Pumped storage unit insulation defect detection method based on robot combined with improved RT-DETR algorithm

Abstract

The invention discloses a pumped storage group insulation defect detection method based on a robot combined with an improved RT-DETR algorithm, and the method comprises the following steps: S1, collecting image data in a pumped storage group chamber through a robot under the condition of not drawing a rotor, and taking the image data as an original data set; s2, marking defect types and defect positions in the original data set, processing the original data set by using a data enhancement technology, and segmenting to obtain a training set, a verification set and a test set; s3, constructing an improved RT-DETR neural network model, training the neural network by using the training set and the verification set, and evaluating the neural network model by using the test set; and S4, deploying the defect detection model into a visual system processing end, processing data returned by the robot during operation, and carrying out defect detection on the rotor of the pumped storage unit. According to the invention, under the condition that the rotor is not taken out, the fault defect of the rotor of the pumped storage generator set, which is easy to process, can be found in time.

Description

Technical Field

[0001] The invention relates to the technical field of insulation defect detection of pumped storage units, and in particular to an insulation defect detection method of a pumped storage unit based on a robot combined with an improved RT-DETR (rtdetr_r18vd_6x_coco) algorithm. Background Art

[0002] With the development of green power generation technology, the proportion of new energy sources such as hydropower and wind power in power supply has gradually increased. However, new energy power generation is greatly restricted by natural conditions and has the characteristics of "relying on the weather", which leads to insufficient stability of power supply. In this context, pumped storage power stations can just meet the urgent demand for flexible power regulation, and can be said to be an important cornerstone for ensuring the safe and stable operation of the new power system with new energy as the main body.

[0003] The generators in operation will inevitably be affected by the high magnetic stress of the power grid, the transient shocks with potential damage (such as voltage deviation, frequency fluctuation and negative phase sequence current, etc.), and the high mechanical torque and centrifugal force from the turbine rotor, which will cause the stator of the generator to vibrate to varying degrees. If the fault defects that are easier to handle can be discovered and eliminated in time, such adverse events can be avoided as much as possible.

[0004] Robot technology continues to develop and progress, and the newly developed advanced robots have performed exceptionally well in the inspection and status evaluation of multiple generators and have played a prominent role. However, in terms of non-disassembly maintenance objects, there are few studies on wall-climbing robots that can complete the maintenance process on site, especially in the field of fault detection robots in large generator chambers, where relevant research is seriously lacking. In addition, most generators in power plants in China currently have the problem of excessive rotor extraction for maintenance, and there is an urgent need to expand research in this field to solve this problem. Therefore, it is urgent to study a reliable defect detection method that can detect unit insulation defects with less or no rotor extraction. Summary of the invention

[0005] The technical problem to be solved by the present invention is to provide a pumped storage unit insulation defect detection method based on a robot combined with an improved RT-DETR algorithm, which can sensitively and reliably realize the detection of rotor insulation defects of a variable speed pumped storage unit.

[0006] In order to solve the above technical problems, the technical solutions adopted by the present invention are as follows.

[0007] A method for detecting insulation defects of a pumped storage unit based on a robot combined with an improved RT-DETR algorithm comprises the following steps: S1. Use the robot as the data acquisition terminal to collect image data in the chamber of the pumped storage unit without pumping the rotor as the original data set; S2. Mark the defect types and defect locations in the original data set, process the original data set using data enhancement technology, and split the data into training set, validation set, and test set; S3. Build an improved RT-DETR neural network model, train the neural network with the training set and the validation set, evaluate the neural network model with the test set, perform multiple rounds of training, and select the best model that meets the industrial standard requirements as the defect detection model in multiple rounds of training and testing; S4. Deploy the defect detection model to the visual system processing end, process the data sent back by the robot at the data acquisition end during operation, and perform defect detection on the rotor of the pumped storage unit.

[0008] Preferably, in step S1, a crawler motor-driven inspection robot equipped with an image acquisition device having high resolution and strong adaptability to low illumination environments is used to collect images inside the chamber of the pumped-storage unit through preset path planning as original data.

[0009] Preferably, marking the defect types and defect locations in the original data set in step S2 includes: manually labeling the original image with lableme software in a supervised manner, and the defect types include damaged insulation on the stator surface, discoloration and displacement of the stator core laminations, visible dust and white deposits on the surface of the slot wedges in the stator bore, traces of electric current damage on the rotor surface, clogged rotor ventilation holes, and loose rotor dynamic balancing bolts.

