A deep learning defect detection system applied to power distribution network completion acceptance

Abstract

The application provides a deep learning defect detection system applied to power distribution network completion acceptance, which comprises a model training module, a synchronous transmission module, an edge detection module and a display report module, the model training module is used for training large-scale models and lightweight models to improve the detection effect of the lightweight models, the synchronous transmission module is used for transmitting lightweight model data to the edge detection module and transmitting the detection result of the edge detection module to the display report module, the edge detection module detects power equipment in real time based on an edge box, and the display report module is used for displaying the detection result of the edge detection module; the system can realize defect identification of power distribution network equipment with a wide range of categories at the edge, quickly feed back the detection result on site, and facilitate defect elimination in time.

Description

Technical Field

[0001] This invention relates to the field of electrical digital data processing, and specifically to a deep learning defect detection system applied to the final acceptance of power distribution networks. Background Technology

[0002] Current research on power distribution network applications largely focuses on algorithm models with over ten million parameters and models deployed on servers with ample computing power, with limited research and improvements on lightweight edge-side implementations. Among the few studies and patents on lightweight target detection models for power distribution network edge applications, the coverage of specific defect types is limited. For example, a single model may only detect insulators and related defects, failing to cover defects related to towers and cable ties. Alternatively, the defect categories detected by the algorithm may be unclear and unrealistic, resulting in detection results that are significant for research but have limited practical application value. Existing deep learning detection algorithms for broader power distribution network scenarios have large parameter counts and high computational costs, leading to long processing times on edge devices with limited computing power, thus failing to demonstrate the advantage of edge devices in reducing cloud communication load.

[0003] The foregoing description of the background art is intended only to facilitate understanding of the invention. This description does not endorse or acknowledge any common general knowledge in the materials mentioned.

[0004] Many power distribution network monitoring systems have been developed. Extensive research and reference have revealed existing systems such as the one disclosed in publication number CN108132418B. These systems generally include a monitoring device, a server, and user terminals. These components are connected via the internet to form a networked monitoring system for power equipment. The monitoring device samples electrical signals from the main or branch switches of the power network. The server uses a predetermined algorithm to decompose the electrical signals into signal components with different characteristics, matching these components with the characteristics of preset reference signal components for each power device in the power network to determine the corresponding signal component. Based on these signal components, the operating status of the corresponding power device in the power network is determined. However, this system detects equipment by detecting electrical signals and cannot detect external anomalies unrelated to the electrical signals. Summary of the Invention

[0005] The purpose of this invention is to address the existing shortcomings by proposing a deep learning-based defect detection system for power distribution network completion acceptance.

[0006] The present invention adopts the following technical solution:

[0007] A deep learning-based defect detection system for power distribution network completion acceptance includes a model training module, a synchronous transmission module, an edge detection module, and a report display module.

[0008] The model training module is used to train large-scale models and lightweight models to improve the detection effect of lightweight models. The synchronous transmission module is used to transmit lightweight model data to the edge detection module and transmit the detection results of the edge detection module to the display report module. The edge detection module performs real-time detection of power equipment based on edge boxes. The display report module is used to display the detection results of the edge detection module.

[0009] The model training module includes a training data management unit, a large-scale model training unit, a lightweight model training unit, a distillation and quantization unit, and a model packaging unit. The training data management unit stores training data and sends it to the large-scale model training unit and the lightweight model training unit. The large-scale model training unit trains the training data based on a model with tens of millions of parameters, and the lightweight model training unit trains the training data based on a model with fewer than tens of millions of parameters. The distillation and quantization unit distills the output of the large-scale model training unit and quantizes the lightweight model training unit based on the distillation results. The model packaging unit packages the quantized lightweight model data.

[0010] The edge detection module includes an image capture unit, a detection storage unit, a detection management unit, a detection execution unit, and a model synchronization unit. The image capture unit is used to capture video stream information of the power distribution network scenario. The detection storage unit is used to save video stream information and detection result information. The detection management unit is used to obtain frame data from the video stream and send it to the detection execution unit. The detection execution unit is used to run a lightweight model to detect the frame data. The model synchronization unit is used to receive packaged data and synchronize the lightweight model in the detection execution unit.

[0011] Furthermore, the distillation quantization unit includes a distillation processor, a number matching processor, an optimization processor, and a quantization processor. The distillation processor performs distillation processing on the output of the large-scale model training unit. The number matching processor matches the distillation results with the output of the lightweight model training unit according to the numbering information. The optimization processor calculates the loss function of the lightweight model training unit based on the matching results and optimizes the parameters in the lightweight model training unit. The quantization processor performs asymmetric quantization processing on the lightweight model.

