Transformer substation edge image recognition and intelligent early warning method based on distillation learning

Abstract

The invention discloses a transformer substation edge image recognition and intelligent early warning method based on distillation learning, and relates to the technical field of intelligent operation and maintenance of power systems. Aiming at the problems of high cloud processing delay, insufficient edge computing power and high false alarm rate of a single data source in the prior art, the method comprises the following steps: arranging visible light and infrared sensing equipment in a transformer substation, and collecting data and transmitting the data to an edge analysis server; the cloud end utilizes a knowledge distillation technology to combine with feature map distillation and parameter pruning to generate a lightweight student model of which the parameter quantity is compressed to be less than 1 / 10 of the original quantity and FLOPs is less than or equal to 1GFLOPs; after Gaussian denoising and ROI extraction are carried out on the image by the edge side, abnormal reasoning is carried out through sliding window time sequence analysis, sound-light alarm is triggered, and through dual verification of electrical quantity and image weighted score (w1 > = w2), synchronous uploading to a master station is carried out. According to the method, edge real-time reasoning (delay is smaller than or equal to 500 ms) is achieved, the false alarm rate is reduced to 3.2% or below, the model supports differential online updating, and the intelligent early warning requirement of the transformer substation is met.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of power system intelligent operation and maintenance, and particularly relates to a substation edge image recognition and intelligent early warning method based on distillation learning. BACKGROUND

[0002] With the development of intelligent power system, substation equipment state monitoring has higher requirements for real-time and accuracy; traditional operation and maintenance relies on manual inspection and fixed threshold monitoring, which has problems such as low efficiency (single station full inspection time ≥ 8 hours), poor real-time (abnormality discovery delay ≥ 30 minutes), high false alarm rate (false alarm rate under environmental interference is more than 20%) and the like; in recent years, intelligent monitoring schemes based on deep learning realize defect identification through visible light / infrared perception devices and cloud AI models, which have improved operation and maintenance efficiency by 40%, but face four technical bottlenecks in engineering application:

[0003] Firstly, in the cloud centralized processing mode, 4K video (code rate 20 Mbps) full upload leads to transmission delay ≥ 1.2 seconds, and the total delay is more than 1.5 seconds after adding the cloud reasoning time, which cannot meet the 500 ms response requirement of emergency faults such as main transformer overheating, and the bandwidth occupation exceeds the bearing capacity of the 30M private line of the substation;

[0004] Secondly, the mainstream deep learning model (such as YOLOv5s) has a parameter amount of 7MB and FLOPs of 12GFLOPs, while the edge server has limited computing power, and the reasoning frame rate is only 8fps (demand ≥ 15fps) when directly deployed, and the missing detection rate is as high as 12%;

[0005] Thirdly, a single data source (infrared or visible light) is significantly affected by the environment: the misjudgment rate of infrared temperature measurement under sunlight interference is 18%, the accuracy of visible light images at night is reduced to below 70%, and the early defect false negative rate is more than 25% when not fusing electrical quantity data, therefore, multi-source data fusion technology must be used to realize cross-modal collaborative analysis;

[0006] Fourthly, the traditional edge model uses offline update, the complete model package is ≥ 50MB, the update time is ≥ 10 minutes, and the fixed threshold early warning mechanism cannot adapt to the dynamic changes of load rate, environmental temperature and the like, and the seasonal false and missing alarm rate fluctuation is more than 30%;

[0007] The essential defect of the prior art is that the collaboration of "lightweight model edge deployment + multi-source data fusion + dynamic threshold self-adaption" is not realized, resulting in insufficient real-time, accuracy and environmental adaptability; the present application fundamentally solves the above problems and meets the intelligent operation and maintenance requirements through innovations such as knowledge distillation compression model, multi-source data double verification and sliding window time series analysis.

[0008] Therefore, the present application is proposed. SUMMARY

[0009] The technical problem to be solved by the present application is to overcome the shortcomings of the prior art and provide a substation edge image recognition and intelligent early warning method based on distillation learning, which solves the problems raised in the above background art.