[0010] Preferably, the processing of the original data set using data enhancement technology in step S2 includes: Use the Albumentations library to perform image transformation operations, and perform geometric transformation, elastic transformation, and grid distortion operations on the tensor of the original image in sequence. Each transformation generates 3 images. The processed data are combined into one data set, and the normal distribution with μ = 0, σ = 1 is used as the difference to randomly adjust the brightness and contrast of each channel. At the same time, Gaussian noise and dynamic blur are used to further improve the robustness of the algorithm. Three images are generated for each transformation. Divide the data into training set, validation set and test set.

[0011] Preferably, the improved RT-DETR neural network model in step S3 includes: Backbone network: The input image is extracted through FasterNet, and finally the feature maps of multiple scales are output to capture different levels of details in the image; Efficient hybrid encoder: It includes a decoder module AIFI that processes the last feature map and a CCFM module that uses a feature pyramid network to achieve cross-scale fusion of the last three feature maps; the CCFM module adds a self-developed small target detection head for processing, including upsampling using a bilinear difference algorithm, depthwise separable convolution, and merging feature maps through the RepC3 module; IoU-aware query selection: constrains the model to produce high classification scores for features with high IoU scores and low classification scores for features with low IoU scores during training; Decoder: receives the output of the encoder and IoU-aware query selection module, and uses self-attention and encoder-decoder attention mechanisms to predict the location and category of the object.

[0012] Preferably, the FasterNet replaces ordinary convolution with a combination of partial convolution and point-by-point convolution; in the partial convolution, ordinary convolution is applied only to part of the input channels to extract spatial features, and only the first or last continuous channel is used as a representative for calculating the entire feature map; in the point-by-point convolution, the size of the convolution kernel is 1x1, the convolution operation is only performed on a single pixel, and a linear transformation is performed on a single pixel of the input feature map; after the graph embedding module is processed, "—FasterNet Block——merging——FasterNet Block——merging——FasterNetBlock——merging——FasterNet Block——Global Pool——1*1Conv——FC—" is sequentially passed through, and finally feature maps of multiple scales are output to capture different levels of detail in the image.

[0013] Preferably, the upsampling using the bilinear difference algorithm in the small target detection head processing includes: finding the value of the unknown function f at point P = (x, y), and the values ​​of the known function f at four points Q11 = (x1, y1), Q12 = (x1, y2), Q21 = (x2, y1) and Q22 = (x2, y2); Step 1: Linear interpolation in the X direction, insert R2 into Q12, Q22, insert R1 into Q11, Q21;

[0014] Step 2: Linear interpolation in the Y direction. Point P is calculated by interpolating R1 and R2 calculated in the first step in the Y direction: .

[0015] Preferably, the depthwise separable convolution is to first process the image with depthwise convolution, and the depthwise convolution kernel is in single-channel mode, and convolve each channel of the input to obtain an output feature map with the same number of channels as the input feature map; then process the image with pointwise convolution; the merging of the feature maps and passing through the RepC3 module means merging the output feature maps after the bilinear interpolation algorithm and the depthwise separable convolution together and then passing through the residual module repeated three times.

[0016] Preferably, the defect detection of the rotor of the pumped storage unit in step S4 includes: Video frame diagnosis: diagnose the returned video frames through the algorithm service deployed by python3; Result storage: The diagnosis results are saved to the local mariadb database. When defects are identified, the frame images are saved to the image database before and after annotation. Performance evaluation: The precision P, recall R, average precision AP and mean average precision mAP are calculated based on the intersection of the predicted box and the true box.

[0017] Preferably, the calculation formula of the precision rate P is:

[0018] in, TP is the number of samples correctly predicted as positive; FP is the number of samples that are incorrectly predicted as positive; The calculation formula of recall rate R is: Among them, FN is the number of samples that are incorrectly predicted as negative classes; The mean average precision (AP) is the average value of the precision at different recall levels, which is obtained by calculating the precision and recall rate for each category based on the intersection-over-union ratio of the predicted box and the true box, and then interpolating the precision as a function of the recall rate and calculating the average value of the interpolated precision. The mean average precision (mAP) is the value obtained by averaging the APs of multiple categories. The calculation formula of the average precision AP is: Where p(r) is the precision when the recall rate is r; The calculation formula of the mean average precision mAP is: Where c is the total number of categories; AP i is the average precision of the ith category; min-AP is the minimum average precision, which is the minimum value of the average precision among all categories.

[0019] Due to the adoption of the above technical scheme, the technical progress achieved by the present invention is as follows.

