Construction method of damper throwing and withdrawing identification network, server and storage medium

Abstract

This application discloses a method for constructing a windshield operation / retreat recognition network, a server, and a storage medium, belonging to the field of automatic control. The method includes: Step 1: Constructing a windshield operation / retreat recognition network, including a master network and slave networks, which are downsampled using overlay convolution; Step 2: Collecting positive and negative samples of several sets of windshield images to form an initial dataset, which is then proportionally divided into an initial training set and an initial validation set; Step 3: Training the windshield operation / retreat recognition network, retaining the optimal weights from the initial validation set; Step 4: Loading the optimal weights into the windshield operation / retreat recognition network for recognition, comparing the recognition results with the true values, and adding misclassified images to the optimization dataset; Step 5: Adjusting the optimal weights using the optimization dataset until a mature standard is reached. This method utilizes a deep convolutional neural network to determine the operation / retreat status of the windshield, exhibiting stronger stability and higher accuracy.

Description

Technical Field

[0001] This application belongs to the field of automatic control, and specifically relates to a method for constructing a windshield deactivation / activation identification network, a server, and a storage medium. Background Technology

[0002] The air brake of the hydro-generator brake is an important component of the mechanical braking system of the hydro-generator unit. When the air brake is engaged, the bottom cylinder pushes the brake plate upward, causing the brake shoes on the brake plate to rub against the brake ring below the generator rotor, thereby reducing the unit speed. When the air brake is disengaged, the bottom cylinder descends to reset the brake plate position, disengaging the brake shoes from the brake ring and maintaining normal unit operation.

[0003] To determine the exact location of the airlock, a contact-type limit switch is typically installed below it. The hydropower plant's computer monitoring system indirectly determines the airlock's activation or deactivation status by collecting data on the operation of these limit switches. However, due to the harsh environment and high humidity inside the turbine's wind tunnel, the contact-type limit switches have a high failure rate, causing the computer monitoring system to be unable to accurately collect the airlock's activation or deactivation signals, resulting in significant inconvenience for maintenance personnel.

[0004] With the development of computer vision and deep learning technologies, deep convolutional networks have replaced contact limit switches for recognizing various mechanical positions, such as the ConvNext network. However, directly applying existing ConvNext networks to detect the operation and shutdown of wind gates presents certain problems: First, data annotation is difficult because the visibility conditions inside the wind gate are poor, making it hard to identify the shutdown status in the captured images; second, the wind gate itself occupies a large area in the captured images, and the receptive field of traditional ConvNext networks is not large enough, preventing the model from learning global image information. Therefore, traditional ConvNext networks cannot effectively identify the shutdown status of turbine unit wind gates.

[0005] Therefore, a technical solution for constructing an identification network for the deactivation of wind gates is needed to solve the above problems. Summary of the Invention

[0006] To address the shortcomings of the existing technology, this application provides an improved method for constructing a cascaded ConvNeXt damper activation / deactivation identification network. This method employs a cascaded and short-circuit structure to increase the model's receptive field, and improves the quality of channel feature integration through the SENet channel attention module. The initialization layer and downsampling module are modified to cover convolutions to enhance the model's recognition capability. A deep convolutional neural network is used to determine the damper's activation / deactivation status, replacing the traditional contact-type limit switch, resulting in stronger stability and higher accuracy.

[0007] The technical effect to be achieved in this application is accomplished through the following solution:

[0008] According to a first aspect of this application, a method for constructing a windshield deactivation identification network is provided, comprising the following steps:

[0009] Step 1: Construct a windshield deployment / removal identification network. The windshield deployment / removal identification network includes a master network and a slave network. The slave network is used to expand the network's receptive field and provide multi-scale features. The master network and the slave network use overlay convolution for downsampling.

[0010] Step 2: Construct the initial dataset, which involves collecting positive and negative samples of several sets of windshield images, performing data augmentation on the positive and negative samples to form the initial dataset, and dividing the initial dataset into an initial training set and an initial validation set according to the proportions.

