Iron tower detection and early warning system based on convolutional neural network

Abstract

The invention discloses an iron tower detection and early warning system based on a convolutional neural network, and relates to the technical field of iron tower detection, the system comprises a collection end, a processing end and an output end, the collection end is used for obtaining real-time images and sensor data of an iron tower, the processing end trains a lightweight CNN recognition model based on collected labeled samples, and the output end outputs the lightweight CNN recognition model. The model is used to carry out abnormity identification on real-time data; meanwhile, an automatic verification mechanism based on time sequence slope change is introduced into the system, abnormal data of the sensor are intelligently judged, and misinformation and missing report are reduced; and the output end sends out early warning signals of different levels according to the identification result in combination with a preset alarm mapping table. Through multi-modal data fusion and an intelligent verification mechanism, the accuracy of iron tower anomaly detection and the early warning response efficiency are remarkably improved, manual intervention is effectively reduced, and the safety guarantee capability of iron tower operation is improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of iron tower detection, and in particular to an iron tower detection and early warning system based on a convolutional neural network. Background Art

[0002] As the number of power and telecommunications towers grows, their safe operation is crucial to maintaining the stability of power grids and communication networks. Currently, tower inspections rely primarily on manual inspections, drone aerial photography, and sensor monitoring. While drones and sensors have improved automation, manual inspections remain inefficient and pose safety risks.

[0003] Patent publication number CN116823192A discloses an intelligent tower inspection system, comprising an inspection information input module, an inspection order generation module, an inspection equipment group, a cloud-based database, an inspection equipment allocation module, and an inspection project verification module. By incorporating inspection equipment that does not require tower climbing, the system enables automated collection and management of inspection data, reducing manual intervention and improving data integrity and accuracy. However, the system fails to fully address the issues of accurately identifying abnormal conditions based on multimodal data and automatically verifying abnormal data.

[0004] Existing detection systems still face challenges in anomaly identification and early warning. First, tower anomaly detection requires comprehensive analysis of multimodal data. Efficiently and accurately integrating image and sensor information for precise identification remains a technical challenge. Second, early warning systems struggle to automatically verify and correct anomaly data, leading to delayed responses and frequent false alarms and missed alerts, compromising safety management effectiveness.

[0005] Based on the above-mentioned deficiencies of the existing technology, there is an urgent need for a tower detection and early warning system based on convolutional neural networks, which can integrate tower images and sensor data, accurately identify abnormal tower conditions, and have the ability to automatically verify and correct abnormal data, thereby improving the intelligence level and reliability of tower safety monitoring. Summary of the Invention

[0006] In order to solve the problems of insufficient accuracy in abnormal state recognition and limited ability to automatically verify abnormal data in existing tower detection and early warning systems.

[0007] To this end, the present invention provides a tower detection and early warning system based on a convolutional neural network, comprising: The collection end is used to collect normal sample data and abnormal sample data, as well as real-time images of the tower and real-time sensor detection values; The processing end is used to train the recognition model based on the annotated abnormal samples and normal samples. After the training is completed, the recognition model is used to identify the real-time images of the tower and the real-time sensor data detected by the sensor, and the recognition results are output; Abnormal sample data includes abnormal images of several towers and abnormal data detected by sensors when the towers are abnormal. Normal sample data includes normal images and normal data obtained by sensors when the towers are normal. The output terminal is used to generate corresponding alarms according to the recognition results when the recognition results are obtained, and to issue different alarms according to the alarm mapping table set in advance. The alarm mapping table contains different abnormal conditions and their corresponding alarm levels; The processing end is further configured to output abnormal conditions and corresponding alarm levels to the output end.

[0008] Furthermore, the method for training a recognition model based on the labeled abnormal samples and normal samples includes the following steps: All abnormal sample data and normal sample data are labeled and divided into training set, validation set and test set in a ratio of 7:1.5:1.5; The labeled training set is put into the recognition model to extract features with the help of the recognition model. Then, the parameters of the trained recognition model are adjusted with the help of the validation set, and the generalization ability of the model is evaluated by the test set, mainly through the recognition accuracy feedback. When the recognition accuracy exceeds the set threshold, it means that the recognition model training is completed.

[0009] Furthermore, before sample labeling, the image data is quality screened to remove blurred, overexposed, and severely occluded images.