[0010] To solve the above technical problems, the basic idea of the technical solution adopted by the present application is:

[0011] A substation edge image recognition and intelligent early warning method based on distillation learning, comprising the following steps:

[0012] Visible light and infrared sensing devices are arranged in the substation, device image and temperature data are collected and transmitted to the edge analysis server; wherein, when the sensing devices are arranged, the main transformer area of the 500kV substation is configured with a binocular temperature measuring pan-tilt and a visible light ball machine, and the indoor screen cabinet area is monitored without dead angle through the pressing plate and indicator light realized by the lifting rod;

[0013] The cloud server generates a lightweight student model through knowledge distillation technology, and the distillation process includes: softening the teacher model output probability distribution based on the temperature parameter T, combining feature map distillation and parameter pruning to compress the model parameter quantity to less than 1 / 10 of the original model; wherein, the temperature parameter T of knowledge distillation is in the range of 5-15, and the floating point operation quantity (FLOPs) of the model after parameter pruning is ≤1GFLOPs; the student model supports online update: after the master station collects misidentified samples, it pushes the model update package through differential upgrade, and the update process does not affect real-time inference; the formula for softening the teacher model output probability distribution based on the temperature parameter T is: Wherein, q i is the softened prediction probability of the edge side student model for the i-th type of fan operating condition / fault type, z i is the original output score of the cloud teacher model for the i-th type of condition / fault, T is the temperature parameter of knowledge distillation, and N is the total number of condition / fault types;

[0014] The edge analysis server uses the student model to perform real-time inference on the preprocessed image, and judges the anomaly by combining sliding window time series analysis; wherein, the image preprocessing includes Gaussian denoising and region of interest (ROI) extraction, and the region of interest (ROI) is the key components of the device (opening and closing mechanism, meter, joint); in the sliding window time series analysis, the smoothing coefficient a is in the range of 0.2-0.5, and the window length is 5-10 time stamps; the edge analysis server and the master station communicate through a 30MMSTP private line, the data encryption uses the SM4 algorithm, and the transmission delay is ≤500ms; the formula for combining sliding window time series analysis is S t =αX t +(1-α)S t-1 , wherein S t is the residual value smoothed by the sliding window at the current time (t), a is the smoothing coefficient, and Xt S is the original residual error for the current time (t) t -1 is the residual error value of the previous time (t-1) smoothed by a sliding window;

[0015] An audible and visual alarm is triggered during an anomaly, a multi-source data fusion mechanism is used to jointly analyze the electrical quantity data and the image recognition result after spatio-temporal alignment, and the electrical quantity data and the image recognition result are double-verified through weighted scoring and uploaded to the main station simultaneously; in the double-verification, the electrical quantity score P e The image score P i Based on the defect detection confidence, the weights w1+w2=1 and w1≥w2; during multi-source data fusion, the infrared temperature measurement data and the image data are synchronized through a GPS clock, and the time error is ≤10 ms; wherein the weighted scoring formula of the electrical quantity data and the image recognition result is: W=w1P e +w2P i wherein, W is the comprehensive fault score (value range 0-1), w1 and w2 are weight coefficients, P e is the fault score of the electrical quantity data, P i is the fault score of the image recognition.

[0016] After adopting the above technical scheme, the present application has the following beneficial effects compared with the prior art, of course, any product implementing the present application does not necessarily need to achieve all the advantages described below:

[0017] The present application generates a lightweight student model through knowledge distillation, and the edge side inference time is reduced from 1.5 seconds of the cloud solution to ≤450 ms, combined with the 30MMSTP special line transmission delay ≤500 ms, the total response time meets the 500 ms early warning requirement of the main transformer overheating and other emergency faults;

[0018] The lightweight technology combining feature map distillation and parameter pruning compresses the model parameter quantity from 7 MB of the traditional YOLOv5s to ≤5 MB, reduces the floating point operation quantity to below 0.8 GFLOPs, and adapts to the edge server computing resource; after deployment, the inference frame rate is increased to 18 fps, the CPU utilization is stably below 30%, 8 channels of 4K video streams can be processed simultaneously, and the missing detection rate is reduced from 12% to below 0.5% compared with the existing edge scheme;