[0011] Step 3: Train the wind gate deployment / retraction recognition network using the initialization training set and the initialization validation set. During the training process, retain the best weights from the initialization validation set.

[0012] Step 4: Load the optimal weights into the wind gate activation / deactivation recognition network to identify the activation / deactivation status of the wind gate, compare the recognition results with the true values, and add misclassified images to the optimization dataset;

[0013] Step 5: After adjusting the optimal weights using the optimized dataset, repeat Step 4 until the wind gate deactivation identification network reaches a mature standard, thus completing the construction of the wind gate deactivation identification network.

[0014] Preferably, in step 1, the input of the main network is a 224×224×3 RGB image, and the RGB image is scaled by a Gaussian pyramid to generate a 112×112×3 RGB image which is then input to the slave network to increase the receptive field of the network.

[0015] Preferably, both the main network and the slave network include an initialization convolution module, several ConvNeXt modules, and several downsampling modules. The feature map after passing through the ConvNeXt module of the slave network is recalibrated by the channel attention module, and then concatenated with the corresponding features of the same scale in the main network, so that the network obtains multi-scale features.

[0016] Preferably, the output of the main network downsampling module is The output from the network downsampling module is

[0017]

[0018] Where: dim is the number of feature channels output by the ConvNeXt module; h is the height of the feature map input to the downsampling module; and w is the width of the feature map input to the downsampling module.

[0019] Preferably, in step 2, at least 50 times the number of surveillance cameras are collected as positive and negative samples. These samples are then sized and normalized before data augmentation. Specifically, the windshield image is first flipped horizontally and vertically, and then a mask is added using the GridMask algorithm. The GridMask algorithm is set with the following parameters: the masking mask is square, parameter d is set to 96, and the spacing r between the masking masks is set to 0.4 and 0.6 respectively, with k = 2r - r 2 The retention ratios k were calculated to be 0.64 and 0.84, respectively. The images were then expanded by 12 times to the original positive and negative samples as the initial dataset. The initial dataset was then randomly divided into an initial training set and an initial validation set in an 8:2 ratio.

[0020] Preferably, in step 3, AdamW is used as the training optimizer, with an initial learning rate of 0.01, a weight decay of 0.05, a training epoch count of 300, a batch size of 8, and a loss function of binary cross-entropy loss, the formula of which is:

[0021] In the formula: N is the number of images in each batch, y i For the label, where y i =1 is the tag for the air damper being put into operation, y i =0 indicates the exit label for the air damper, p(y i ) is y i The predicted probabilities of the labels are trained using the formula described above to obtain the optimal weights.

[0022] Preferably, in step 4, the true value is the damper status information collected by the PLC; during training, damper image information from a monitoring camera and damper status information collected by the PLC are captured every 5 seconds, the damper image information is input into the damper withdrawal recognition network for recognition, and the recognition result is compared with the true value. The damper image that is incorrectly compared is taken as a misclassified image; when the number of misclassified images in the optimized dataset reaches 5 * the number of monitoring cameras, the optimized dataset is established.

[0023] Preferably, in step 5, adjusting the optimal weights using the optimized dataset specifically involves:

[0024] The optimization dataset was randomly divided into an optimization training set and an optimization validation set in an 8:2 ratio.

[0025] The images in the optimized training set and optimized validation set are subjected to size processing and normalization.

[0026] The network was trained using an optimized training set and an optimized validation set. The hyperparameters were set as follows: the training optimizer was AdamW, the initial learning rate was 0.0005, the weight decay was 0.05, the number of training epochs was 10, the batch size was 8, and the loss function was FocalLoss. During training, the weights that performed best in the optimized validation set were retained as the optimal weights, and the optimal weights were adjusted.