[0010] Furthermore, the recognition model adopts a lightweight CNN network as the backbone structure, and the CBAM attention module is embedded in the network backbone structure to enhance the feature extraction capability of key components of the tower, and improves the small component detection effect through the FPN feature pyramid and anchor point adaptation mechanism.

[0011] Furthermore, the training strategy of the recognition model includes: The loss function integrates classification cross entropy, positioning loss CloU and weight balance term; At the same time, SMOTE oversampling or focal loss mechanism is used for abnormal samples; The training parameters are set as follows: learning rate 3e-4, batch size 16, optimizer AdamW, and early stopping threshold 5 rounds; Model training is performed on a training platform consisting of four NVIDIA A100 GPUs, each with 80GB of video memory. It uses the PyTorch framework and supports mixed precision training.

[0012] Furthermore, when any sensor detects an abnormality in the data, the data will be automatically defined as abnormal data, and the top N corresponding abnormal data will be automatically obtained. prevThe sensor detection value at each time point, the N prev The preset number of retained data points is used to form a data series; A two-dimensional coordinate system is constructed with the time point as the horizontal coordinate and the corresponding sensor detection value as the vertical coordinate. The slope between adjacent data points in the data sequence is calculated to obtain several slope values, of which the last slope is defined as the tail slope. The difference between the tail term slope and its previous slope, as well as the difference between the maximum and minimum slopes other than the tail term slope, is used to determine whether to trigger a verification signal.

[0013] Furthermore, the step of determining whether to trigger a verification signal includes: Obtain the difference between the maximum and minimum values of all slopes except the tail slope, and define it as the mean difference; Get the difference between the slope of the tail term and its previous slope, which is defined as the tail term difference; When the result of the tail term difference minus the mean term difference exceeds the preset threshold Δ threshold When , a verification signal is triggered, indicating that the data detected by the sensor at this time may be abnormal and needs to be verified; The verification signal, the corresponding sensor and the detection data are integrated to form verification information; the processing end is used to output the verification information to the output end.

[0014] Compared with the prior art, the present invention has the following beneficial effects: 1. This invention significantly improves the recognition accuracy and real-time response capability of abnormal tower conditions by effectively integrating real-time images of the tower with multi-source sensor data, combined with a lightweight convolutional neural network and an attention mechanism, breaking through the bottleneck of insufficient recognition accuracy of traditional single data sources. 2. The introduction of an automatic verification mechanism based on the slope change of the sensor data time series can intelligently determine the authenticity of anomaly detection data, reduce the occurrence of false positives and missed positives, and enhance the early warning system's error correction capability for anomaly data and overall stability; 3. This invention automates the entire process, from data collection and model training to warning output, reducing the frequency and workload of manual inspections and improving detection efficiency and safety. Furthermore, a hierarchical warning mechanism can issue timely alerts for different abnormal situations, assisting operations and maintenance personnel in responding quickly and enhancing the intelligence and reliability of tower safety management. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 This is a flow chart of the tower detection and early warning system based on convolutional neural network of the present invention. DETAILED DESCRIPTION