[0020] The present invention can timely discover relatively easy-to-handle fault defects of the rotor of a pumped-storage generator set without removing the rotor, thereby not only reducing the number of rotor penetrations and reducing process operation risks, but also reducing component and seal wear and tear and lowering maintenance costs. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 is a flow chart of the present invention; Figure 2 This is a flow chart of model training and deployment of the present invention; Figure 3 mAP50-95 curve diagram for training of an embodiment of the present invention; Figure 4 mAP50 curve diagram for training of an embodiment of the present invention; Figure 5 A recall curve diagram for training of an embodiment of the present invention; Figure 6 This is a curve chart of the accuracy of training according to an embodiment of the present invention. DETAILED DESCRIPTION

[0022] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0023] A pumped storage unit insulation defect detection method based on robot combined with improved RT-DETR algorithm Figure 1 to Figure 2 As shown, the following steps are included: S1. Use the robot as the data acquisition terminal to collect image data in the chamber of the pumped storage unit without pumping the rotor as the original data set.

[0024] Specifically, the robot collects images inside the pumped-storage unit chamber through preset path planning as raw data. The robot can be a tracked motor-driven inspection robot equipped with an image acquisition device. The image acquisition device can be a camera with high resolution and strong adaptability to low-light environments.

[0025] S2. Mark the defect types and defect locations in the original data set, and use data enhancement technology to process the original data set and split the data into training set, validation set and test set.

[0026] Marking the defect types and defect locations in the original data set includes: using labelme software to perform supervised manual labeling on the original images, labeling the defect locations and defect types that need to be detected. The defect types are as follows (in actual applications, numbers are used to represent the types of defects for labeling and training): 1. Damaged insulation on the stator surface; 2. Discoloration and displacement of the stator core laminations; 3. Visible dust and white deposits on the surface of the slot wedges in the stator chamber; 4. There are traces of electric current damage on the rotor surface; 5. The rotor ventilation holes are blocked; 6. The rotor dynamic balancing bolts are loose.

[0027] The data augmentation techniques used to process the original data set include: Use the Albumentations library to perform image transformation operations, and perform geometric transformation, elastic transformation, and grid distortion operations on the tensor of the original image in sequence. Each transformation generates 3 images. The processed data are combined into one data set, and the normal distribution with μ = 0, σ = 1 is used as the difference to randomly adjust the brightness and contrast of each channel. At the same time, Gaussian noise and dynamic blur are used to further improve the robustness of the algorithm. Three images are generated for each transformation. Divide the data into training set, validation set and test set.

[0028] S3. Build an improved RT-DETR neural network model, train the neural network with the training set and validation set, evaluate the neural network model with the test set, perform multiple rounds of training, and select the best model that meets the industrial standard requirements as the defect detection model in multiple rounds of training and testing.

[0029] The improved RT-DETR neural network model includes: (1) Backbone network: The input image is passed through FasterNet (FasterNet CVPR2023) to replace the backbone PRestNet18 in the original RT-DETR for feature extraction, and finally outputs feature maps of multiple scales that capture different levels of detail in the image.

[0030] Specifically, FasterNet replaces ordinary convolution with a combination of partial convolution (PConv) and point-wise convolution (PWConv). In partial convolution, ordinary convolution is applied to only some of the input channels to extract spatial features, and the remaining channels remain unchanged; for continuous memory access, only the first or last continuous channel is used as the representative for calculating the entire feature map. In point-by-point convolution, the size of the convolution kernel is 1x1, and the convolution operation is only performed on a single pixel. A linear transformation is performed on a single pixel of the input feature map without changing the spatial size of the feature map; after the graph embedding module is processed, it passes through "—(FasterNet Block)——merging——(FasterNet Block)——merging——(FasterNet Block)——merging——(FasterNet Block)——(Global Pool)——(1*1Conv)——FC—" in sequence, where merging is merging; Global Pool is global pooling; 1*1Conv is 1×1 convolution; FC is file comparison, and finally outputs feature maps of multiple scales, which capture different levels of detail in the image.

[0031] (2) Efficient Hybrid Encoder: It mainly consists of two modules: 1. The decoder module AIFI that processes the last feature map; 2. The CCFM (CNN-based cross-scale feature-fusion module) module that uses the feature pyramid network to achieve cross-scale fusion of the last three feature maps.

