in P and a 16.5% increase in mAP50. Simultaneously, inference speed is reduced from 27.1ms to 6.9ms, dramatically enhancing real-time detection capabilities. Compared to YOLOv3-tiny, P, R, and mAP50 are improved by 9.7%, 6.9%, and 7.9%, respectively, with an acceptable increase in computational load. Furthermore, the inference time is reduced from 17.3ms to 6.9ms, demonstrating faster detection speed while maintaining high accuracy. Compared to the YOLOv5 series, although the model has slightly more parameters and FLOPs, P increases from approximately 73% to 82.5%, and mAP50 also shows a slight improvement, indicating higher accuracy and fewer false positives. The inference time is also significantly faster than YOLOv5n's 11.2ms and YOLOv5s's 16.5ms. The time to inference is further reduced from 8.5ms, indicating that the model has improved real-time performance while also improving accuracy. Compared with YOLOv6, the model of this invention still leads in all indicators under the lightweight design, especially with a P-value that is 5.4 percentage points higher, showing more stable performance. The inference time is 6.9ms, which is slightly faster than YOLOv6's 7.1ms, indicating that the model maintains efficient inference ability while ensuring high accuracy. Compared with more advanced lightweight models such as YOLOv8n and YOLOv10n, the model of this invention further improves the mAP50 to 81.6% and the P-value to 82.5%. Although the inference time is slightly higher than YOLOv8n's 4.9ms and YOLOv10n's 5.3ms, considering both accuracy and speed, it is still an efficient and high-precision solution, demonstrating stronger detection accuracy and robustness. In summary, although the model of this invention has slightly higher parameter quantity and computational cost, it is superior in key indicators such as P, R and mAP50, indicating that the model is more suitable for detecting abnormal conditions in coal-fired power plants with high requirements for detection accuracy and real-time performance.

[0071] To further verify the effectiveness of the model of this invention, we selected eight typical behaviors as detection objects: wearing a safety vest, not wearing a safety vest, wearing a safety helmet, not wearing a safety helmet, climbing dangerous areas, falling, walking, and standing normally, corresponding to categories 1 to 8 respectively. The models involved in Table 1 were then used to perform accuracy tests on these eight categories, and the results are as follows: Figure 8 As shown, from Figure 8 It can be seen that the model of this invention has achieved superior performance improvement in several difficult-to-detect categories.

[0072] Figure 9 This section describes the anomaly detection results obtained by using different models to detect abnormal situations in images of coal-fired power plants. From... Figure 9 It can be seen that the model of this invention exhibits stronger detection capabilities for various behavioral targets, especially in categories with complex occlusion or blurred features.