[0019] The present application combines the double verification mechanism of infrared temperature measurement, visible light image and electrical quantity data, reduces the false alarm rate of a single data source from 18% / 25% to 3.2%, and reduces the false negative rate from 25% to below 0.5%;

[0020] The sliding window timing analysis combines Kalman filtering to dynamically adjust the early warning threshold, eliminates interference such as sunshine and load fluctuation, and reduces the false alarm rate in summer from 30% to below 5%, and reduces the low-temperature missed detection rate in winter from 15% to below 2%; the model difference upgrading mechanism realizes real-time optimization of the edge model, and the new type of defect recognition accuracy improves at a speed 20 times faster than the traditional offline update, ensuring long-term stable operation of the system;

[0021] The GPS clock synchronization mechanism ensures the spatiotemporal consistency of infrared and image data, avoids misjudgment caused by clock deviation, the SM4 encryption communication is combined with the 30M MSTP private line, the transmission delay is less than or equal to 500ms, and the data security meets the requirements of Q / GDW11337-2014, and meets the substation safety protection specification.

[0022] The specific embodiments of the application will be further described in detail below with reference to the accompanying drawings. BRIEF DESCRIPTION OF DRAWINGS

[0023] The drawings in the following description are only some embodiments, and other drawings can be obtained by those skilled in the art without creative labor on the basis of the drawings. In the drawings:

[0024] Figure 1 The flow chart of the substation edge image recognition and intelligent early warning method based on distillation learning.

[0025] It should be noted that these drawings and written descriptions are not intended to limit the scope of the concept of the application in any way, but to illustrate the concept of the application to those skilled in the art by referring to specific embodiments. DETAILED DESCRIPTION

[0026] The application will now be further described in detail with reference to the accompanying drawings.

[0027] Please refer to Figure 1 In the embodiment, a substation edge image recognition and intelligent early warning method based on distillation learning is provided, which includes the following steps:

[0028] Visible light and infrared sensing devices are arranged in the substation, device images and temperature data are collected and transmitted to the edge analysis server; wherein, when the sensing devices are arranged, the main transformer area of the 500kV substation is configured with a binocular temperature measuring cloud platform and a visible light ball machine, and the indoor screen cabinet area is monitored by the lifting rod to realize the non-blind plate and indicator light monitoring;

[0029] The cloud server generates a lightweight student model through a knowledge distillation technique, and the distillation process includes: softening the teacher model output probability distribution based on a temperature parameter T, combining feature map distillation and parameter pruning to compress the model parameter quantity to less than 1 / 10 of the original model; wherein the temperature parameter T of the knowledge distillation is in the range of 5-15, and the floating point operation quantity (FLOPs) of the model after parameter pruning is ≤1 GFLOPs; the student model supports online updating: after the master station collects misidentified samples, it pushes a model update package through differential upgrading, and the updating process does not affect real-time inference; the formula for softening the teacher model output probability distribution based on the temperature parameter T is: wherein q i is the softened prediction probability of the edge side student model for the i-th type of fan operating condition / fault type, z i is the original output score of the cloud teacher model for the i-th type of condition / fault, T is the temperature parameter of knowledge distillation, and N is the total number of condition / fault categories; the cloud server uses ResNet50 as the teacher model and pre-trains it on a dataset containing 100,000 substation equipment images;

[0030] The edge analysis server uses the student model to perform real-time inference on the image after preprocessing, and judges the anomaly by combining sliding window time series analysis; wherein the image preprocessing includes Gaussian denoising and region of interest (ROI) extraction, and the region of interest (ROI) is the key component of the equipment (opening and closing mechanism, meter, joint); in the sliding window time series analysis, the smoothing coefficient a is in the range of 0.2-0.5, and the window length is 5-10 time stamps; the edge analysis server communicates with the master station through a 30MMSTP private line, the data encryption uses the SM4 algorithm, and the transmission delay is ≤500ms; the formula for combining sliding window time series analysis is S t = aX t + (1-a)S t-1 , wherein S t is the residual value smoothed by the sliding window at the current time (t), a is the smoothing coefficient, X t is the original residual at the current time (t), and S t-1 is the residual value of the previous time (t-1) smoothed by the sliding window; the edge analysis server preprocesses the collected images: after Gaussian denoising (kernel size 3x3, standard deviation 1.5), the main transformer joint, sleeve and other regions of interest (ROI) are extracted; sliding window time series analysis (window length 8 timestamps, smoothing coefficient a = 0.3) is used; the value of the sliding window smoothing coefficient a is related to the device state change rate: for the main transformer joint (temperature changes slowly), when a = 0.2, the historical data weight accounts for 80%, which can filter short-term interference (such as temperature fluctuations caused by sunlight), and the measured temperature curve smoothness is improved by 40%; for the pressure plate state (sudden failure), when a = 0.5, the current data weight is higher, and the abnormal response time is shortened from 1.2s to 0.6s, meeting the requirements of DL / T1498-2015 for rapid alarm; calculate the temperature residual: St = 0.3Xt + 0.7St-1;

