A method for image recognition and temperature compensation of thermal defects in busbars based on edge computing

By performing local closed-loop processing for bus terminal identification and temperature compensation at the edge terminal, the problems of large infrared temperature measurement error and high cloud latency in substation busbars are solved, enabling accurate identification and real-time monitoring of bus terminals and improving the accuracy and reliability of thermal defect identification.

CN123118548APending Publication Date: 2026-11-17SUZHOU PUHUI NEW ENERGY TECH CO LTD
View PDF 2 Cites 0 Cited by

Patent Information

Application Number
CN202611346700.5
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-09-01
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing technologies for infrared temperature measurement of busbars in substations are easily affected by equipment emissivity, measurement distance, and ambient temperature, resulting in large measurement errors and data distortion. This makes it difficult to meet the needs of refined and real-time operation and maintenance monitoring of power equipment. Furthermore, cloud processing has high latency, which cannot meet the need for millisecond-level real-time early warning.

Method used

A lightweight CNN convolutional neural network is used to identify bus terminals and automatically lock ROIs at the edge terminal. A dynamic temperature compensation model with multiple environmental parameters is constructed, including ambient temperature deviation, temperature measurement distance attenuation and emissivity correction, to achieve on-site closed-loop processing of image recognition, target area locking, temperature compensation and defect judgment at the edge terminal.

Benefits of technology

It achieves accurate identification and temperature compensation of bus terminals under complex backgrounds, eliminates background radiation interference, improves the accuracy and reliability of thermal defect identification, meets the millisecond-level real-time monitoring requirements of substation equipment, and ensures the safe and stable operation of bus equipment.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123118548A_ABST
    Figure CN123118548A_ABST
Patent Text Reader

Abstract

This invention discloses a method for image recognition and temperature compensation of busbar thermal defects based on edge computing, belonging to the fields of power equipment condition monitoring, edge computing, and machine vision temperature measurement technology. The method completes infrared image acquisition and preprocessing at the edge, without uploading the original infrared image to the cloud. The preprocessed image is input into an improved MobileNetV2 lightweight convolutional neural network, which uses the SE attention module and FPN multi-scale feature fusion to detect targets, outputting busbar terminal bounding boxes and generating ROIs. A linear empirical correction model is constructed by fusing ambient temperature, measurement distance, and emissivity. Using the original temperature value within the ROI as input, the compensation coefficient is solved in a closed loop through on-site calibration and variable substitution to complete the temperature correction. The compensated temperature is compared with a preset threshold, and thermal defects are determined locally at the edge, outputting an early warning signal. This invention forms a complementary error closed loop between ROI locking and temperature compensation, achieving millisecond-level local early warning and improving the real-time performance and temperature measurement accuracy of busbar thermal defect monitoring.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention belongs to the fields of power equipment condition monitoring, edge computing and machine vision temperature measurement technology. Specifically, it relates to a method for intelligent identification and temperature compensation correction of busbar thermal defects that completes image processing at the edge side. It is mainly applicable to infrared temperature measurement monitoring and thermal defect early warning scenarios of power equipment such as substation busbars, terminals, and busbar joints. Background Technology

[0002] Currently, infrared thermography is commonly used to monitor equipment heating defects in power busbars. However, infrared thermography results are easily affected by external parameters such as equipment emissivity, measurement distance, and ambient temperature. Fixed-parameter temperature measurement modes cannot adapt to the complex on-site environment of substations, resulting in large measurement errors, data distortion, and an inability to accurately reflect the actual temperature rise of the busbar.

[0003] In the prior art, patent document CN114356892A discloses a method for detecting defects in infrared thermal images of power equipment based on deep learning. It uses the YOLOv5 network to detect the equipment area in the infrared thermal image and uploads the detected area to the cloud server for temperature analysis and defect judgment. Although this scheme improves the accuracy of target detection, its entire processing flow depends on cloud computing power and has the following defects: (1) The original infrared image needs to be uploaded to the cloud via fiber optic or 4G network. The data transmission delay is usually in the second range, which cannot meet the millisecond-level real-time early warning requirements of substation bus equipment; (2) In the centralized cloud processing mode, a large amount of invalid background data (cabinet, cable, debris, etc.) is mixed into the calculation process, which increases unnecessary bandwidth occupation and calculation overhead.