[0032] Specifically, the CCFM module adds a self-developed small target detection head to the feature image output by the attention-based intra-scale feature interaction (AIFM) for processing, so as to adapt to the application scenario of pumped storage unit insulation defect detection. The small target detection head includes upsampling using a bilinear difference algorithm, depthwise separable convolution, and merging feature maps through the RepC3 module, as follows: ① Use the bilinear interpolation algorithm for upsampling. The algorithm is as follows: Find the value of the unknown function f at point P = (x, y). The values ​​of the known function f at four points Q11 = (x1, y1), Q12 = (x1, y2), Q21 = (x2, y1) and Q22 = (x2, y2).

[0033] Step 1: Linear interpolation in the X direction, insert R2 into Q12, Q22, insert R1 into Q11, Q21;

[0034] Step 2: Linear interpolation in the Y direction. Point P is calculated by interpolating R1 and R2 calculated in the first step in the Y direction: .

[0035] ② Depthwise separable convolution: First, use depthwise convolution to process the image. The depthwise convolution kernel is in single-channel mode. Convolution is performed on each channel of the input to obtain an output feature map with the same number of channels as the input feature map; then use point-by-point convolution to process the image.

[0036] ③ After the feature maps are merged, they are passed through the RepC3 module, which means that the output feature maps after the bilinear interpolation algorithm and the depth-wise separable convolution are merged (concat) together and then passed through the residual module repeated three times.

[0037] (3) IoU-aware Query Selection: Due to the inconsistent distribution of classification scores and location confidences, the model is constrained to produce high classification scores for features with high IoU scores (a metric for evaluating the accuracy of predicted boxes by object detection models) and low classification scores for features with low IoU scores during training. Therefore, the predicted boxes corresponding to the top K encoder features selected by the model based on the classification scores have both high classification scores and high IoU scores.

[0038] (4) Decoder: Receives the output of the encoder and IoU-aware query selection module, and uses self-attention and encoder-decoder attention mechanisms to predict the location and category of the object.

[0039] S4. Deploy the defect detection model to the visual system processing end, process the data sent back by the robot at the data acquisition end during operation, and perform defect detection on the rotor of the pumped storage unit.

[0040] Defect detection of pumped storage unit rotors includes: Video frame diagnosis: Diagnose the returned video frames through the algorithm service deployed by python3.

[0041] Specifically, when the robot crawls in the stator chamber, the camera is used to collect videos of the stator and rotor surfaces, and transmit the collected results to the computer in real time. At the same time, the collected video frames are diagnosed through the algorithm service deployed by python3.

[0042] Result storage: The diagnosis results are saved to the local mariadb database. When defects are identified, the frame images are saved to the image database before and after annotation. Performance evaluation: The precision P, recall R, average precision AP and mean average precision mAP are calculated based on the intersection of the predicted box and the true box.

[0043] Specifically, the precision rate P refers to the proportion of samples predicted by the model to be positive that are actually positive. The calculation formula is:

[0044] in, TP (True Positive) is the number of samples correctly predicted as positive; FP (False Positive) is the number of samples that are incorrectly predicted as positive; the precision reflects the accuracy of the model's prediction of the positive class.

[0045] The recall rate R refers to the proportion of samples that are actually positive that are correctly predicted as positive by the model. The calculation formula is: Among them, FN (False Negative) is the number of samples that are incorrectly predicted as negative (that is, they are actually positive but predicted as negative). The recall rate reflects the model's ability to identify positive samples.

[0046] The average precision (AP) is the average of the precision at different recall levels. It is obtained by calculating the integral of the precision within the range of recall, and reflects the overall performance of the model at different recall levels.

[0047] Specifically, the average precision AP can be calculated by the following steps: for each category, the precision and recall are calculated based on the intersection over union (IoU) of the predicted box and the true box; the precision is interpolated as a function of the recall rate to evaluate the precision at different recall levels; the average of these interpolated precision rates is calculated to obtain the AP of the category. The mean average precision mAP is the value obtained by averaging the AP of multiple categories, and is used in multi-category classification tasks to measure the overall performance of the model on all categories.

[0048] The calculation formula of average precision AP is:

[0049] Where p(r) is the precision when the recall rate is r; The calculation formula of the mean average precision mAP is: Where c is the total number of categories; AP iis the average precision of the ith category; min-AP is the minimum average precision, which is the minimum value of the average precision among all categories and is used to evaluate the performance of the model in the worst category.