[0012] Furthermore, the distillation processor performs distillation processing according to the following formula to obtain the tag information:

[0013]

[0014] Where z(i) represents the i-th output of the large-scale model training unit, q(i) represents the i-th label information, T is the distillation temperature, exp() represents the exponential function with the natural constant as the base, and m is the number of outputs of the large-scale model training unit in a training batch.

[0015] Furthermore, the detection management unit includes a time consumption monitoring processor, a data throughput performance monitoring processor, and a frame information control processor. The time consumption monitoring processor is used to monitor the time consumed by the detection execution unit from receiving a frame information to outputting a detection result. The data throughput performance monitoring processor is used to monitor the speed at which the edge box outputs data. The frame information control processor is used to control the process of obtaining frame information from the video stream register.

[0016] Furthermore, the frame information control processor performs curve fitting based on discrete points on Time(t) and Tout(t) to obtain complete Time(t) and Tout(t) curves, and calculates the balance index P(t) according to the following formula:

[0017]

[0018] The frame information control processor is configured with a default detection time point t. i and detection radius r i The frame information control processor needs to [t] i -r i , t i +r i Within a given time period, a frame of information is selected for detection. The frame information control processor calculates the screening index Se(t) according to the following formula:

[0019] Se(t) = e P(t) ·(|tt i |+1);

[0020] The frame information control processor in each [t] i -r i , t i +r i Find the time point t' with the minimum filtering index within the time period, and obtain the corresponding frame information to send to the detection execution unit.

[0021] The beneficial effects achieved by this invention are:

[0022] This system targets the defect identification scenario of power distribution equipment on edge devices with limited computing power. First, it uses a pre-trained model with medium parameters to improve the detection capability of the model with small parameters through teacher-student knowledge distillation. Then, it uses the KL divergence calibration method to quantize and calibrate the model with small parameters, further accelerating its computing speed at the cost of slight accuracy loss. For image data traffic of different tasks, it maximizes the use of edge device data throughput performance based on the edge device's data throughput performance, and balances the time consumption of deep learning model operation and data throughput in parallel.

[0023] To further understand the features and technical content of the present invention, please refer to the following detailed description and drawings of the present invention. However, the drawings provided are for reference and illustration only and are not intended to limit the present invention. Attached Figure Description

[0024] Figure 1 This is a schematic diagram of the overall structural framework of the present invention;

[0025] Figure 2 This is a schematic diagram of the model training module of the present invention;

[0026] Figure 3 This is a schematic diagram of the edge detection module of the present invention;

[0027] Figure 4 This is a schematic diagram of the distillation quantification unit of the present invention;

[0028] Figure 5 This is a schematic diagram of the detection management unit of the present invention. Detailed Implementation

[0029] The following specific embodiments illustrate the implementation of the present invention. Those skilled in the art can understand the advantages and effects of the present invention from the content disclosed in this specification. The present invention can be implemented or applied through other different specific embodiments, and various details in this specification can also be modified and changed based on different viewpoints and applications without departing from the spirit of the present invention. Furthermore, the accompanying drawings of the present invention are for simple illustrative purposes only and are not depictions of actual dimensions; this is stated beforehand. The following embodiments will further describe the relevant technical content of the present invention in detail, but the disclosed content is not intended to limit the scope of protection of the present invention.

[0030] Example 1.

[0031] This embodiment provides a deep learning-based defect detection system for power distribution network completion acceptance, combined with... Figure 1 It includes a model training module, a synchronous transmission module, an edge detection module, and a display report module;

[0032] The model training module is used to train large-scale models and lightweight models to improve the detection effect of lightweight models. The synchronous transmission module is used to transmit lightweight model data to the edge detection module and transmit the detection results of the edge detection module to the display report module. The edge detection module performs real-time detection of power equipment based on edge boxes. The display report module is used to display the detection results of the edge detection module.

[0033] The model training module includes a training data management unit, a large-scale model training unit, a lightweight model training unit, a distillation and quantization unit, and a model packaging unit. The training data management unit stores training data and sends it to the large-scale model training unit and the lightweight model training unit. The large-scale model training unit trains the training data based on a model with tens of millions of parameters, and the lightweight model training unit trains the training data based on a model with fewer than tens of millions of parameters. The distillation and quantization unit distills the output of the large-scale model training unit and quantizes the lightweight model training unit based on the distillation results. The model packaging unit packages the quantized lightweight model data.

