An ai unmanned inspection system and method for water supply and drainage pump station

Abstract

The application provides a kind of for water supply and drainage pump station AI unmanned inspection system and method, it solves the problem of internal monitoring of water supply and drainage pump station, it includes the camera being arranged in pump house and the edge box being connected with camera, camera is connected with push stream service module, camera is connected with video center, edge box is connected with video center by edge management platform, video center is connected with internet of things platform by center end identification algorithm module and internet of things platform is connected with edge box, internet of things platform is connected with information model, information model is connected with video center, push stream service module and information model are connected with inspection module.The application has good compatibility, high recognition accuracy and other advantages.

Description

Technical Field

[0001] This invention belongs to the field of water station monitoring technology, specifically relating to an AI-powered unmanned inspection system and method for water supply and drainage pumping stations. Background Technology

[0002] Urban water supply and drainage systems are vital to a city's economy and have risen to a strategic level of urban safety, incorporated into the overall urban security concept. Water supply and drainage pumping stations, as the hub and heart of these systems, are crucial for ensuring the safe operation of urban water systems. With urban development, water supply and drainage networks are characterized by numerous points, long lines, wide coverage, and long-term uninterrupted operation. The complex factors of the natural environment, human damage, and social activities have resulted in hundreds of defects and hidden dangers in pumping stations, affecting the safe operation of the water supply and drainage network. Safety inspections of water supply and drainage pumping stations are crucial, but traditional methods such as manual line patrols and monitoring are inefficient, with high rates of missed and false inspections, failing to meet the current operation and maintenance needs of water supply and drainage pumping stations in my country.

[0003] To address the shortcomings of existing technologies, people have conducted long-term explorations and proposed various solutions. For example, Chinese patent literature discloses a water pump station monitoring system [201810630784.4], which includes a display screen, an intercom host, and intercom extensions. The display screen shows a primary wiring diagram and monitoring videos fed back by cameras. Cameras are respectively installed at the water pump station gate, water pump room, each water pump, power distribution room, transformer, drain outlet, and water inlet. The intercom host is located next to the display screen, and intercom extensions are located next to each camera. The primary wiring diagram includes parameter information, which includes a preset label format, a preset parameter attribute name format, and the corresponding values ​​of the parameter attribute names. The parameter attribute names include at least one of the following: water pump upper shaft temperature, water pump lower shaft temperature, water pump temperature, transformer winding temperature, transformer temperature, power factor of reactive power compensator, frequency of reactive power compensator, water level at the inlet, and water level at the outlet.

[0004] The above solution has solved the problem of pump station monitoring to a certain extent, but it still has many shortcomings, such as relying on too many types of sensing elements and poor compatibility with water supply and drainage pump stations. Summary of the Invention

[0005] The purpose of this invention is to address the above-mentioned problems by providing an AI-powered unmanned inspection system for water supply and drainage pumping stations that offers excellent monitoring performance and strong compatibility.

[0006] Another objective of this invention is to address the aforementioned problems by providing an AI-based unmanned inspection method for water supply and drainage pumping stations with high recognition accuracy.

[0007] To achieve the above objectives, the present invention adopts the following technical solution: an AI unmanned inspection system for water supply and drainage pumping stations, comprising a camera installed in the pumping station and an edge box connected to the camera, the camera being connected to a streaming service module and a video center, the edge box being connected to the video center through an edge management platform, the video center being connected to an IoT platform through a central end recognition algorithm module, and the IoT platform being connected to the edge box, the IoT platform being connected to an information model, the information model being connected to the video center, and the streaming service module and the information model being connected to the inspection module.

[0008] In the aforementioned AI-powered unmanned inspection system for water supply and drainage pumping stations, the inspection module includes an AI monitoring unit and an AI inspection unit.

[0009] In the aforementioned AI-powered unmanned inspection system for water supply and drainage pumping stations, the system includes a business application layer, which is connected to an information model layer and an AI recognition layer. The information model layer collects data from IoT data sources via the Internet of Things (IoT), and the AI ​​recognition layer collects data from video data sources via a security platform.