[0027] According to a second aspect of this application, a server is provided, comprising: a memory and at least one processor;

[0028] The memory stores a computer program, and the at least one processor executes the computer program stored in the memory to implement the above-described method for constructing the windshield deactivation and activation identification network.

[0029] According to a third aspect of this application, a computer-readable storage medium is provided, wherein a computer program is stored in the computer-readable storage medium, and when the computer program is executed, it implements the above-described method for constructing the windshield deactivation and deactivation identification network.

[0030] According to one embodiment of this application, the technical advantage of using the construction method of this windshield deactivation and activation identification network is as follows:

[0031] The system utilizes a deep convolutional neural network to determine the activation / deactivation status of the damper, replacing the traditional contact limit switch. This provides greater stability and facilitates maintenance by operators.

[0032] An improved cascaded ConvNeXt wind gate activation / deactivation identification network was constructed, which has higher accuracy in wind gate activation / deactivation identification tasks compared with the original ConvNeXt network;

[0033] Data labeling is completed by linking the video monitoring systems of the hydropower plant's computer monitoring system and mechanical braking system, freeing up manpower and enabling automatic model updates. Attached Figure Description

[0034] To more clearly illustrate the embodiments of this application or the existing technical solutions, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0035] Figure 1 This is a flowchart illustrating a method for constructing a wind gate deactivation / reactivation identification network according to an embodiment of this application;

[0036] Figure 2 This is a structural block diagram of the damper deactivation / activation identification network in one embodiment of this application;

[0037] Figure 3 This is a structural block diagram of the ConvNeXt module in one embodiment of this application;

[0038] Figure 4 This is a structural block diagram of the downsampling module in one embodiment of this application;

[0039] Figure 5 This is a schematic diagram of the structure of a server according to one embodiment of this application. Detailed Implementation

[0040] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0041] The method for constructing the wind gate deactivation and activation identification network in one embodiment of this application requires the deployment of monitoring hardware before implementation. For example, it involves constructing a video monitoring system for the mechanical braking system of a hydropower plant turbine unit, including the front-end installation and back-end deployment of monitoring cameras.

[0042] The main method of front-end installation is as follows: a base for fixing the camera is welded on the support arm of the lower frame of the wind tunnel under the turbine. The detachable macro PoE camera, which meets the State Grid Level 4 radiation standard, is installed on the camera base. A switch with PoE power supply function is deployed in the electromagnetic shielding box next to the support arm of the lower frame, and all macro cameras are connected to the switch with an eight-core PoE network cable.

[0043] The main deployment method for the backend is as follows: server racks are deployed around the wind tunnel of the turbine unit, and video recorders are deployed inside the server racks, connected to the PoE switches using twisted-pair cables; servers with a certain AI computing power are deployed in the server racks, and connected to the video recorders using twisted-pair cables; a monitoring video acquisition program is deployed on the server to acquire real-time monitoring video streams using the RTSP protocol; the control unit PLC of the turbine unit is connected to the server via fiber optic cable; and a host computer communication program is deployed on the server to receive the wind gate status monitoring switch quantities from the PLC using the Modbus TCP protocol, providing real-value data for subsequent training.

[0044] After completing the above arrangements, the method for constructing the wind gate deactivation and activation identification network is as follows: Figure 1 As shown, it includes the following steps:

[0045] S110: Construct a windshield activation / deactivation identification network, which includes a master network and a slave network. The slave network is used to expand the network's receptive field and provide multi-scale features. The master network and the slave network use overlay convolution for downsampling.

[0046] The damper deactivation / activation identification network in this step is as follows: Figure 2 As shown, both the master network and the slave network include, in sequence:

[0047] The initialization convolution module is used for initializing convolutions and layer normalization. In the diagram, k is the kernel size parameter; for example, a convolution kernel with k=3 is a 3×3 convolution kernel. s is the convolution stride; p is the zero-padding coefficient for the feature map edges. For example, p=2 indicates that the feature map edges are zero-padding coefficients. Figure 4 The week is filled with 2 rows or 2 columns of 0. p = (2,1,2,1) means that the feature map is filled with 2 rows of 0 at the top, 1 row at the bottom, 2 columns on the left and 1 column on the right.