[0034] The edge detection module includes an image capture unit, a detection storage unit, a detection management unit, a detection execution unit, and a model synchronization unit. The image capture unit is used to capture video stream information of the power distribution network scenario. The detection storage unit is used to save video stream information and detection result information. The detection management unit is used to obtain frame data from the video stream and send it to the detection execution unit. The detection execution unit is used to run a lightweight model to detect the frame data. The model synchronization unit is used to receive packaged data and synchronize the lightweight model in the detection execution unit.

[0035] Furthermore, the distillation quantization unit includes a distillation processor, a number matching processor, an optimization processor, and a quantization processor. The distillation processor performs distillation processing on the output of the large-scale model training unit. The number matching processor matches the distillation results with the output of the lightweight model training unit according to the numbering information. The optimization processor calculates the loss function of the lightweight model training unit based on the matching results and optimizes the parameters in the lightweight model training unit. The quantization processor performs asymmetric quantization processing on the lightweight model.

[0036] Furthermore, the distillation processor performs distillation processing according to the following formula to obtain the tag information:

[0037]

[0038] Where z(i) represents the i-th output of the large-scale model training unit, q(i) represents the i-th label information, T is the distillation temperature, exp() represents the exponential function with the natural constant as the base, and m is the number of outputs of the large-scale model training unit in a training batch.

[0039] Furthermore, the detection management unit includes a time consumption monitoring processor, a data throughput performance monitoring processor, and a frame information control processor. The time consumption monitoring processor is used to monitor the time consumed by the detection execution unit from receiving a frame information to outputting a detection result. The data throughput performance monitoring processor is used to monitor the speed at which the edge box outputs data. The frame information control processor is used to control the process of obtaining frame information from the video stream register.

[0040] Furthermore, the frame information control processor performs curve fitting based on discrete points on Time(t) and Tout(t) to obtain complete Time(t) and Tout(t) curves, and calculates the balance index P(t) according to the following formula:

[0041]

[0042] The frame information control processor is configured with a default detection time point t. i and detection radius r i The frame information control processor needs to [t] i -r i , t i +r i Within a given time period, a frame of information is selected for detection. The frame information control processor calculates the screening index Se(t) according to the following formula:

[0043] Se(t) = e P(t) ·(|tt i |+1);

[0044] The frame information control processor in each [t] i -r i , t i +r i Find the time point t' with the minimum filtering index within the time period, and obtain the corresponding frame information to send to the detection execution unit.

[0045] Example 2.

[0046] This embodiment includes all the contents of Embodiment 1, and provides a deep learning defect detection system for power distribution network completion acceptance, including a model training module, a synchronous transmission module, an edge detection module and a display report module;

[0047] The model training module is used to train large-scale models and lightweight models to improve the detection effect of lightweight models. The synchronous transmission module is used to transmit lightweight model data to the edge detection module and transmit the detection results of the edge detection module to the display report module. The edge detection module performs real-time detection of power equipment based on edge boxes. The display report module is used to display the detection results of the edge detection module.

[0048] Combination Figure 2 The model training module includes a training data management unit, a large-scale model training unit, a lightweight model training unit, a distillation and quantization unit, and a model packaging unit. The training data management unit stores training data and sends it to the large-scale model training unit and the lightweight model training unit. The large-scale model training unit trains the training data based on a model with tens of millions of parameters, and the lightweight model training unit trains the training data based on a model with fewer than tens of millions of parameters. The distillation and quantization unit distills the output of the large-scale model training unit and quantizes the lightweight model training unit based on the distillation results. The model packaging unit packages the quantized lightweight model data.

[0049] The synchronous transmission module includes a model transmission unit and a detection transmission unit. The model transmission unit is used to send the packaged data in the model packaging unit to the edge detection module, and the detection transmission unit is used to transmit the detection results of the edge detection module to the display report module.

[0050] Combination Figure 3 The edge detection module includes an image capturing unit, a detection storage unit, a detection management unit, a detection execution unit, and a model synchronization unit. The image capturing unit is used to capture video stream information of the power distribution network scenario. The detection storage unit is used to save video stream information and detection result information. The detection management unit is used to obtain frame data from the video stream and send it to the detection execution unit. The detection execution unit is used to run a lightweight model to detect the frame data. The model synchronization unit is used to receive the packaged data sent by the model transmission unit and synchronize the lightweight model in the detection execution unit.