[0010] In the aforementioned AI-powered unmanned inspection system for water supply and drainage pumping stations, the central identification algorithm module includes:

[0011] Input layer: performs image input stitching, adaptive anchor box training, and adaptive image scaling;

[0012] Backbone network: Extracts multi-level features from the input image;

[0013] Neck network: Multi-scale features extracted from the backbone network are fused;

[0014] Detection head: Convolves multi-scale feature maps with 1×1 convolution to generate a prediction tensor containing anchor box coordinates, confidence and category information;

[0015] Output layer: performs prediction box decoding and non-maximum suppression (NMS).

[0016] An AI-powered unmanned inspection method for water supply and drainage pumping stations, employing the aforementioned AI-powered unmanned inspection system for water supply and drainage pumping stations, includes the following training steps:

[0017] S1: Forward propagation, the input image passes through the backbone network, the neck network, and the head detection network, and the prediction tensors at three scales are obtained at the network output.

[0018] S2: Match the target with the prediction to determine the real object that the prediction box is responsible for detecting;

[0019] S3: Perform image detection and introduce an attention mechanism;

[0020] S4: Perform loss calculations, including bounding box regression loss, confidence loss, and classification loss;

[0021] S5: Backpropagation and parameter update.

[0022] In the above-mentioned AI-powered unmanned inspection method for water supply and drainage pumping stations, step S2 involves prior anchor box matching, which matches the ground truth box with multiple preset anchor boxes at each location, and selects the best-matching anchor box as a positive sample; a ground truth box is assigned to a prediction box to be responsible for prediction, and the selected prediction box is a positive sample; prediction boxes that do not match a ground truth box are used as negative samples and participate in the calculation of confidence loss.

[0023] In the above-mentioned AI-based unmanned inspection method for water supply and drainage pumping stations, step S2 calculates the distance between the actual bounding box and the predicted bounding box. If the distance is less than a preset threshold, the box is selected. The distance calculation is as follows:

[0024] ;

[0025] in, The degree of matching between the ground truth bounding box and the predicted bounding box. This represents the position of the actual bounding box. The location of the predicted bounding box. Let be the covariance matrix of the predicted observation space.

[0026] In the above-mentioned AI-based unmanned inspection method for water supply and drainage pumping stations, step S3 includes the following steps:

[0027] S31: Obtain feature maps of different resolutions through Backbone, upsample the low-resolution feature map and fuse it with the high-resolution feature map;

[0028] S32: Perform average pooling operation layer by layer on the high-resolution feature maps at each spatial location and fuse them with the low-resolution feature maps.

[0029] S33: Interaction of channel components is calculated using one-dimensional convolution;

[0030] S34: Use the Sigmoid function for nonlinear mapping.

[0031] In the above-mentioned AI-based unmanned inspection method for water supply and drainage pumping stations, the total loss in step S4 is:

[0032] ;

[0033] in, For the total loss, For bounding box regression loss, For confidence loss, For classifying losses, , , To balance the weighting coefficients of different loss terms;

[0034] Bounding box regression loss:

[0035] ;

[0036] ;

[0037] in, The intersection-union ratio (IU) of the predicted bounding box and the ground truth bounding box. and These represent the center coordinates of the predicted bounding box and the ground truth bounding box, respectively. The distance between the center points of the predicted bounding box and the ground truth bounding box is Euclidean distance. The diagonal distance of the smallest closure region that includes both the predicted bounding box and the ground truth bounding box. These are the weighting coefficients. A parameter for measuring aspect ratio consistency, It is the number of positive samples;

[0038] Confidence loss:

[0039] ;

[0040] in, This is the original confidence score of the predicted bounding box. This represents the confidence target for the predicted bounding boxes, with 1 for positive samples and 0 for negative samples. N is the total number of predicted bounding boxes involved in the calculation. It is the Sigmoid function, which maps the predicted value to the (0,1) interval;

[0041] Classification loss:

[0042] ;

[0043] in, This is the original class score output for the predicted bounding boxes of positive samples. The true class label corresponding to the positive sample. It represents the total number of categories.

[0044] In the above-mentioned AI-based unmanned inspection method for water supply and drainage pumping stations, step S5 includes:

[0045] S51: Backpropagation, calculate the gradient of the total loss Ltotal with respect to each weight parameter W in the network. ;

[0046] S52: Optimizer update: The optimizer updates all network parameters according to the calculated gradient and the learning rate.

[0047] .