[0048] The ConvNeXt module, where dim is the number of feature channels output by the ConvNeXt module;

[0049] C represents feature concatenation, which means that the feature maps from the network are recalibrated by the SENet channel attention module after passing through the ConvNeXt module, and then concatenated with the corresponding features of the same scale in the main network, such as... Figure 2 The 56×56×32, 28×28×64 and 14×14×128 feature maps output from the network are stitched into the main network;

[0050] The downsampling module, the main network downsampling module output is The output from the network downsampling module is h represents the height of the feature map of the input downsampling module; w represents the width of the feature map of the input downsampling module.

[0051] SENet is a channel attention module, and r is its compression coefficient;

[0052] During configuration, the main network input is a 224×224×3 RGB image that reaches the initialization convolution module. After passing through multiple ConvNeXt modules, it is concatenated with the features of the secondary network and then reaches the downsampling module. After passing through several more ConvNeXt modules and concatenating with the features, global average pooling and layer normalization are performed. After linear regression calculation, the output is completed.

[0053] The main network inputs a 224×224×3 RGB image, which is then scaled using a Gaussian pyramid to generate a 112×112×3 RGB image, which is then input to the slave network. This increases the network's receptive field and improves network performance.

[0054] Compared to the original ConvNeXt network, the initial convolutions of the main and slave networks have been optimized by using overlay convolutions for downsampling, reducing image information loss during the downsampling process. After passing through the ConvNeXt module, the slave network feature maps undergo feature recalibration via the SENet channel attention module, and are then concatenated with corresponding features of the same scale from the main network, enabling the network to obtain multi-scale features. The downsampling module of this network, similar to the initial convolutions, uses overlay convolutions to reduce image information loss.

[0055] like Figure 3 As shown, in this embodiment, the ConvNeXt module sequentially performs depthwise convolution, layer normalization, convolution processing, GELU activation function, convolution processing, scaling, DropPath regularization, and finally feature accumulation (+ sign) of the short-circuit structure to complete the data processing; the parameters of depthwise convolution are k=7, s=4, p=3; the parameters of convolution processing are k=1, s=1.

[0056] like Figure 4 As shown, in this embodiment, the downsampling module performs layer normalization and convolution processing sequentially, and the convolution output in the main network is... The output from the network's convolution is The convolution parameters are all k=7, s=2, p=(3,2,3,2).

[0057] S120: Construct the initial dataset, specifically: collect positive and negative samples of several sets of windshield images, perform data augmentation on the positive and negative samples to form the initial dataset, and divide the initial dataset into an initial training set and an initial validation set according to the proportion;

[0058] In this step, the host computer communication program collects the changes in the switch quantity of the air damper status monitoring, delays for 1 minute, captures the current monitoring video image, and marks it with the corresponding status of the switch quantity. The image marked with "1" represents "air damper in", and the image marked with "0" represents "air damper out". Positive and negative samples are collected at a rate of 50 times the number of monitoring cameras.

[0059] The positive and negative samples are cropped and scaled to adjust the image scale to 224×224, and then the image is normalized. Data augmentation is performed on the normalized positive and negative samples: First, the image is flipped horizontally and vertically. The purpose of flipping is to: 1. Expand the dataset, as flipping the same image in two directions can generate 4 images in different directions (including the original image); 2. To a certain extent, it can overcome the overfitting of image spatial information and enhance the generalization of the trained model.