[0050] The following experiments verify the effectiveness of the proposed defect detection method: Construction of training experiment platform: The CPU of data collection end is 6-core NVIDIA Carmel Arm ® v8.2 64-bit CPU 6MB L2 +4MB L3, GPU is 384-core NVlDlA VoltaTM architecture GPU with 48Tensor Cores, system environment is Deepstream6.1.1, Cuda11.4.14, Python3.8. The CPU of the data processing set is 12th Gen Intel Core i9-12900 Processor, GPU is NVIDIA RTX A4000, system environment is Windows1 1, Cuda12.0, Python3.1 1.3.

[0051] In the present invention, efficient model training is achieved through the designed training configuration file, and the specific parameters are set as follows: the total number of training rounds is 600, the early stopping threshold is 10, the batch size is 16, and the input image size is 640x640 pixels. The optimizer uses AdamW, the initial learning rate is 0.001, the cosine learning rate scheduling is adopted, the momentum is 0.937, and the weight decay is 0.0005. The data enhancement strategy includes hue change ±0.1, saturation change ±0.1, brightness change ±0.1, rotation ±10 degrees, translation ±0.1, and scaling range [0.8, 1.2].

[0052] Take the performance index of the model with the best performance in each training result as the result of this training, repeat the training five times, take the average value of each performance index and compare it with the performance of the model with the best performance in these five times, and take the highest value as the final result. The formula is as follows: (x is a certain performance index, the mth model performs best, and the performance index is x m , the final result is X) } Through the above algorithm, Figures 3 to 6 As shown in the figure, the final performance indicator parameters are: accuracy (precision) is 0.98222; recall (recall) is 0.79888; mean average precision (mAP): mAP50 is 0.92159, mAP50-95 is 0.52543.

[0053] A pumped storage unit insulation defect detection device based on a robot combined with an improved RT-DETR algorithm comprises a crawler motor-driven inspection robot equipped with a high-resolution image acquisition device with strong adaptability to low-light environments at a data acquisition end, and a computer at a computer vision processing end.

[0054] When the present invention is in use, it is possible to timely discover relatively easy-to-handle fault defects of the rotor of a pumped-storage generator set without removing the rotor, thereby not only reducing the number of times the rotor is pumped through and reducing process operation risks, but also reducing component and seal wear and tear and lowering maintenance costs.

Claims

1. A method for detecting insulation defects of pumped storage units based on a robot combined with an improved RT-DETR algorithm, characterized in that: The following steps are involved: S1. Use the robot as the data acquisition terminal to collect image data in the chamber of the pumped storage unit without pumping the rotor as the original data set; S2. Mark the defect types and defect locations in the original data set, process the original data set using data enhancement technology, and split the data into training set, validation set, and test set; S3. Build an improved RT-DETR neural network model, train the neural network with the training set and the validation set, evaluate the neural network model with the test set, perform multiple rounds of training, and select the best model that meets the industrial standard requirements as the defect detection model in multiple rounds of training and testing; S4. Deploy the defect detection model to the visual system processing end, process the data sent back by the robot at the data acquisition end during operation, and perform defect detection on the rotor of the pumped storage unit.

2. According to claim 1, a pumped storage unit insulation defect detection method based on robot combined with improved RT-DETR algorithm is characterized in that: In the step S1, a crawler motor-driven inspection robot equipped with an image acquisition device having high resolution and strong adaptability to low illumination environments collects images of the interior of the pumped-storage unit through a preset path planning as original data.

3. The method for detecting insulation defects of pumped storage units based on a robot combined with an improved RT-DETR algorithm according to claim 1, characterized in that: The step S2 of marking the defect types and defect locations in the original data set includes: using labelme software to perform supervised manual annotation on the original image, and the defect types include damaged insulation on the stator surface, discoloration and displacement of the stator core laminations, visible dust and white deposits on the surface of the slot wedges in the stator bore, traces of electric current damage on the rotor surface, blocked rotor ventilation holes, and loose rotor dynamic balancing bolts.

4. The method for detecting insulation defects of pumped storage units based on a robot combined with an improved RT-DETR algorithm according to claim 1, characterized in that: The processing of the original data set using the data enhancement technology in step S2 includes: Use the Albumentations library to perform image transformation operations, and perform geometric transformation, elastic transformation, and grid distortion operations on the tensor of the original image in sequence. Each transformation generates 3 images. The processed data are combined into one data set, and the normal distribution with μ = 0, σ = 1 is used as the difference to randomly adjust the brightness and contrast of each channel. At the same time, Gaussian noise and dynamic blur are used to further improve the robustness of the algorithm. Three images are generated for each transformation. Divide the data into training set, validation set and test set.