[0048] Compared with existing technologies, the advantages of this invention are as follows: The system consists of multiple components such as cameras, edge boxes, video centers, and IoT platforms, which are connected through standard interfaces, supporting plug-and-play functionality. This allows the system to be seamlessly integrated with existing pumping station equipment without requiring large-scale infrastructure modifications, and it has good compatibility. The central-end recognition algorithm module is based on an improved YOLO architecture, introducing multi-scale feature fusion, attention mechanisms, and improved target matching and optimized loss functions, resulting in higher recognition accuracy and a lower false alarm rate. The separate architecture of the edge management platform and the central-end recognition algorithm module allows for batch and remote distribution of new models to the edge boxes of all pumping stations after training them on the central platform when new defect types or optimized algorithms are needed. This eliminates the need for technicians to upgrade each pumping station individually on-site, significantly reducing operation and maintenance costs and time. Attached Figure Description

[0049] Figure 1 This is a schematic diagram of the system structure of the present invention;

[0050] Figure 2 This is a system application structure diagram of the present invention;

[0051] Figure 3 This is a graph showing the training process metrics of this invention. Detailed Implementation

[0052] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0053] like Figure 1-3 As shown, an AI-powered unmanned inspection system for water supply and drainage pumping stations includes a camera installed in the pumping station and an edge box connected to the camera. The camera is connected to a push-flow service module and a video center. The edge box is connected to the video center through an edge management platform. The video center is connected to an IoT platform through a central-end recognition algorithm module, and the IoT platform is connected to the edge box. The IoT platform is connected to an information model, which is connected to the video center. The push-flow service module and the information model are connected to an inspection module. The above system is integrated into an AI inspection cabinet and its configured AI inspection platform. The AI ​​inspection cabinet collects data through front-end devices.

[0054] Among them, the edge box is a micro-server designed for edge AI computing scenarios, equipped with an AI computing chip and a hard decoder. It can pull the camera RTSP stream for decoding and second-level algorithm recognition, and upload the recognition data to the Internet of Things platform. The edge management platform can configure and manage edge boxes, cameras, algorithm images, etc., and the algorithm can be configured for edge boxes and cameras through the platform. The video center can configure and manage information such as camera entities and recognition target entities, and can obtain screenshots at regular intervals from the camera for algorithm recognition at the central end.

[0055] At the same time, the push stream service module is used for the camera video playback of the inspection system, and the central end recognition algorithm module is used for low-frequency algorithm recognition. It receives the camera screenshots pushed by the video center for recognition and uploads the recognition data to the Internet of Things platform. The Internet of Things platform is used for the upload, reception, management, and query of IOT data. The information model supports the definition and configuration of entities, categories, alarm rules, etc., makes alarm judgments on the recognition data of the Internet of Things platform, and provides the recognition data and alarm information to the inspection system for display. The core functions of the inspection system are AI duty and AI inspection. AI duty realizes real-time early warning through alarm judgment based on real-time second-level recognition data of edge computing. AI inspection makes alarm judgments and data statistics on the data of edge computing and central end computing at regular intervals, and reports and displays the information of all inspection objects.

[0056] Furthermore, this AI unmanned inspection system further includes a business application layer, which is connected to an information model layer and an AI recognition layer. The information model layer collects data from Internet of Things data sources through the Internet of Things, and the AI recognition layer collects data from video data sources through the security platform. The functions of the business application layer include, but are not limited to, AI inspection, VR inspection, process drawings, real-time monitoring, real-time alarm, safety disclosure, and linkage review. The information model layer includes category management, entity management, network management, and threshold management. The AI recognition layer has inspection tasks, inspection plans, video management, and access control management. The Internet of Things can achieve access management, remote device management, remote management, and data distribution. The security platform includes comprehensive control configuration, video monitoring configuration, and device management. The Internet of Things data sources include interconnected LANs and edge gateways, as well as 4G / 5G / NB-IOT.