[0060] Then, the GridMask algorithm is used to add a mask. Specifically, masking masks are evenly distributed throughout the image. Based on the previous quadrupling and expansion of the data, an additional 8 times the image data is added, making the total dataset 12 times larger. The GridMask algorithm parameters are set as follows: the masking mask is a square, the parameter d is set to 96 (i.e., its side lengths δx and δy are random integers >0 and <96), and the spacing r between the masking masks is set to 0.4 and 0.6 respectively. This is achieved by using k = 2r - r 2 The retention ratios k were calculated to be 0.64 and 0.84, respectively. The images were then expanded by 12 times to the original positive and negative samples as the initial dataset. The initial dataset was then randomly divided into an initial training set and an initial validation set in an 8:2 ratio.

[0061] S130: Train the wind gate deployment / retraction identification network using the initialization training set and the initialization validation set, and retain the best weights in the initialization validation set during the training process;

[0062] In this step, AdamW is used as the training optimizer, with an initial learning rate of 0.01, weight decay of 0.05, 300 training epochs, a batch size of 8, and a binary cross-entropy loss function, the formula of which is:

[0063] In the formula: N is the number of images in each batch, y i For the label, where y i =1 is the tag for the air damper being put into operation, y i =0 indicates the exit label for the air damper, p(y i ) is y i The predicted probabilities of the labels are trained using the formula described above to obtain the optimal weights.

[0064] S140: Load the optimal weights into the wind gate operation / deactivation identification network to identify the wind gate operation / deactivation status, compare the identification results with the true values, and add misclassified images to the optimization dataset;

[0065] In this step, the optimal weights are loaded into the damper deployment / retraction recognition network as a trained network to detect the damper status in real time and construct an optimized dataset. The true values ​​are the damper status information collected by the PLC.

[0066] Every 5 seconds, capture the windshield image information from a monitoring camera and the windshield status information collected by the PLC. Input the windshield image information into the windshield withdrawal and activation recognition network for recognition, and compare the recognition result with the actual value.

[0067] The incorrectly classified windshield images are used as misclassified images, and the misclassified images and their corresponding true labels are added to the optimization dataset. When the number of misclassified images in the optimization dataset reaches 5 times the number of surveillance cameras, the optimization dataset is considered complete.

[0068] S150: After adjusting the optimal weights using the optimized dataset, repeat step 4 until the wind gate withdrawal identification network reaches a mature standard, thus completing the construction of the wind gate withdrawal identification network.

[0069] In this step, optimizing the data and fine-tuning the optimal weights specifically involves randomly dividing the optimization dataset into an optimization training set and an optimization validation set in an 8:2 ratio.

[0070] The training and validation images are subjected to the same data augmentation processing as in S120, namely cropping, scaling, and normalization preprocessing.

[0071] The wind gate withdrawal recognition network was further trained using an optimized training set and an optimized validation set. The hyperparameters were set as follows: the training optimizer was AdamW, the initial learning rate was 0.0005, the weight decay was 0.05, the number of training epochs was 10, and the batch size was 8. Since the positive and negative samples in the optimized dataset were imbalanced, the loss function was set to FocalLoss. During training, the best performing weights in the optimized validation set were retained to replace the original best weights.

[0072] The adjusted optimal weights are reloaded into the wind gate deployment and withdrawal recognition network. Steps S140 and S150 are repeated until the growth rate of the optimized training set images is less than 0.083 frames per hour. At this point, the wind gate deployment and withdrawal recognition network training is considered to have reached a mature standard.

[0073] The well-trained damper activation / deactivation identification network was put into production and operation. Every 5 seconds, the activation / deactivation results of the dampers in all monitoring images were predicted, completing one round of inspection. The activation / deactivation results were converted into I / O signals and transmitted to the PLC through the host computer communication program.

[0074] By linking the hydropower plant's computer monitoring system and mechanical braking system, data annotation is completed, freeing up manpower and enabling automatic updates and upgrades to the wind gate deactivation / deactivation identification network. Employing cascaded and short-circuit structures increases the model's receptive field, and the SENet channel attention module improves the quality of channel feature integration. Modifying the initialization layer and downsampling module to cover convolution enhances the model's recognition capabilities. In practical applications, the recognition accuracy for turbine wind gate pushers reaches 99.62%.

