CN-122024004-A - Ground knife opening and closing state intelligent detection system and method based on improved YOLO model

CN122024004ACN 122024004 ACN122024004 ACN 122024004ACN-122024004-A

Abstract

The invention discloses an intelligent detection system and method for the state of a ground knife switch-on/off based on an improved YOLO model, which belong to the technical field of visual recognition detection of the state of power equipment and are suitable for an automatic monitoring scene of a transformer substation. The method comprises the following steps of data set construction, model configuration, model training, model optimization and screening and real-time detection. The system and the method solve the problems of high safety risk, difficult compromise between precision and speed and poor robustness of the traditional deep learning method by two types of marked data set design, light weight and balance model combination and multi-index model screening.

Inventors

PAN BIN
Lv kai
WANG XINJUN
NIE SHENGDONG
ZHAO CHUNYU
Guo Shifa
LU HONGSHEN
ZHAI QINGYU
LIU ZHIYUAN

Assignees

山东泰开成套电器有限公司

Dates

Publication Date: 20260512
Application Date: 20251229

Claims (10)

1. An intelligent detection system for the opening and closing state of a ground knife based on an improved YOLO model is characterized by comprising the following modules: the data acquisition module is used for acquiring a ground knife switching-on/off operation video and intercepting an image sample; The data labeling module is used for carrying out two types of labeling on the image sample, wherein the first type of labeling takes the whole moving knife part of the ground knife as a labeling target, and the second type of labeling takes three moving knife heads of the ground knife as labeling targets respectively and labels of the types; The data set dividing module is used for dividing the marked image sample into a training set, a testing set and a verification set according to the ratio of 6:3:1; The model training module is used for loading a YOLO11s or YOLO11n model, configuring training parameters, training the model by utilizing a training set, and adjusting the model parameters through a verification set; The model evaluation module is used for performing performance evaluation on the trained model by utilizing the test set and outputting the accuracy rate, recall rate, mAP50 and mAP50-95 indexes; And the state detection module is used for loading the screened optimal model, receiving the image to be detected and outputting the ground knife state detection result.
2. The intelligent detection system for the opening and closing state of the earth knife based on the improved YOLO model of claim 1 is characterized by further comprising the following modules: The data preprocessing module is used for denoising, size normalization and data enhancement processing on the intercepted image samples, wherein the data enhancement processing comprises random overturn, brightness adjustment and Gaussian blur.
3. The intelligent detection method for the opening and closing state of the earth knife based on the improved YOLO model is characterized by being implemented by adopting the intelligent detection system for the opening and closing state of the earth knife based on the improved YOLO model, and comprises the following steps: Step one, constructing a data set; collecting a ground knife switching-on/off operation video, and intercepting an image sample from the ground knife switching-on/off operation video to construct an original data set; Two types of labeling are carried out on the image sample, wherein the first type of labeling takes the whole moving knife part of the ground knife as a labeling target, and the second type of labeling takes three moving knife heads of the ground knife as labeling targets respectively; Dividing the marked sample into a training set, a testing set and a verification set according to a preset proportion, wherein marking categories comprise 0 class, 1 class and 2 class; Step two, configuring a model; selecting a YOLO11 series model as a basic detection model, including a YOLO11s balance model and a YOLO11n lightweight model, and then configuring model training parameters; step three, model training; Respectively inputting the data sets corresponding to the two labeling modes in the first step into the two YOLO11 models configured in the second step, and performing model training to obtain four groups of trained ground knife state detection models; step four, model optimization and screening; Performing performance evaluation on the four groups of trained models through a test set, wherein evaluation indexes comprise accuracy rate, recall rate, mAP50 and mAP50-95, and screening out a model with optimal comprehensive performance as a final ground cutter state detection model; step five, detecting in real time; And inputting the ground knife operation image to be detected into a final ground knife state detection model, and outputting detection results of opening, closing or other abnormal states of the ground knife.
4. The method for intelligently detecting the opening and closing states of a ground knife based on an improved YOLO model as claimed in claim 3, wherein in the step of constructing a data set, the original data set contains 1056 image samples.
5. The intelligent detection method for the ground knife opening and closing state based on the improved YOLO model is characterized in that in the step of constructing a data set, class 0 represents opening, class 1 represents closing, class 2 represents other abnormal states, and the other abnormal states comprise semi-closing, mechanical failure and foreign matter shielding.
6. The intelligent detection method for the switch-on and switch-off states of the earth knife based on the improved YOLO model of claim 3 is characterized in that in the step of constructing a data set, the preset ratio of the training set, the test set and the verification set is 6:3:1.
7. The intelligent detection method for the switch-on and switch-off states of the earth-knife based on the improved YOLO model of claim 3, wherein in the step of constructing a data set, the image samples are obtained through multi-scene earth-knife operation video interception, and different illumination conditions, equipment operation angles and environment interference scenes are covered.
8. The intelligent detection method for the ground knife opening and closing state based on the improved YOLO model according to claim 3, wherein in the step of model configuration, the step of configuring model training parameters is characterized in that the image input size is uniformly set to 640×640, the training batch size is 16, the training round is 100, and the self-adaptive moment estimation optimizer and CIoU loss function are adopted to optimize the bounding box regression accuracy and the model convergence speed.
9. The intelligent detection method for the states of opening and closing of the earth-boring cutter based on the improved YOLO model of claim 3 is characterized in that in the step of model training, four groups of earth-boring cutter state detection models are YOLO11s+first class labels, YOLO11s+second class labels, YOLO1n+first class labels and YOLO1n+second class labels in sequence.
10. The intelligent detection method for the ground knife opening and closing state based on the improved YOLO model, which is disclosed in claim 3, is characterized in that in the step of model optimization and screening, the model with optimal comprehensive performance meets the conditions that the accuracy rate of opening and closing states is more than or equal to 95%, the recall rate is more than or equal to 99%, the mAP50 is more than or equal to 98%, the accuracy rate of other abnormal states is more than or equal to 93%, the recall rate is more than or equal to 80%, and the mAP50 is more than or equal to 92%.