[0073] 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 abnormal conditions in coal-fired power plants based on multi-scale features, characterized in that: S1. Construct an image database of abnormal situations in coal-fired power plants as a dataset; S2. Construct an anomaly identification network for coal-fired power plants based on multi-scale features. This network includes a backbone network, a neck network, and a detection head, specifically including the following sub-steps: S2.1 Input the abnormal situation image of the coal-fired power plant into the backbone network. The backbone network includes four dual-branch feature extraction subnetworks DFES1, DFES2, DFES3 and DFES4, which are connected in sequence. S2.1.1 and DFES1 inputs are images of abnormal conditions in coal-fired power plants. They contain one feature extraction module RepVGG1_1, one feature extraction module RepVGG1_2, one dual-branch module Dufex1_1, and one convolutional module Conv1_1, all connected serially. The RepVGG network adopts a training-inference separation structure. During training, the RepVGG network structure includes four sets of convolutional units GConv1, as well as GConv2, GConv3, and GConv4 with identical structures, all connected serially. During inference, the RepVGG network structure includes four sets of convolutional units GConv5, GConv6, GConv7, and GConv8, all connected serially. Dufex1_1 contains two parallel branches. The left branch contains one convolutional module Conv1_2; the right branch contains one convolutional module Conv1_3 and two feature extraction modules RepVGG1_3 and RepVGG1_4, which are connected in sequence. The left and right branches concatenate features. The DFES1 module outputs one feature IM1_1. The input to S2.1.2 and DFES2 is IM1_1, which contains one Maxpool1_1, one dual-branch module Dufex1_2, and one convolutional module Conv1_4, connected in series. Dufex1_2 contains two parallel branches: the left branch contains one convolutional module Conv1_5; the right branch contains one convolutional module Conv1_6 and two feature extraction modules RepVGG1_5 and RepVGG1_6, connected in series. The left and right branches concatenate the features. The DFES2 module outputs one feature IM1_2. S2.1.3 and DFES3 take IM1_2 as input, which contains one Maxpool1_2, one dual-branch module Dufex1_3, and one convolutional module Conv1_7, connected in series. Dufex1_3 contains two parallel branches: the left branch contains one convolutional module Conv1_8, and the right branch contains one convolutional module Conv1_9 and two feature extraction modules RepVGG1_7 and RepVGG1_8, connected in series. Features are concatenated between the left and right branches. The DFES3 module outputs one feature IM1_3. The input to S2.1.4 and DFES4 is IM1_3, which contains one Maxpool1_3, one dual-branch module Dufex1_4, and one convolutional module Conv1_10, connected in series. Dufex1_4 contains two parallel branches: the left branch contains one convolutional module Conv1_11; the right branch contains one convolutional module Conv1_12 and two feature extraction modules RepVGG1_9 and RepVGG1_10, connected in series. The left and right branches concatenate the features. The DFES4 module outputs one feature IM1_4. S2.2 The main channel feature IM1_4 output from the backbone network is input into the neck network; the neck network consists of 5 sets of convolutional processing units CP1, CP2, CP3, CP4 and CP5, which are connected in sequence. S2.2.1, CP1 contains one main branch MB2_1 and one side branch LB2_1, with features concatenated from the main and side branches; the input of MB2_1 is IM1_4, and it contains one SPP module, one convolutional module Conv2_1, and one upsample2_1 module, which are connected in sequence; the input of LB2_1 is IM1_3, and it contains one convolutional module Conv2_2; the SPP module outputs one feature IM2_1, which is the input of Conv2_1; the CP1 unit outputs one feature IM2_2; S2.2.2, CP2 contains one main branch MB2_2 and one side branch LB2_2, with features concatenated from the main and side branches; the input of MB2_2 is IM2_2, which contains one RCS-OSA2_1 module, one convolutional module Conv2_3, and one upsample2_2 module, connected in sequence; the input of LB2_2 is IM1_2, which contains one convolutional module Conv2_4; the RCS-OSA2_1 module outputs one feature IM2_3, which is the input of Conv2_3; the CP2 unit outputs one feature IM2_4. S2.2.3 and CP3 contain one main branch MB2_3 and one side branch LB2_3, with features concatenated from the main and side branches. The input of MB2_3 is IM2_4, which contains one RCS-OSA2_2 module and one convolutional module Conv2_5, connected in sequence. The output of LB2_3 is IM2_3. Among them, the RCS-OSA2_2 module outputs one feature IM2_5, which is the input of Conv2_5. The CP3 unit outputs one feature IM2_6. S2.2.4 and CP4 each contain one main branch MB2_4 and one side branch LB2_4, with the main and side branch features concatenated. The input of MB2_4 is IM2_6, which contains one RCS-OSA2_3 module and one convolutional module Conv2_6, connected in sequence. The output of LB2_4 is IM2_1. Among them, the RCS-OSA2_3 module outputs one feature IM2_7, which is the input of Conv2_6. The CP4 unit outputs one feature IM2_8. The inputs of S2.2.5 and CP5 are IM2_8, which contains one RCS-OSA2_4 module, one Maxpool2_1 module, and one E_ELAN module, connected serially; among them, the RCS-OSA2_4 module outputs one feature IM2_9, which is the input of Maxpool2_1; the CP5 unit outputs one feature IM2_10; S2.

3. Input the five features (IM1_1 output from the backbone network and IM2_5, IM2_7, IM2_9, and IM2_10 output from the neck network) into the detection head network. The detection head network contains five sets of TSCODE modules connected in parallel. The TSCODE modules have the same structure, and the inputs of the five sets of TSCODE modules correspond to the five features mentioned above. Each set of TSCODE modules contains a dual-branch structure, with the upper branch containing a semantic context coding unit (SCE) and the lower branch containing a detail-preserving coding unit (DPE), and they are output in parallel. S3. Train a deep network. Use the dataset constructed in step S1 to train the coal-fired power plant abnormal situation identification network constructed in step S2, and obtain the network parameters and model. S4. Perform abnormal situation image recognition of coal-fired power plants. Use the abnormal situation recognition network of coal-fired power plants obtained in S3 to recognize the newly collected abnormal situation images of coal-fired power plants and obtain the abnormal situation recognition results.

2. The method for identifying abnormal conditions in coal-fired power plants based on multi-scale features according to claim 1, characterized in that: Images of abnormal situations at coal-fired power plants include instances of unauthorized crossings of dangerous areas, power plant personnel falling, and failure to wear safety helmets or safety vests; the image samples cover different shooting conditions, including various lighting conditions during the day and at night.

3. The method for identifying abnormal conditions in coal-fired power plants based on multi-scale features according to claim 1, characterized in that: The RepVGG network adopts a training-inference separation structure design: During the training phase, the RepVGG network structure includes four sets of convolutional units GConv1, GConv2, GConv3 and GConv4, which are connected in series; GConv1 contains one dual-branch module GC_Dufex1 and one ReLU function, which are connected in series; wherein, GC_Dufex1 contains two parallel branches, each of which contains one 3×3 convolutional kernel with a stride of 2, and the features of the left and right branches are concatenated; GConv2, GConv3, and GConv4 have the same structure, each containing a three-branch module and a ReLU function, which are connected in sequence. The three-branch module contains three parallel branches: the left branch contains a 3×3 convolutional kernel with a stride of 2; the middle branch contains a 1×1 convolutional kernel with a stride of 2; and the right branch contains the input of the module. The left, middle, and right branches are used to concatenate features. During the inference phase, the RepVGG network structure consists of four sets of identical convolutional units GConv5, GConv6, GConv7, and GConv8, connected in series. Each set of convolutional units includes a 3×3 convolutional kernel with a stride of 2 and a ReLU function, connected in series.