[0016] The technical solutions of the present invention will be clearly and completely described below in conjunction with the embodiments. Obviously, the embodiments described are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0017] Example 1 See also Figure 1 ,This application provides a tower detection and early warning system based on convolutional neural network,,the system includes: an acquisition end, a processing end and an output end; The collection end is used to collect tower data, mainly real-time tower images of key parts, such as the tower body, guy wires, insulators, etc., which are pre-set by the user. This can be achieved by installing waterproof and dustproof cameras around the tower that support night infrared. It can also be used with high-definition cameras on drones to shoot the tower at preset angles along a preset route. At the same time, data is collected with the help of integrated sensors, such as the tower's inclination and vibration frequency, through tilt sensors and vibration sensors; the scenes can cover sunny days, rainy days, foggy days, nights, and different seasons; At the same time, a number of abnormal sample data are collected. The abnormal sample data includes several abnormal images and abnormal data. Specifically, the abnormal sample data refers to the simulated tower tilt, which is reflected by the tilt data obtained by the tilt sensor and the image of the tower tilt. Faults such as missing parts such as screws and insulators, rust, and foreign objects hanging are reflected here through abnormal images. The acquisition end is used to transmit abnormal sample data and normal sample data to the processing end, wherein the normal sample data is the image and sensor data of the tower in a normal state, and the abnormal sample data is the image and corresponding sensor data in an abnormal state; The processing end is used to train and identify the recognition model based on the sample data formed by the fusion of abnormal sample data and normal sample data. The specific training and identification methods are as follows: First, the image data is quality screened to remove blurry, overexposed, or severely occluded images from abnormal samples; Use Labelmg or CVAT tools to annotate the tower's key components with bounding boxes and generate a COCO dataset. The annotations include information about the tower in normal and tilted states, as well as damage to specific components. The annotation error is kept within 5 pixels. The labeled content is defined as a dataset, and then the dataset is divided into training set, validation set and test set according to the ratio of 7:1.5:1.5; The recognition model is used to train the dataset. The main method is to feed the labeled training set into the recognition model, extract features with the help of the recognition model, and then adjust the parameters of the trained recognition model using the validation set. The model is then fed into the test set to evaluate the generalization ability of the model, mainly through feedback on the recognition accuracy. When the recognition accuracy exceeds the set threshold, it means that the recognition model passes and can be used for recognition. The recognition model described here uses a lightweight CNN network as the backbone structure, such as EfficientDet and YOLOv5s, to balance speed and accuracy. It also outputs tower classification, such as normal or abnormal, and fault location, through bounding box calibration. The CBAM attention module is integrated into the network backbone structure to enhance the feature extraction capability of key tower components; at the same time, the detection effect of small components is improved through feature pyramid FPN and anchor point adaptive adjustment.

[0018] The training strategy of the recognition model includes: using the training set to train the model, and using classification cross entropy loss, positioning loss CloU and weight balance term to form a loss function.

[0019] At the same time, SMOTE oversampling or focal loss strategy is applied to abnormal samples to alleviate sample imbalance.

[0020] The training parameters are set as follows: learning rate is 3e-4, batch size is 16, optimizer is AdamW, early stopping strategy is set, and the early stopping round threshold is 5.

[0021] Model training is performed on a training platform consisting of four NVIDIA A100 GPUs, each with 80GB of video memory. It uses the PyTorch framework and supports mixed precision training.

[0022] The trained recognition model is used to identify the tower's real-time data and obtain the recognition results; the real-time data includes the tower's real-time image and the real-time sensor data detected by the sensor; When the recognition result is obtained, the corresponding alarm is generated according to the recognition result, and different alarms are issued according to the alarm mapping table set in advance. The alarm mapping table contains different abnormal conditions and their corresponding alarm levels; The processing end is further configured to output abnormal conditions and their corresponding alarm levels to the output end to implement early warning notifications.

[0023] When any sensor detects abnormal data, the system automatically marks the data type as abnormal data and automatically obtains the N previous data of the corresponding abnormal data. prev data, N prev The number of data points reserved for the preset.

[0024] To form a data sequence, the system will automatically collect the real-time detection values of the sensor at a set period and retain them in sequence. If the number of retained data points exceeds the preset N prev After that, if no exception occurs, the earliest data will be automatically eliminated when the new data is collected and retained, and only the latest N prev Retain values to form a sliding data window of fixed length; Once an anomaly is detected, the system will prev The retained values are used as the data series before the abnormality for subsequent slope calculation and verification analysis.

[0025] When a sensor detects abnormal data at a certain moment, the system constructs a two-dimensional coordinate system with the time point as the horizontal axis and the corresponding sensor detection value as the vertical axis. The system then sequentially calculates the slope between two adjacent points in the data sequence to obtain several slope values, the last of which is the tail slope, corresponding to the time point when the abnormality occurred. Then, the system performs the following verification and judgment operations: Obtain the difference between the maximum and minimum values of all slopes except the tail slope, and define it as the mean difference; Then the difference between the slope of the tail term and its previous slope is obtained, which is defined as the tail term difference; When the result of the tail term difference minus the mean term difference exceeds the preset threshold Δ threshold When the sensor detects an abnormality in the data, the system automatically triggers the verification signal. When the verification signal is triggered, the system integrates the verification signal, the corresponding sensor number and its detection value to form a verification message, and the processing end outputs the verification information to the output end for further analysis or manual review by the on-duty personnel or the back-end system.