Description

Ground knife opening and closing state intelligent detection system and method based on improved YOLO model Technical Field The invention relates to the technical field of visual identification and detection of power equipment states, in particular to an intelligent detection system and method for a ground knife opening and closing state based on an improved YOLO model, which are suitable for automatic monitoring scenes of transformer substations and can realize high-precision, rapid and robust detection of the ground knife state. Background The grounding switch (called a ground knife for short) is a key device in the high-voltage switch cabinet, and the accuracy of the opening and closing state of the grounding switch is directly related to the safe and stable operation of the power system. Traditional ground knife state confirmation mainly relies on manual inspection, and staff judges the ground knife state through on-site observation or by means of a simple instrument, but the mode has the problems of high safety risk, low efficiency, high labor intensity, easiness in being influenced by human factors and environmental conditions and the like, and especially in complex transformer substation environments, misjudgment and missed judgment conditions are easy to occur, so that hidden danger is brought to the safety of a power system. With the development of computer vision and deep learning technology, the power equipment state detection method based on image recognition gradually becomes a research hot spot. The existing ground knife state detection system and method based on deep learning have the following general problems: the data set is marked singly, namely, the whole ground knife (such as a moving knife part) is marked, the details of the knife head are not concerned, and the recognition capability of an abnormal state (such as a half-closing) is weak; the model selection pertinence is insufficient, the precision and the speed are not considered, the heavy model has high precision but poor real-time performance, and the light model has high speed but easy omission; the robustness is poor, the adaptability to multiple scenes (different illumination, equipment angles and environmental interference) is insufficient, and the actual requirements of automatic monitoring of a transformer substation can not be met. Disclosure of Invention The invention aims to solve the technical problem of providing an intelligent detection method for the opening and closing state of a ground knife based on an improved YOLO model, which solves the problems of high safety risk, difficult compromise between the precision and the speed of the existing deep learning method and poor robustness of the traditional method by means of two marked data set designs, light weight and balance model combination and multi-index model screening. In order to achieve the above purpose, the present invention adopts the following technical scheme: An intelligent detection method for the state of a ground knife switch on/off based on an improved YOLO model realizes the efficient and accurate detection of the state of the ground knife through an innovative data set construction mode and an optimized model training strategy, and comprises the following steps: Step one, a data set construction step; And acquiring ground knife opening and closing operation videos under different scenes, wherein the videos comprise scenes with different illumination intensities, equipment operation angles, environmental interference and the like, and capturing 1056 clear images from the scenes as original data set samples. Labeling the sample by adopting two labeling modes: The first type of marking, namely directly marking the state type of the whole movable knife by taking the whole movable knife part of the ground knife as a marking target; Marking, namely respectively marking the state type of each cutter head by taking three movable cutter heads of the ground cutter as independent marking targets; The labeling categories are uniformly divided into three categories, namely class 0 (opening), class 1 (closing) and class 2 (other abnormal states including semi-closing, mechanical failure, foreign object shielding and the like). Randomly dividing the marked samples into a training set, a testing set and a verification set according to the proportion of 6:3:1, and ensuring the distribution balance of the data set. Step two, configuring a model; selecting a YOLO11 series model as a basic detection model, wherein the series model has the characteristics of light weight, high detection speed and high precision, and the method comprises the following steps: The YOLO11s model has balanced speed and precision, and is suitable for scenes with higher requirements on detection precision; The YOLO11n model has higher light weight degree, smaller parameter quantity and higher detection speed, and is suitable for real-time monitoring of scenes; The training parameters of the mo