[0004] Furthermore, patent document CN116776543A discloses an infrared temperature measurement compensation method based on edge computing, which performs ambient temperature compensation calculations on infrared temperature measurement data at the edge. Although this scheme pushes some calculations down to the edge, it still uses fixed emissivity parameters and a single ambient temperature compensation model. It neither incorporates temperature measurement distance attenuation into the compensation model nor locks the temperature measurement area in any spatial way, so background radiation from cabinets, cables, etc., will inevitably mix into the temperature measurement data. At the same time, this scheme does not disclose the complete calibration process of the compensation coefficient, so those skilled in the art cannot reproduce the complete compensation process based on this.

[0005] Meanwhile, due to interference from complex backgrounds such as cabinets, cables, and on-site debris, existing solutions cannot accurately locate the core temperature measurement area of ​​the busbar. A large amount of invalid data is mixed into the calculation process, reducing the accuracy and reliability of busbar thermal defect identification and making it difficult to meet the needs of refined and real-time operation and maintenance monitoring of power equipment. Summary of the Invention

[0006] Purpose of the invention The purpose of this invention is to overcome the shortcomings of existing technologies and provide a method for image recognition and temperature compensation of bus thermal defects based on edge computing. This method solves the technical problems of large temperature measurement errors, high cloud processing latency, and inaccurate target recognition in complex backgrounds in traditional solutions. The method deploys image recognition, target area locking, temperature compensation, and defect determination on the edge terminal, completing the "collection-processing-analysis-early warning" process locally without relying on cloud computing power.

[0007] Compared with existing technologies, this invention proposes solutions to two specific technical obstacles in busbar monitoring scenarios: First, the on-site shooting distance in substations varies greatly within the range of 0.5m to 5m. Coupled with complex background interference such as cabinets, cables, and debris, the busbar terminals exhibit significant multi-scale and small target characteristics in infrared images, making it difficult to stably lock onto the core temperature measurement area. Second, the infrared temperature measurement results are simultaneously affected by three factors: equipment emissivity, temperature measurement distance, and ambient temperature. Fixed parameter modes cannot adapt to on-site working conditions.

[0008] Therefore, the present invention adopts the following design: (1) Feature maps are extracted from layers 3, 5, 7, and 9 of the backbone network, and multi-scale fusion is completed by upsampling and element-wise addition. Target detection is performed on the multi-scale fused feature map so that the network can adapt to the scale changes of the bus terminal under different shooting distances and angles; (2) The bounding box output by the target recognition is used as the spatial constraint for temperature measurement. Only the original temperature measurement value within the ROI coverage area is used as the input of the compensation model, thereby eliminating the pollution of the temperature measurement result by background radiation; (3) A linear empirical correction model is established that simultaneously accommodates environmental temperature deviation, temperature measurement distance attenuation, and emissivity correction. Its coefficients are obtained by fitting the field calibration test in one go. The above (2) and (3) constitute an on-site closed loop of "identification-location-temperature measurement-judgment": ROI locking provides an accurate temperature measurement area for temperature compensation, and temperature compensation provides a reliable temperature basis for defect judgment. The two are completed in series within the edge terminal.

[0009] Technical solution To achieve the above objectives, the present invention adopts the following technical solution: A method for image recognition and temperature compensation of thermal defects in busbars based on edge computing includes the following steps: S1. Edge-side infrared image acquisition and preprocessing: Infrared thermal images of the busbar are acquired at the busbar monitoring site using an infrared image sensor, and the raw infrared images are transmitted to the edge terminal. The edge terminal performs image noise reduction, grayscale equalization, and pixel normalization preprocessing locally to obtain the preprocessed infrared image. The raw single-channel temperature data output by the infrared image sensor is mapped into a three-channel image through pseudo-color encoding (such as an ironbow palette) before participating in subsequent processing. All image preprocessing is completed on the edge side, and the raw images do not need to be uploaded to the cloud.

[0010] S2. Bus terminal identification and automatic ROI locking based on a lightweight CNN convolutional neural network: A preprocessed infrared image of size 224×224×3 is input into the lightweight convolutional neural network. The lightweight convolutional neural network is a lightweight backbone network improved based on MobileNetV2, containing 9 convolutional layers, with SE attention modules added after the 4th and 7th convolutional layers respectively. Below the backbone network is an FPN multi-scale feature fusion layer, which draws features from the 3rd, 5th, 7th, and 9th layers of the backbone network. Multi-scale feature maps are extracted layer by layer, and feature fusion is completed by upsampling and element-wise addition to obtain a multi-scale fused feature map. Detection heads are set on the fused feature map at each scale. The detection heads include parallel classification branches and regression branches, which output the target category confidence and bounding box coordinates at each spatial location at that scale, respectively. The detection results output at each scale are fused by non-maximum suppression to obtain the target category and bounding box coordinates of the bus terminal. Regions of interest (ROIs) are generated based on the bounding box coordinates, and the core temperature measurement area of ​​the bus terminal is automatically selected, while invalid background areas such as shielding cabinets, redundant cables, and environmental obstructions are excluded.