[0051] The image capture unit, detection storage unit, detection management unit, detection execution unit, and model synchronization unit are integrated into an edge box, and each edge box is responsible for monitoring a power distribution network scenario;

[0052] The report display module includes a storage backup unit, a report display unit, and a scene retrieval unit. The storage backup unit is used to receive and store the detection results sent by each edge box. The report display unit is used to filter out abnormal reports from the detection results and display them. The scene retrieval unit is used to retrieve all report information for the corresponding scene.

[0053] The training data management unit includes a training data register and a data output processor. The training data register is used to store training data. The data output processor is used to retrieve a copy of training data from the training data register, number it, copy the numbered training data, and send it to the large-scale model training unit and the lightweight model training unit respectively.

[0054] Combination Figure 4 The distillation quantization unit includes a distillation processor, a number matching processor, an optimization processor, and a quantization processor. The distillation processor performs distillation processing on the output of the large-scale model training unit. The number matching processor matches the distillation results with the output of the lightweight model training unit according to the numbering information. The optimization processor calculates the loss function of the lightweight model training unit based on the matching results and optimizes the parameters in the lightweight model training unit. The quantization processor performs asymmetric quantization processing on the lightweight model.

[0055] The distillation processor performs distillation processing according to the following formula to obtain the label information:

[0056]

[0057] Where z(i) represents the i-th output of the large-scale model training unit, q(i) represents the i-th label information, T is the distillation temperature, exp() represents the exponential function with the natural constant as the base, and m is the number of outputs of the large-scale model training unit in a training batch.

[0058] In this embodiment, the distillation temperature T is set to 3;

[0059] The training data is assigned a corresponding output result and label information. The number matching processor sets the label information with the same number and the output result of the lightweight model training unit as a pair of matching data.

[0060] The optimization processor optimizes the parameters in the lightweight model training unit by including the following steps:

[0061] S1. Calculate the loss value based on the loss function;

[0062] S2. Calculate the gradient of each parameter by backpropagating based on the loss value;

[0063] S3. Update parameters based on gradient;

[0064] In step S1, the loss value is calculated according to the following formula:

[0065]

[0066] Where s(i) is the data result of the lightweight model training unit matched with q(i);

[0067] The gradient of the parameters is the partial derivative of the loss value with respect to the parameters. The parameter update is obtained by the following formula:

[0068]

[0069] Where w is the original parameter value, w' is the new parameter value, and α is the learning rate;

[0070] The quantization processor quantizes the parameters according to the following formula:

[0071]

[0072] Among them, V int V is the quantized integer parameter. float These are floating-point parameters before quantization, ZeroPoint is the zero point, Sc is the quantization scaling factor, and round() is the rounding function;

[0073] The values ​​of the zero point and quantization scaling factor were obtained by calibration using the KL divergence calibration method;

[0074] The image capturing unit includes a webcam and a rotation control processor. The webcam is used to capture video stream information, and the rotation control processor is used to control the webcam to rotate and capture images of the entire power distribution network scene.

[0075] The detection storage unit includes a video stream register and a detection result register. The video stream register is used to store the captured video stream information, and the detection result register is used to store the detection result information.

[0076] Combination Figure 5 The detection management unit includes a time consumption monitoring processor, a data throughput performance monitoring processor, and a frame information control processor. The time consumption monitoring processor is used to monitor the time consumed by the detection execution unit from receiving a frame information to outputting a detection result. The data throughput performance monitoring processor is used to monitor the speed at which the edge box outputs data. The frame information control processor is used to control the process of obtaining frame information from the video stream register.

[0077] The time taken for the network camera to rotate for one cycle is denoted as T, and any point in time within that cycle is denoted as t. The time detected by the time consumption monitoring processor is denoted as Time(t), which represents the time consumed by the detection execution unit in processing the frame information at time t. The data throughput performance monitoring processor calculates the time Tout(t) required to output the frame information at time t and the corresponding detection result according to the following formula:

[0078]

[0079] Where Dv(t) is the amount of frame information and corresponding detection results at time t, and v is the data output speed;

[0080] Dv(t) changes with the captured image, while v remains constant and is related to the edge box hardware performance.