[0057] Furthermore, the central end recognition algorithm module includes:

[0058] Input layer: Perform image input splicing to enrich background information to improve the detection ability of small targets, perform adaptive anchor box training and adaptive image scaling. At the beginning of training, re-cluster based on the training set data to generate anchor box sizes more suitable for the data set, uniformly scale the input image to a resolution of 640×640, while maintaining the original aspect ratio to reduce information distortion;

[0059] Backbone Network: Extracts multi-level features from the input image; the Focus module of the backbone network slices the input image, breaking down each 2×2 adjacent pixel region and recombining them onto the channel dimension. Its purpose is to downsample without losing information and reduce computational cost; the C3 module is based on the CSPNet idea, which divides the feature map into two parts on the channel. One part is processed through multiple Bottleneck residual blocks for deep feature extraction, and the other part is directly short-circuited. Finally, the two parts are concatenated. Its purpose is to significantly reduce computational cost and memory consumption while maintaining feature extraction capability and avoiding gradient repetition; the SPPF module performs three 5×5 max pooling operations on the input in sequence and concatenates all pooling results with the original input. Its purpose is to fuse features at different scales without significantly increasing computational cost, greatly increasing the network's receptive field and enabling the model to better understand global contextual information.

[0060] Neck Network: Multi-scale features extracted by the backbone network are fused. The FPN structure upsamples the high semantic features from deep layers and concatenates them with the high-resolution features from shallow layers. This path transfers strong semantic information from high layers to low layers, improving the recognition ability of small targets. The PAN structure downsamples the features fused from shallow layers and concatenates them with the corresponding deep features. This path transfers precise localization information from low layers to the entire network, further improving the localization accuracy of small targets, while also enriching the features of large and medium-sized targets.

[0061] Detection head: Convolves multi-scale feature maps with 1×1 convolution to generate a prediction tensor containing anchor box coordinates, confidence and category information;

[0062] Output layer: performs predicted bounding box decoding and non-maximum suppression (NMS). The network predicts the offset relative to the anchor box. During inference, the predicted value is decoded to obtain the actual bounding box. NMS eliminates duplicate boxes through confidence ranking and IoU suppression, achieving redundant detection.

[0063] An AI-powered unmanned inspection method for water supply and drainage pumping stations, employing the aforementioned AI-powered unmanned inspection system for water supply and drainage pumping stations, includes the following training steps:

[0064] S1: Forward propagation, the input image passes through the backbone network, the neck network, and the head detection network, and the prediction tensors at three scales are obtained at the network output.

[0065] S2: Target and prediction matching: Determine the real object that the predicted box is responsible for detecting, and associate the predicted box with the real box through a matching strategy to ensure that the network can learn the effective object location and category;

[0066] S3: Perform image detection and introduce an attention mechanism;

[0067] S4: Perform loss calculations, including bounding box regression loss, confidence loss, and classification loss;

[0068] S5: Backpropagation and parameter update.

[0069] Furthermore, in step S2, prior anchor box matching, the ground truth box is matched with multiple preset anchor boxes at each location, and the best matching anchor box is selected as a positive sample; a ground truth box is assigned to a prediction box to be responsible for prediction, and the selected prediction box is a positive sample; prediction boxes that do not match ground truth boxes are used as negative samples and participate in the calculation of confidence loss.

[0070] In addition, step S2 calculates the distance between the ground truth bounding box and the predicted bounding box. If the distance is less than a preset threshold, the box is selected. The distance calculation is as follows:

[0071] ;

[0072] in, The degree of matching between the ground truth bounding box and the predicted bounding box. This represents the position of the actual bounding box. The location of the predicted bounding box. This is the covariance matrix of the predicted observation space. This step considers the uncertainty of the predicted box and represents it through the covariance matrix, making the matching process more robust, especially in the case of object deformation or occlusion. Compared with traditional... In comparison, this metric handles unaligned bounding boxes better, improving matching accuracy.

[0073] Meanwhile, step S3 enhances the model's ability to capture key information by weighting the important regions of the feature map, making it suitable for complex scenarios in pump station inspections, such as equipment failure or anomaly detection. It includes the following steps:

[0074] S31: Feature maps of different resolutions are obtained through Backbone. The low-resolution feature map is upsampled and fused with the high-resolution feature map. Feature fusion improves the multi-scale object detection performance by combining deep semantic features and shallow detail features through upsampling and stitching operations, which is especially suitable for equipment parts of different sizes in pumping stations.

[0075] S32: Perform average pooling operation layer by layer on the high-resolution feature maps of each spatial location and fuse them with the low-resolution feature maps. This average pooling operation reduces the size of the feature maps while retaining key information. After fusing with the low-resolution features, it enhances the model's understanding of the global context and improves detection consistency.