[0031] Abnormal triggering of sound and light alarm, and double verification through weighted scoring of electrical quantity data and image recognition results, synchronous upload to the main station; in the double verification, the electrical quantity score P e Based on current and voltage deviation calculation, image score P i Based on defect detection confidence, weights w1+w2=1 and w1≥w2; when multiple sources of data are fused, infrared temperature measurement data and image data are synchronized through GPS clock, and the time error is ≤10ms; wherein the weighted scoring formula of electrical quantity data and image recognition results is: W = w1P e +w2P i , wherein W is the comprehensive fault score (value range 0-1), w1 and w2 are weight coefficients, P e is the fault score of electrical quantity data, and P i is the fault score of image recognition.

[0032] Advantages:

[0033] I. Solve the problem of high cloud processing delay and realize edge real-time early warning

[0034] Through knowledge distillation to generate a lightweight student model (parameter compression to less than 1 / 10 of the original, FLOPs≤1GFLOPs), the edge side inference time is reduced from 1.5 seconds of the cloud solution to ≤450ms (of which the model inference time is ≤190ms), combined with the 30MMSTP transmission delay ≤500ms, the total response time meets the 500ms early warning requirement of the main transformer overheating and other emergency failures; the measured data shows that the edge side abnormal detection delay of a certain 500kV substation is improved by 70% compared with the cloud solution, meeting the requirements of DL / T1498-2015 for rapid alarm;

[0035] II. Break through the limitation of edge device computing power and realize low resource and efficient inference

[0036] The lightweight technology of feature map distillation combined with parameter pruning compresses the model parameter from 7MB of traditional YOLOv5s to ≤5MB, reduces the floating point operation to below 0.8GFLOPs, and adapts to the computing resources of edge servers (such as NVIDIA Jetson AGX Orin); after deployment, the inference frame rate is improved to 18fps (meeting the ≥15fps requirement), the CPU utilization is stable below 30%, 8-way 4K video streams can be processed simultaneously, and the false rejection rate is reduced from 12% to below 0.5% compared with existing edge solutions;

[0037] III. Multi-source data fusion reduces false rejection and false alarm rates and improves decision reliability

[0038] The dual verification mechanism of infrared temperature measurement, visible light image and electrical quantity data fusion (weighted scoring formula W = w1P e +w2P i , w1≥w2) reduces the false rejection rate of a single data source from 18% (infrared) / 25% (visible light) to 3.2%, and the false rejection rate from 25% to below 0.5%; actual measurement shows that the detection accuracy of early overheating (temperature not exceeding threshold but current deviation 15%) of the main transformer joint is 98.7%, meeting the detection requirements of early defects in 《DL / T1998-2019》;

[0039] IV. Dynamic threshold and online updating mechanism to improve environmental adaptability

[0040] Sliding window time series analysis (α = 0.2-0.5, window length 5-10) combined with Kalman filter dynamically adjusts the warning threshold, eliminating interference such as sunlight and load fluctuations, reducing false alarm rate from 30% to below 5% in summer and false detection rate from 15% to below 2% in winter; Model differential upgrade mechanism (update package ≤1MB, time consumption ≤450ms) realizes real-time optimization of edge model, and the identification accuracy of new type of defects is improved 20 times faster than traditional offline update, ensuring long-term stable operation of the system;

[0041] V. Multi-source data synchronization and secure transmission guarantee

[0042] GPS clock synchronization mechanism (time error ≤10ms) ensures the spatio-temporal consistency of infrared and image data, avoiding false positives caused by clock deviation; SM4 encryption communication combined with 30MMSTP private line, transmission delay ≤500ms and data security meet the requirements of 《Q / GDW11337-2014》, meeting the requirements of substation security protection specification.