[0011] S3. Dynamic Temperature Compensation Correction for Multiple Environmental Parameters: The edge terminal synchronously acquires the ambient temperature collected by the ambient temperature sensor and the temperature measurement distance output by the ranging module, and reads the inherent emissivity parameters of the bus equipment; it takes the characteristic temperature of the infrared temperature data within the ROI coverage area as the original temperature measurement value Traw; using the original temperature measurement value Traw as input, a linear empirical correction model is constructed; the linear empirical correction model introduces the ambient temperature deviation coefficient, the temperature measurement distance attenuation coefficient, and the emissivity correction coefficient to uniformly correct the original temperature measurement data and output the compensated true temperature data of the bus.

[0012] S4. Local bus thermal defect judgment and early warning output at the edge: The actual temperature after compensation is compared with the preset bus safe temperature rise threshold and temperature difference threshold. The edge terminal judges locally whether the bus has overheated or abnormal temperature rise thermal defects. When a thermal defect is judged to exist, the monitoring results and early warning signals are output in real time to achieve millisecond-level local early warning.

[0013] Furthermore, the lightweight convolutional neural network training method in step S2 is as follows: collect a massive dataset of infrared samples of bus joints and terminals in substations, label the samples, divide them into training set, validation set, and test set, complete network training locally or in the cloud, and deploy the model to edge terminal after training is completed.

[0014] Furthermore, in step S3, the method for extracting the original temperature measurement value Traw is as follows: take the average temperature of the top 5% of pixels with the highest temperature within the ROI coverage area; compared to directly taking the maximum value of the infrared temperature matrix within the ROI coverage area, the method of taking the average temperature of the top 5% of pixels can suppress the influence of isolated noise points and transient interference in the infrared temperature matrix on the original temperature measurement value, thereby improving the stability and reproducibility of the original temperature measurement value.

[0015] Furthermore, the dynamic temperature correction model described in step S3 has the following correction calculation logic: Treal = f(Traw, kt, kd, ke) In the formula, Treal is the compensated true temperature; Traw is the original temperature measurement value within the ROI area; kt is the ambient temperature deviation coefficient; kd is the temperature measurement distance attenuation coefficient; ke is the emissivity correction coefficient; each coefficient is obtained through the following field calibration test, the specific steps of which are as follows: (1) Select a standard heat source with a known temperature (such as a constant temperature heating block) at the substation site, and place the standard heat source at different known distances d1, d2, and d3 from the busbar equipment under test (corresponding to the standard output value of the ranging module). (2) At each distance, adjust the ambient temperature to different set values ​​T_env1, T_env2, T_env3 (covering the temperature range that may occur on site), and record the original temperature measurement value Traw output by the infrared sensor in the ROI area and the actual temperature Tstd of the standard heat source; (3) Based on the collected data of multiple sets (Traw, Tstd, ΔT, d), the coefficients kt, kd, and ke in the formula Treal = Traw×ke×(1+kt×ΔT+kd / d) are fitted. Considering that the formula is nonlinear for kt, kd, and ke as a whole, intermediate variables A = ke, B = ke·kt, and C = ke·kd are introduced to transform it into a linear regression model for the three feature terms [Traw,Traw·ΔT, Traw / d]. A, B, and C are solved at once by the least squares method, and the three coefficients are restored by ke = A, kt = B / A, and kd = C / A, so that the sum of squared residuals Σ(Tstd − Traw×ke×(1+kt×ΔT+kd / d))² is minimized. (4) Since the parameters to be solved are linear after the above transformation, the least squares solution has a unique closed solution. The calibration results do not depend on the initial value of the parameters, the calibration order or the iteration path, and the same set of data can be stably reproduced. The obtained kt, kd and ke coefficients are stored in the local configuration file of the edge terminal for direct call in real-time inference.

[0016] Furthermore, the above-mentioned modified model is an engineering linear empirical model. It should be noted that the physical basis of infrared radiation thermometry is the Stefan-Boltzmann law, which states that radiative exitance is proportional to the fourth power of the thermodynamic temperature. Therefore, when substituting into the above formula for calculation, Traw and T_env should be uniformly converted to thermodynamic temperature (unit: K) before being included in addition and multiplication operations. The resulting Treal is then subtracted by 273.15 to convert it to Celsius output, and ΔT is the converted thermodynamic temperature deviation. After the above processing, the dimensions of kt in the formula are 1 / K, the dimension of kd is meters (m), and ke is a dimensionless emissivity correction coefficient. All dimensions are consistent, and the fitting and solution process remains unchanged. The kd / d term in the formula is used for a first-order approximation to characterize the atmospheric transmission attenuation caused by increased measurement distance, as well as the combined effect of background radiation mixing when the target does not completely fill the sensor's field of view. It has sufficient fitting accuracy within a close-range monitoring range of 0.5m to 5m.