[0076] S33: It adopts one-dimensional convolution to calculate the interaction of channel components, and adjusts the channel weights through compression and activation operations to highlight important feature channels, suppress redundant information, and improve feature representation efficiency;

[0077] S34: The Sigmoid function is used for non-linear mapping to achieve soft attention of the feature map, enabling the model to adaptively focus on key regions and improve detection accuracy.

[0078] As can be seen, the total loss in step S4 is:

[0079] ;

[0080] in, For the total loss, The bounding box regression loss is responsible for accurate localization. To differentiate between foreground and background for confidence loss, To ensure correct category prediction for classification loss, , , To balance the weight coefficients of different loss terms, the weight coefficients are used to adjust the relative importance of each loss term, prevent any one term from dominating the training process, and ensure that the model achieves a balance between localization, confidence, and classification. This is suitable for diverse detection tasks in pump station inspection.

[0081] Bounding box regression loss:

[0082] ;

[0083] ;

[0084] in, The intersection-union ratio (IU) of the predicted bounding box and the ground truth bounding box. and These represent the center coordinates of the predicted bounding box and the ground truth bounding box, respectively. The distance between the center points of the predicted bounding box and the ground truth bounding box is Euclidean distance. The diagonal distance of the smallest closure region that includes both the predicted bounding box and the ground truth bounding box. These are the weighting coefficients. A parameter for measuring aspect ratio consistency, It is the number of positive samples; the loss is based on CIoU loss, which not only considers the overlapping area, but also introduces the center point distance and aspect ratio consistency, making the bounding box regression more accurate, especially suitable for the detection of irregularly shaped objects in pump station equipment;

[0085] Confidence loss:

[0086] ;

[0087] in, This is the original confidence score of the predicted bounding box. This represents the confidence target for the predicted bounding boxes, with 1 for positive samples and 0 for negative samples. N is the total number of predicted bounding boxes involved in the calculation. The sigmoid function maps the predicted values ​​to the (0,1) interval; the confidence loss uses binary cross-entropy to supervise whether the predicted box contains an object; the sigmoid function converts the original output into probabilities to ensure training stability; this loss helps the network distinguish between objects and background, reducing false positives;

[0088] Classification loss:

[0089] ;

[0090] The classification loss employs multi-label binary cross-entropy, with each category processed independently, making it suitable for multi-class detection tasks. This is the original class score output for the predicted bounding boxes of positive samples. The true class label corresponding to the positive sample. It represents the total number of categories.

[0091] Clearly, step S5 includes:

[0092] S51: Backpropagation, calculate the gradient of the total loss Ltotal with respect to each weight parameter W in the network. The gradient is calculated using the chain rule, and the gradient is backtracked from the output layer to determine the contribution of each parameter to the loss, providing direction for parameter updates.

[0093] S52: Optimizer update: The optimizer updates all network parameters according to the calculated gradient and the learning rate.

[0094] .

[0095] The optimizer uses gradient information to adjust parameters, and the learning rate η controls the update step size, ensuring training convergence while avoiding oscillations. This process is repeated iteratively, gradually improving the model's detection capability during pump station inspections.

[0096] like Figure 3 The information shown in the actual training output reflects the changing trends of various indicators during the training process and is used to evaluate the training effect and model performance.

[0097] box_loss (bounding box regression loss), obj_loss (confidence loss), and cls_loss (classification loss). A decrease followed by a leveling off indicates that the model has converged.

[0098] Precision: The proportion of all targets predicted by the model that are correctly predicted.

[0099] Recall: The proportion of all true targets that are successfully predicted by the model.

[0100] mAP@0.5: The average accuracy at an IoU threshold of 0.5. This is one of the most critical metrics for evaluating object detection models.

[0101] mAP@0.5:0.95: Average mAP across multiple criteria with IoU thresholds ranging from 0.5 to 0.95 (in steps of 0.05).

[0102] Precision, Recall, and mAP metrics steadily increase and eventually stabilize during training, indicating that the model has converged.

[0103] Example 1

[0104] In this embodiment, the inspection target is the water pump unit itself, requiring the simultaneous detection of both large anomalies and small but critical anomalies. Traditional methods struggle to accurately capture defects of varying sizes simultaneously on a single scale.