[0026] In this way, the present invention can achieve rapid verification of sensor abnormal data, effectively improve the accuracy and reliability of abnormality detection, and reduce the risks of false positives and missed positives.

[0027] The above embodiments are only used to illustrate the technical method of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical method of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical method of the present invention.

Claims

1. The tower detection and early warning system based on convolutional neural network is characterized by: include: The collection end is used to collect normal sample data and abnormal sample data, as well as real-time images of the tower and real-time sensor detection values; The processing end is used to train the recognition model based on the annotated abnormal samples and normal samples. After the training is completed, the recognition model is used to identify the real-time images of the tower and the real-time sensor data detected by the sensor, and the recognition results are output; Abnormal sample data includes abnormal images of several towers and abnormal data detected by sensors when the towers are abnormal. Normal sample data includes normal images and normal data obtained by sensors when the towers are normal. The output terminal is used to generate corresponding alarms according to the recognition results when the recognition results are obtained, and to issue different alarms according to the alarm mapping table set in advance. The alarm mapping table contains different abnormal conditions and their corresponding alarm levels; The processing end is further configured to output abnormal conditions and corresponding alarm levels to the output end.

2. The tower detection and early warning system based on convolutional neural network according to claim 1 is characterized in that: The method for training a recognition model based on labeled abnormal samples and normal samples includes the following steps: All abnormal sample data and normal sample data are labeled and divided into training set, validation set and test set in a ratio of 7:1.5:1.5; The labeled training set is put into the recognition model to extract features with the help of the recognition model. Then, the parameters of the trained recognition model are adjusted with the help of the validation set, and the generalization ability of the model is evaluated by the test set, mainly through the recognition accuracy feedback. When the recognition accuracy exceeds the set threshold, it means that the recognition model training is completed.

3. The tower detection and early warning system based on convolutional neural network according to claim 2, characterized in that: Before sample labeling, the image data is quality screened to remove blurred, overexposed, and severely occluded images.

4. The tower detection and early warning system based on convolutional neural network according to claim 3 is characterized in that: The recognition model adopts a lightweight CNN network as the backbone structure. The CBAM attention module is embedded in the network backbone structure to enhance the feature extraction capability of key components of the tower, and improves the small component detection effect through the FPN feature pyramid and anchor point adaptation mechanism.

5. The tower detection and early warning system based on convolutional neural network according to claim 4 is characterized in that: The training strategy of the recognition model includes: The loss function integrates classification cross entropy, positioning loss CloU and weight balance term; At the same time, SMOTE oversampling or focal loss mechanism is used for abnormal samples; The training parameters are set as follows: learning rate 3e-4, batch size 16, optimizer AdamW, and early stopping threshold 5 rounds; Model training is performed on a training platform consisting of four NVIDIA A100 GPUs, each with 80GB of video memory. It uses the PyTorch framework and supports mixed precision training.

6. The tower detection and early warning system based on convolutional neural network according to claim 1 is characterized in that: When any sensor detects an abnormal data, it will automatically define the data as abnormal data and automatically obtain the top N corresponding abnormal data. prev The sensor detection value at each time point, the N prev The preset number of retained data points is used to form a data series; A two-dimensional coordinate system is constructed with the time point as the horizontal coordinate and the corresponding sensor detection value as the vertical coordinate. The slope between adjacent data points in the data sequence is calculated to obtain several slope values, of which the last slope is defined as the tail slope. The difference between the tail term slope and its previous slope, as well as the difference between the maximum and minimum slopes other than the tail term slope, is used to determine whether to trigger a verification signal.

7. The tower detection and early warning system based on convolutional neural network according to claim 6 is characterized in that: The step of determining whether to trigger the verification signal includes: Obtain the difference between the maximum and minimum values of all slopes except the tail slope, and define it as the mean difference; Get the difference between the slope of the tail term and its previous slope, which is defined as the tail term difference; When the result of the tail term difference minus the mean term difference exceeds the preset threshold Δ threshold When , a verification signal is triggered, indicating that the data detected by the sensor at this time may be abnormal and needs to be verified; The verification signal, the corresponding sensor and the detection data are integrated to form verification information; the processing end is used to output the verification information to the output end.

Images