5. The method for detecting insulation defects of pumped storage units based on a robot combined with an improved RT-DETR algorithm according to claim 1, characterized in that: The improved RT-DETR neural network model in step S3 includes: Backbone network: The input image is extracted through FasterNet, and finally the feature maps of multiple scales are output to capture different levels of details in the image; Efficient hybrid encoder: It includes a decoder module AIFI that processes the last feature map and a CCFM module that uses a feature pyramid network to achieve cross-scale fusion of the last three feature maps; the CCFM module adds a self-developed small target detection head for processing, including upsampling using a bilinear difference algorithm, depthwise separable convolution, and merging feature maps through the RepC3 module; IoU-aware query selection: constrains the model to produce high classification scores for features with high IoU scores and low classification scores for features with low IoU scores during training; Decoder: receives the output of the encoder and IoU-aware query selection module, and uses self-attention and encoder-decoder attention mechanisms to predict the location and category of the object.

6. The method for detecting insulation defects of pumped storage units based on a robot combined with an improved RT-DETR algorithm according to claim 5, characterized in that: The FasterNet replaces ordinary convolution with a combination of partial convolution and point-by-point convolution; in the partial convolution, ordinary convolution is applied to only part of the input channels to extract spatial features, and only the first or last continuous channel is used as the representative for calculating the entire feature map; in the point-by-point convolution, the size of the convolution kernel is 1x1, the convolution operation is only performed on a single pixel, and a linear transformation is performed on a single pixel of the input feature map; after the graph embedding module is processed, it passes through "—FasterNet Block——merging——FasterNet Block——merging——FasterNet Block——merging——FasterNet Block——Global Pool——1*1Conv——FC—" in sequence, and finally outputs feature maps of multiple scales to capture different levels of detail in the image.

7. The method for detecting insulation defects of pumped storage units based on a robot combined with an improved RT-DETR algorithm according to claim 5, characterized in that: The upsampling using the bilinear difference algorithm in the small target detection head processing includes: finding the value of the unknown function f at the point P = (x, y), and the values ​​of the known function f at four points Q11 = (x1, y1), Q12 = (x1, y2), Q21 = (x2, y1) and Q22 = (x2, y2); Step 1: Linear interpolation in the X direction, insert R2 into Q12, Q22, insert R1 into Q11, Q21; Step 2: Linear interpolation in the Y direction. Point P is calculated by interpolating R1 and R2 calculated in the first step in the Y direction: 。 8. The method for detecting insulation defects of pumped storage units based on a robot combined with an improved RT-DETR algorithm according to claim 5, characterized in that: The depthwise separable convolution is to first process the image with depthwise convolution, and the depthwise convolution kernel is in single-channel mode, and convolve each channel of the input to obtain an output feature map with the same number of channels as the input feature map; then process the image with point-by-point convolution; the merging of the feature maps and passing through the RepC3 module means merging the output feature maps after the bilinear interpolation algorithm and the depthwise separable convolution together and then passing through the residual module repeated three times.

9. The method for detecting insulation defects of pumped storage units based on a robot combined with an improved RT-DETR algorithm according to claim 5, characterized in that: The defect detection of the rotor of the pumped storage unit in step S4 includes: Video frame diagnosis: diagnose the returned video frames through the algorithm service deployed by python3; Result storage: The diagnosis results are saved to the local mariadb database. When defects are identified, the frame images are saved to the image database before and after annotation. Performance evaluation: The precision P, recall R, average precision AP and mean average precision mAP are calculated based on the intersection of the predicted box and the true box.

10. The method for detecting insulation defects of pumped storage units based on a robot combined with an improved RT-DETR algorithm according to claim 9, characterized in that: The calculation formula of the precision rate P is: in, TP is the number of samples correctly predicted as positive; FP is the number of samples that are incorrectly predicted as positive; The calculation formula of recall rate R is: Among them, FN is the number of samples that are incorrectly predicted as negative classes; The mean average precision (AP) is the average value of the precision at different recall levels, which is obtained by calculating the precision and recall rate for each category based on the intersection-over-union ratio of the predicted box and the true box, and then interpolating the precision as a function of the recall rate and calculating the average value of the interpolated precision. The mean average precision (mAP) is the value obtained by averaging the APs of multiple categories. The calculation formula of the average precision AP is: Where p(r) is the precision when the recall rate is r; The calculation formula of the mean average precision mAP is: Where c is the total number of categories; AP i is the average precision of the ith category; min-AP is the minimum average precision, which is the minimum value of the average precision among all categories.

Images