[0105] First, multi-scale feature extraction and fusion are performed. A high-resolution image of the water pump unit is input, and after passing through the backbone, three feature maps of different resolutions are obtained. The low-resolution feature map effectively identifies the large target of the pump body and large areas of oil contamination. The high-resolution feature map locates loose bolts and small cracks. Through upsampling and fusion of the neck network, the identification information of the pump body and bolt positions are combined, enabling the algorithm to determine whether the bolts are in normal condition while locating them.

[0106] After feature fusion, a channel attention mechanism is introduced. This mechanism learns autonomously and assigns higher weights to feature channels associated with anomalies. For oil stain detection, feature channels associated with brown patches and reflective properties are enhanced; for metal parts, channels associated with regular edges and rust textures are highlighted, thereby effectively suppressing background interference from pump stations.

[0107] Finally, loss calculation and optimization are performed so that the system can detect large areas of oil stains on the pump body in real time in a single image and accurately locate the missing or loose bolts.

[0108] Example 2

[0109] This embodiment is used to periodically record the readings of pressure gauges and flow meters, and to check whether valves are open or closed. Traditional OCR and template matching methods have poor robustness to changes in lighting, dial contamination, and viewing angle tilt.

[0110] Since instrument readings and valve handwheel markings are fine-grained targets, the algorithm fully utilizes high-resolution feature maps to preserve these details. Through feature fusion, it combines the pressure gauge type information understood by the deep network with the pixel-level information of the needle tip markings captured by the shallow network, achieving end-to-end detection and identification.

[0111] To address issues such as glare from the dial glass and water stains, the channel attention mechanism learns to eliminate interfering information such as reflective areas and reduces the weight of corresponding feature channels, while increasing the weight of key channels such as pointer edges and scale numerals. For valve status recognition, the attention mechanism focuses on decisive features such as the direction of the handwheel groove or the extension length of the valve stem, without being overly affected by the valve's color or excessive corrosion.

[0112] For valves, they can be directly classified into two categories: open and closed. For instruments, a segmented classification strategy can be adopted: the range is divided into N equal intervals, transforming the reading problem into an N-class classification problem. This approach is more stable and has stronger anti-interference capabilities compared to regression fitting.

[0113] In summary, the principle of this embodiment is as follows: First, features are extracted using a forward propagation network architecture integrating multi-scale feature fusion and attention mechanisms. Then, a top-down and bottom-up feature pyramid fusion is performed via a neck network to enhance the synergistic expression of semantic information and spatial details. Next, a channel attention mechanism is introduced to achieve adaptive recalibration of feature channels through one-dimensional convolution and Sigmoid mapping to focus on key information. Based on this, an advanced matching strategy based on Mahalanobis distance is adopted to match ground truth boxes with anchor boxes that consider prediction uncertainty to accurately define positive and negative samples. Furthermore, a composite loss function is used to comprehensively quantify prediction errors, incorporating a bounding box regression loss that considers center point distance and aspect ratio consistency, a confidence loss combining binary cross-entropy, and a classification loss. Finally, using the total loss gradient calculated during backpropagation, the optimizer drives iterative updates of all network weights, thereby systematically improving the robustness and accuracy of the model in accurately detecting and classifying equipment status in complex pumping station environments.

[0114] The specific embodiments described herein are merely illustrative of the spirit of the invention. Those skilled in the art to which this invention pertains may make various modifications or additions to the described specific embodiments or use similar methods to substitute them, without departing from the spirit of the invention or exceeding the scope defined by the appended claims.

[0115] Although this document uses terms such as camera, edge box, and streaming service module frequently, the possibility of using other terms is not excluded. These terms are used merely for the convenience of describing and explaining the essence of this invention; interpreting them as any additional limitation would contradict the spirit of this invention.