4. The method for identifying abnormal conditions in coal-fired power plants based on multi-scale features according to claim 1, characterized in that: In S2.1, The parameters for Conv1_2 are: kernel size 1×1, stride 1, valid padding, and 32 output channels; the parameters for Conv1_3 are: kernel size 1×1, stride 1, valid padding, and 32 output channels; the parameters for the convolution module Conv1_1 are: kernel size 1×1, stride 1, valid padding, and 64 output channels. The parameters for Maxpool1_1 are: 3×3 max pooling kernel, stride 2, same padding, and 64 output channels; the parameters for Conv1_5 are: 3×3 convolutional kernel, stride 1, same padding, and 64 output channels; the parameters for Conv1_6 are: 3×3 convolutional kernel, stride 1, same padding, and 64 output channels; the parameters for the convolutional module Conv1_4 are: 1×1 convolutional kernel, stride 1, valid padding, and 128 output channels. The parameters for Maxpool1_2 are: 3×3 max pooling kernel, stride 2, same padding, and 128 output channels; the parameters for Conv1_8 are: 3×3 convolutional kernel, stride 1, same padding, and 128 output channels; the parameters for Conv1_9 are: 3×3 convolutional kernel, stride 1, same padding, and 128 output channels; the parameters for the convolutional module Conv1_7 are: 1×1 convolutional kernel, stride 1, valid padding, and 256 output channels. The parameters for Maxpool1_3 are: 3×3 max pooling kernel, stride 2, same padding, and 256 output channels; the parameters for Conv1_11 are: 3×3 convolutional kernel, stride 1, same padding, and 256 output channels; the parameters for Conv1_12 are: 3×3 convolutional kernel, stride 1, same padding, and 256 output channels; the parameters for the convolutional module Conv1_10 are: 1×1 convolutional kernel, stride 1, valid padding, and 512 output channels.

5. The method for identifying abnormal conditions in coal-fired power plants based on multi-scale features according to claim 1, characterized in that: In S2.2, Conv2_1 parameters are: kernel size 1×1, stride 1, valid padding, and 128 output channels; Up sample2_1 module has 128 output channels; Conv2_2 parameters are: kernel size 1×1, stride 1, valid padding, and 128 output channels. Conv2_3 parameters are: kernel size 1×1, stride 1, valid padding, and 64 output channels; Up sample2_2 module outputs 64 channels; Conv2_4 parameters are: kernel size 1×1, stride 1, valid padding, and 64 output channels. Conv2_5 parameters are: kernel size 3×3, stride 2, same padding, and output channels 128. Conv2_6 parameters are: kernel size 3×3, stride 2, same padding, and output channels 256. The parameters for Maxpool2_1 are: 3×3 max pooling kernel, 2-step stride, same padding, and 256 output channels; the E_ELAN output channel count is 256.

6. The method for identifying abnormal conditions in coal-fired power plants based on multi-scale features according to claim 1, characterized in that: The RCS-OSA module structure consists of one RepVGG2_1 network, one dual-branch module Dufex2_1, and one RepVGG2_2 network, connected serially. In Dufex2_1, the left branch is the output of RepVGG2_1; the right branch contains n / 2 RCS modules 2_1 and one dual-branch module Dufex2_2, connected serially. In Dufex2_2, the left branch is the output of n / 2 RCS modules 2_1; the right branch is n / 2 RCS modules 2_2, connected serially. The output features of RepVGG2_1, n / 2 RCS modules 2_1, and n / 2 RCS modules 2_2 are channel-shuffled and then concatenated. Here, n is a positive integer representing the number of RCS modules in the stacked structure.

7. The method for identifying abnormal conditions in coal-fired power plants based on multi-scale features according to claim 1, characterized in that: In the TSCODE module structure SCE contains one main branch MB3_1 and one side branch LB3_1, with the main and side branches' features concatenated. The input to MB3_1 is the feature map of the current layer. It contains one Downsample3_1 module; the output of LB3_1 is the feature map of the next layer. ; DPE contains one main branch MB3_2, one side branch LB3_2 and one side branch LB3_3, with the main and side branches being spliced ​​together. The output of MB3_2 is the feature map of the current layer. LB3_2 input is the feature map of the next layer. It contains one Upsample3_1; the input of LB3_3 is the feature map of the previous layer. It includes one splicing module and one Down sample3_2 module, connected serially; the splicing module includes one Up sample3_2 module, specifically... After processing with Up sample3_2 and Perform feature splicing.

Images