[0081] The frame information control processor performs curve fitting based on discrete points on Time(t) and Tout(t) to obtain complete Time(t) and Tout(t) curves, and calculates the balance index P(t) according to the following formula:

[0082]

[0083] The frame information control processor is configured with a default detection time point t. i and detection radius r i The frame information control processor needs to [t] i -r i , t i +r i Within a given time period, a frame of information is selected for detection. The frame information control processor calculates the screening index Se(t) according to the following formula:

[0084] Se(t) = e P(t) ·(|tt i |+1);

[0085] The frame information control processor in each [t] i -r i , t i +r i Find the time point t' with the minimum filtering index within the time period, and obtain the corresponding frame information to send to the detection execution unit;

[0086] The detection execution unit includes a model loading processor and an input / output processor. The model loading processor is used to load a lightweight model, and the input / output processor is used to receive frame information and output model results.

[0087] In the above text, i and j are both ordinal numbers representing sequence numbers.

[0088] The content disclosed above is only a preferred and feasible embodiment of the present invention, and is not intended to limit the scope of protection of the present invention. Therefore, all equivalent technical changes made based on the content of the present invention specification and drawings are included within the scope of protection of the present invention. Furthermore, the elements therein can be updated as technology develops.

Claims

1. A deep learning-based defect detection system for power distribution network completion acceptance, characterized in that, It includes a model training module, a synchronous transmission module, an edge detection module, and a display report module; The model training module is used to train large-scale models and lightweight models to improve the detection effect of lightweight models. The synchronous transmission module is used to transmit lightweight model data to the edge detection module and transmit the detection results of the edge detection module to the display report module. The edge detection module performs real-time detection of power equipment based on edge boxes. The display report module is used to display the detection results of the edge detection module. The model training module includes a training data management unit, a large-scale model training unit, a lightweight model training unit, a distillation and quantization unit, and a model packaging unit. The edge detection module includes an image capture unit, a detection storage unit, a detection management unit, a detection execution unit, and a model synchronization unit; The detection management unit includes a time consumption monitoring processor, a data throughput performance monitoring processor, and a frame information control processor. The time consumption monitoring processor is used to monitor the time consumed by the detection execution unit from receiving a frame information to outputting a detection result. The data throughput performance monitoring processor is used to monitor the speed at which the edge box outputs data. The frame information control processor is used to control the process of obtaining frame information from the video stream register. The time taken for the network camera to rotate for one cycle is denoted as T, and any point in time within that cycle is denoted as t. The time detected by the time consumption monitoring processor is denoted as Time(t), which represents the time consumed by the detection execution unit in processing the frame information at time point t. The data throughput performance monitoring processor calculates the time Tout(t) required to output the frame information at time point t and the corresponding detection result according to the following formula: ； Where Dv(t) is the amount of frame information and corresponding detection results at time t, and v is the data output speed; Dv(t) changes with the captured image, while v remains constant and is related to the edge box hardware performance. The frame information control processor performs curve fitting based on discrete points on Time(t) and Tout(t) to obtain complete Time(t) and Tout(t) curves, and calculates the balance index P(t) according to the following formula: ； The frame information control processor is configured with a default detection time point t. i and detection radius r i The frame information control processor needs to [t] i -r i , t i +r i Within a given time period, a frame of information is selected for detection. The frame information control processor calculates the screening index Se(t) according to the following formula: ； The frame information control processor finds the time point with the minimum filtering index within each time period [ti-ri, ti+ri] and obtains the corresponding frame information to send to the detection execution unit; The detection execution unit includes a model loading processor and an input / output processor. The model loading processor is used to load a lightweight model, and the input / output processor is used to receive frame information and output model results.

2. The deep learning defect detection system for power distribution network completion acceptance as described in claim 1, characterized in that, The distillation and quantization unit includes a distillation processor, a number matching processor, an optimization processor, and a quantization processor. The distillation processor performs distillation processing on the output of the large-scale model training unit. The number matching processor matches the distillation results with the output of the lightweight model training unit according to the numbering information. The optimization processor calculates the loss function of the lightweight model training unit based on the matching results and optimizes the parameters in the lightweight model training unit. The quantization processor performs asymmetric quantization processing on the lightweight model.

3. The deep learning defect detection system for power distribution network completion acceptance as described in claim 2, characterized in that, The distillation processor performs distillation processing according to the following formula to obtain the tag information: ； Where z(i) represents the i-th output of the large-scale model training unit, q(i) represents the i-th label information, T is the distillation temperature, exp() represents the exponential function with the natural constant as the base, m is the number of outputs of the large-scale model training unit in a training batch, and i and j are ordinal numbers used to represent the sequence number.

Images