Claims

1. An AI-powered unmanned inspection system for water supply and drainage pumping stations, comprising a camera installed in the pumping station and an edge box connected to the camera, characterized in that, The camera is connected to a streaming service module and a video center. The edge box is connected to the video center via an edge management platform. The video center is connected to an IoT platform via a central-end recognition algorithm module, and the IoT platform is connected to the edge box. The IoT platform is connected to an information model, which is connected to the video center. The streaming service module and the information model are connected to the inspection module. The inspection method used in this inspection system employs the following training steps: S1: Forward propagation, the input image passes through the backbone network, the neck network, and the head detection network, and the prediction tensors at three scales are obtained at the network output. S2: Match the target with the prediction to determine the real object that the prediction box is responsible for detecting; Prior anchor box matching involves matching the ground truth bounding box with multiple pre-set anchor boxes at each location, and selecting the best matching anchor box as a positive sample. A ground truth bounding box is assigned to a prediction box to be responsible for prediction, and the selected prediction box is a positive sample. Prediction boxes that do not match ground truth bounding boxes are used as negative samples and participate in the calculation of confidence loss. Calculate the distance between the ground truth bounding box and the predicted bounding box. If the distance is less than a preset threshold, the box is selected. The distance calculation is as follows: ； in, The degree of matching between the ground truth bounding box and the predicted bounding box. This represents the position of the actual bounding box. The location of the predicted bounding box. The covariance matrix of the predicted observation space; S3: Perform image detection and introduce an attention mechanism; S4: Perform loss calculations, including bounding box regression loss, confidence loss, and classification loss; S5: Backpropagation and parameter update.

2. The AI-powered unmanned inspection system for water supply and drainage pumping stations according to claim 1, characterized in that, The inspection module includes an AI monitoring unit and an AI inspection unit.

3. The AI-powered unmanned inspection system for water supply and drainage pumping stations according to claim 1, characterized in that, This AI-powered unmanned inspection system includes a business application layer, which is connected to an information model layer and an AI recognition layer. The information model layer collects data from IoT data sources through the Internet of Things (IoT), and the AI ​​recognition layer collects data from video data sources through a security platform.

4. The AI-powered unmanned inspection system for water supply and drainage pumping stations according to claim 1, characterized in that, The central identification algorithm module includes: Input layer: performs image input stitching, adaptive anchor box training, and adaptive image scaling; Backbone network: Extracts multi-level features from the input image; Neck network: Multi-scale features extracted from the backbone network are fused; Detection head: Convolves multi-scale feature maps with 1×1 convolution to generate a prediction tensor containing anchor box coordinates, confidence and category information; Output layer: performs prediction box decoding and non-maximum suppression (NMS).

5. The AI-powered unmanned inspection system for water supply and drainage pumping stations according to claim 1, characterized in that, Step S3 includes the following steps: S31: Obtain feature maps of different resolutions through Backbone, upsample the low-resolution feature map and fuse it with the high-resolution feature map; S32: Perform average pooling operation layer by layer on the high-resolution feature maps at each spatial location and fuse them with the low-resolution feature maps. S33: One-dimensional convolution is used to calculate the interaction between channels; S34: Use the Sigmoid function for nonlinear mapping.

6. The AI-powered unmanned inspection system for water supply and drainage pumping stations according to claim 1, characterized in that, Total loss in step S4: ； in, For the total loss, For bounding box regression loss, For confidence loss, For classifying losses, , , To balance the weighting coefficients of different loss terms; Bounding box regression loss: ； ； in, The intersection-union ratio (IU) of the predicted bounding box and the ground truth bounding box. and These represent the center coordinates of the predicted bounding box and the ground truth bounding box, respectively. The distance between the center points of the predicted bounding box and the ground truth bounding box is Euclidean distance. The diagonal distance of the smallest closure region that includes both the predicted bounding box and the ground truth bounding box. These are the weighting coefficients. A parameter for measuring aspect ratio consistency, It is the number of positive samples; Confidence loss: ； in, This is the original confidence score of the predicted bounding box. This represents the confidence target for the predicted bounding boxes, with 1 for positive samples and 0 for negative samples. N is the total number of predicted bounding boxes involved in the calculation. It is the Sigmoid function, which maps the predicted value to the interval (0, 1); Classification loss: ； in, This is the original class score output for the predicted positive bounding boxes. The true class label corresponding to the positive sample. It represents the total number of categories.

7. The AI-powered unmanned inspection system for water supply and drainage pumping stations according to claim 1, characterized in that, Step S5 includes: S51: Backpropagation, calculate the gradient of the total loss Ltotal with respect to each weight parameter W in the network. ; S52: Optimizer update: The optimizer updates all network parameters according to the calculated gradient and the learning rate. 。

Images