[0043] The present application is not limited to the above-mentioned embodiments, and any person should know that the structural changes made under the inspiration of the present application, any technical solutions with the same or similar to the present application, fall within the scope of protection of the present application. The technical, shape, structure parts not described in detail in the present application are well-known technologies.

Claims

1. A method for substation edge image recognition and intelligent early warning based on distillation learning, characterized in that, Includes the following steps: Visible light and infrared sensing devices are deployed in the substation to collect equipment images and temperature data and transmit them to the edge analysis server. The cloud server generates a lightweight student model through knowledge distillation technology. The distillation process includes: softening the output probability distribution of the teacher model based on the temperature parameter T, and combining feature map distillation and parameter pruning to compress the number of model parameters to less than 1 / 10 of the original model. The edge analysis server uses a student model to preprocess images and then performs real-time inference, combining sliding window time series analysis to identify anomalies. When an anomaly occurs, an audible and visual alarm is triggered. The system performs dual verification by fusing electrical quantity data with image recognition results from multiple sources, and also performs dual verification through weighted scoring, and simultaneously uploads the data to the main station.

2. The method for substation edge image recognition and intelligent early warning based on distillation learning according to claim 1, characterized in that, When deploying sensing equipment, the main transformer area of ​​the 500kV substation is equipped with a binocular temperature measuring pan-tilt unit and a visible light PTZ camera, while the indoor cabinet area is monitored without blind spots by a lifting rod to achieve pressure plate and indicator light monitoring.

3. The method for substation edge image recognition and intelligent early warning based on distillation learning according to claim 1, characterized in that, The temperature parameter T for knowledge distillation ranges from 5 to 15, and after parameter pruning, the floating-point computation (FLOPs) of the model is ≤1 GFLOPs.

4. The method for substation edge image recognition and intelligent early warning based on distillation learning according to claim 1, characterized in that, Image preprocessing includes Gaussian denoising and region of interest extraction, where the region of interest is the key component of the device.

5. The method for substation edge image recognition and intelligent early warning based on distillation learning according to claim 1, characterized in that, In sliding window time series analysis, the smoothing coefficient α is set to 0.2-0.5, and the window length is 5-10 timestamps.

6. The method for substation edge image recognition and intelligent early warning based on distillation learning according to claim 1, characterized in that, In dual verification, the electrical quantity score P e Image score P is calculated based on current and voltage deviation. i Based on the defect detection confidence, the weights are w1+w2=1 and w1≥w2, where w1 is the electrical quantity weight coefficient and w2 is the image weight coefficient.

7. The method for substation edge image recognition and intelligent early warning based on distillation learning according to claim 1, characterized in that, When fusing multi-source data, infrared temperature measurement data and image data are synchronized via GPS clock, with a time error of ≤10ms.

8. The method for substation edge image recognition and intelligent early warning based on distillation learning according to claim 1, characterized in that, The edge analytics server communicates with the main station via a 30M MSTP dedicated line, and the data encryption uses the SM4 algorithm with a transmission latency of ≤500ms.

9. The method for substation edge image recognition and intelligent early warning based on distillation learning according to claim 1, characterized in that, The student model supports online updates: After the main site collects misidentified samples, it pushes the model update package through differential upgrade, and the update process does not affect real-time inference.

10. A method for substation edge image recognition and intelligent early warning based on distillation learning according to claim 1, characterized in that, The formula for the output probability distribution of the teacher softening model based on temperature parameter T is: Where, q i Let z be the softened prediction probability of the edge-side student model for the i-th type of wind turbine operating condition / fault type. i denoted as the raw output score of the cloud-based teacher model for the i-th type of operating condition / fault, where T is the temperature parameter of knowledge distillation and N is the total number of operating condition / fault categories.

Images