[0075] like Figure 5As shown, one embodiment of this application provides a server, including: a memory 501 and at least one processor 502; the processor 502 is equipped with a GPU to accelerate computing; and a serial communication module to communicate with a PLC.

[0076] The memory 501 stores a computer program, and the at least one processor 502 executes the computer program stored in the memory 501 to implement the above-mentioned method for constructing the windshield deactivation and activation identification network.

[0077] According to a third aspect of this application, a computer-readable storage medium is provided, wherein a computer program is stored in the computer-readable storage medium, and when the computer program is executed, it implements the above-described method for constructing the windshield deactivation and deactivation identification network.

[0078] According to one embodiment of this application, the technical advantage of using the construction method of this windshield deactivation and activation identification network is as follows:

[0079] The system utilizes a deep convolutional neural network to determine the activation / deactivation status of the damper, replacing the traditional contact limit switch. This provides greater stability and facilitates maintenance by operators.

[0080] An improved cascaded ConvNeXt wind gate activation / deactivation identification network was constructed, which has higher accuracy in wind gate activation / deactivation identification tasks compared with the original ConvNeXt network;

[0081] Data labeling is completed by linking the video monitoring systems of the hydropower plant's computer monitoring system and mechanical braking system, freeing up manpower and enabling automatic model updates.

[0082] It should be noted that the above detailed descriptions are exemplary and intended to provide further explanation of this application. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains.

[0083] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the exemplary embodiments according to this application. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0084] It should be noted that the terms "first," "second," etc., used in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such terms can be used interchangeably where appropriate so that the embodiments of this application described herein can be implemented in sequences other than those illustrated or described herein.

[0085] Furthermore, the terms “comprising” and “having”, and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or apparatus that includes a series of steps or units is not necessarily limited to those steps or units that are explicitly listed, but may include other steps or units that are not explicitly listed or that are inherent to such process, method, product, or apparatus.

[0086] For ease of description, spatial relative terms such as "above," "on top of," "on the upper surface of," "above," etc., are used herein to describe the spatial positional relationship of a device or feature as shown in the figures to other devices or features. It should be understood that spatial relative terms are intended to encompass different orientations in use or operation beyond the orientation of the device as described in the figures. For example, if the device in the figures were inverted, a device described as "above" or "on top of" other devices or structures would subsequently be positioned as "below" or "under" other devices or structures. Thus, the exemplary term "above" can include both "above" and "below." The device may also be positioned in other different ways, such as rotated 90 degrees or in other orientations, and the spatial relative descriptions used herein will be interpreted accordingly.

[0087] In the detailed description above, reference has been made to the accompanying drawings, which form part of this document. In the drawings, similar symbols typically identify similar parts unless the context otherwise indicates otherwise. The illustrated embodiments described in the detailed specification, drawings, and claims are not intended to be limiting. Other embodiments may be used and other changes may be made without departing from the spirit or scope of the subject matter presented herein.

[0088] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for constructing a wind gate deactivation / activation identification network, characterized in that, Includes the following steps: Step 1: Construct a windshield deployment / removal identification network. The windshield deployment / removal identification network includes a master network and a slave network. The slave network is used to expand the network's receptive field and provide multi-scale features. The master network and the slave network use overlay convolution for downsampling. The main network inputs an RGB image, and the RGB image is scaled down by a Gaussian pyramid to generate a smaller RGB image, which is then input to the slave network. Both the main network and the slave network include an initialization convolution module, several ConvNeXt modules, and several downsampling modules. The feature map after passing through the ConvNeXt module of the slave network is recalibrated by the channel attention module and then concatenated with the corresponding features of the same scale in the main network. The output of the main network downsampling module is The output from the network downsampling module is ; in: dim The number of feature channels output by the ConvNeXt module; h The height of the feature map of the input downsampling module; w The width of the feature map for the input downsampling module; Step 2: Construct the initial dataset, which involves collecting positive and negative samples of several sets of windshield images, performing data augmentation on the positive and negative samples to form the initial dataset, and dividing the initial dataset into an initial training set and an initial validation set according to the proportions. Step 3: Train the wind gate deployment / retraction recognition network using the initialization training set and the initialization validation set. During the training process, retain the best weights from the initialization validation set. Step 4: Load the optimal weights into the wind gate activation / deactivation recognition network to identify the activation / deactivation status of the wind gate, compare the recognition results with the true values, and add misclassified images to the optimization dataset; Step 5: After adjusting the optimal weights using the optimized dataset, repeat Step 4 until the wind gate deactivation identification network reaches a mature standard, thus completing the construction of the wind gate deactivation identification network.

2. The method for constructing the wind gate deactivation / activation identification network according to claim 1, characterized in that, In step 2, at least 50 times the number of surveillance cameras are collected as positive and negative samples. These samples are then resized and normalized before data augmentation. Specifically, the windshield image is first flipped horizontally and vertically, and then a mask is added using the GridMask algorithm. The GridMask algorithm is set with the following parameters: the masking mask is a square, parameter d is set to 96, and the spacing r between the masking masks is set to 0.4 and 0.6 respectively. The masking distance is calculated using k = 2r - r. 2 The retention ratios k were calculated to be 0.64 and 0.84, respectively. The images were then expanded by 12 times to the original positive and negative samples as the initial dataset. The initial dataset was then randomly divided into an initial training set and an initial validation set in an 8:2 ratio.

3. The method for constructing the wind gate deactivation / activation identification network according to claim 1, characterized in that, In step 3, AdamW is used as the training optimizer with an initial learning rate of 0.01, weight decay of 0.05, 300 training epochs, a batch size of 8, and a binary cross-entropy loss function, the formula of which is: ； In the formula: N y represents the number of images in each batch. i For the label, where y i =1 is the tag for the air damper being put into operation, y i =0 indicates the exit label for the air damper, p(y i ) is y i The predicted probabilities of the labels are trained using the formula described above to obtain the optimal weights.

4. The method for constructing the wind gate deactivation / activation identification network according to claim 1, characterized in that, In step 4, the true value is the damper status information collected by the PLC. During training, damper image information from a monitoring camera and damper status information collected by the PLC are captured every 5 seconds. The damper image information is input into the damper withdrawal recognition network for recognition, and the recognition result is compared with the true value. The damper image that is incorrectly compared is taken as a misclassified image. When the number of misclassified images in the optimized dataset reaches 5 times the number of monitoring cameras, the optimized dataset is established.

5. The method for constructing the wind gate deactivation / activation identification network according to claim 1, characterized in that, In step 5, adjusting the optimal weights using the optimized dataset specifically involves: randomly dividing the optimized dataset into an optimized training set and an optimized validation set at an 8:2 ratio; and performing size processing and normalization processing on the images in the optimized training set and the optimized validation set. The network was trained using an optimized training set and an optimized validation set. The hyperparameters were set as follows: the training optimizer was AdamW, the initial learning rate was 0.0005, the weight decay was 0.05, the number of training epochs was 10, the batch size was 8, and the loss function was FocalLoss. During training, the weights that performed best in the optimized validation set were retained as the optimal weights, and the optimal weights were adjusted.

6. A server, characterized in that, include: Memory and at least one processor; The memory stores a computer program, and the at least one processor executes the computer program stored in the memory to implement the method for constructing the wind gate deactivation and deactivation identification network according to any one of claims 1 to 5.

7. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, which, when executed, implements the method for constructing the wind gate deactivation and deactivation identification network according to any one of claims 1 to 5.

Images