[0017] Furthermore, in step S4, the warning signal includes: equipment number, ROI location coordinates, original temperature measurement value, compensated temperature, environmental parameters, and defect level information.

[0018] Furthermore, the edge terminal does not upload the original infrared images to the cloud, but only selectively uploads the recognition results, compensated temperature, and alarm events.

[0019] This invention also provides a bus thermal defect monitoring system based on edge computing to implement the aforementioned method, including: an infrared image sensor, an ambient temperature sensor, a ranging module, and an edge computing terminal; the edge computing terminal has a built-in lightweight CNN network inference module, an ROI locking module, a temperature compensation calculation module, and a defect judgment and alarm output module.

[0020] The present invention also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the above-described edge computing-based bus thermal defect image recognition and temperature compensation method.

[0021] Beneficial effects Compared with the prior art, the present invention has the following beneficial effects: First, the present invention adopts a fully edge-side localized processing architecture, in which image acquisition, preprocessing, target recognition, ROI locking, temperature compensation, and defect determination are all completed on the edge terminal, without relying on the large computing power of the cloud, avoiding the risks of latency and data loss caused by cloud transmission, and meeting the millisecond-level real-time monitoring and early warning requirements of substation equipment.

[0022] Second, an improved MobileNetV2 lightweight convolutional neural network is adopted, with the addition of an SE attention module and an FPN multi-scale feature fusion layer. Detection is performed on the multi-scale fused feature map to adapt to the low computing power hardware conditions of edge terminals. It can accurately identify bus terminals in complex substation backgrounds, automatically generate ROI temperature measurement areas, eliminate invalid background interference, and improve the accuracy of thermal defect identification.

[0023] Third, a linear temperature correction model is constructed that simultaneously accommodates three types of influencing factors: ambient temperature, temperature measurement distance, and equipment emissivity, to uniformly correct for multi-source interference in infrared thermometry. After variable substitution, the model can be transformed into a linear regression problem, and its coefficients are solved in a closed loop by on-site calibration tests. The calibration results are unique, reproducible, and independent of initial values ​​and calibration order, overcoming the shortcomings of existing fixed emissivity and single ambient temperature compensation methods that have large errors, and obtaining more realistic temperature data that is closer to the equipment's operating conditions.

[0024] Fourth, this invention takes into account the accuracy of identification, the accuracy of temperature measurement, and real-time performance. It can detect potential overheating hazards of busbars in advance, ensure the safe and stable operation of substation busbar equipment, and adapt to the needs of intelligent and refined operation and maintenance of power.

[0025] Fifth, in this invention, ROI locking and temperature compensation form an irreplaceable local closed loop. It should be noted that the temperature compensation model can only correct parametric deviations introduced by emissivity, measurement distance, and ambient temperature. If ROI locking is not performed and temperature is taken directly from the entire infrared image, or if the ROI fails to accurately cover the bus terminals, causing radiation from cabinets, cables, or obstructions to mix into the measurement area, the original measured temperature value (Traw) itself has already deviated from its target—this type of deviation is considered "measuring the wrong object," and its magnitude is far greater than parametric deviations, and cannot be eliminated by any temperature compensation model. Conversely, if there is only a precise ROI without compensation, deviations introduced by the environment, distance, and emissivity still exist. Therefore, ROI locking and temperature compensation are mutually dependent and cannot be substituted for each other in this method. They form a closed loop in series within the edge terminal, and their combined effect is superior to either component implemented alone. Attached Figure Description

[0026] Figure 1 is a flowchart of a bus thermal defect image recognition and temperature compensation method based on edge computing according to the present invention. The figure shows the complete data flow from image acquisition to early warning output: Step S1 (edge-side infrared image acquisition and preprocessing) acquires the original infrared image and performs preprocessing operations such as noise reduction, grayscale equalization, and pixel normalization; Step S2 (lightweight CNN recognition and ROI auto-locking) pointed to by the solid arrow, the lightweight convolutional neural network completes the identification of bus terminals and generates regions of interest; Step S3 (multi-parameter dynamic temperature compensation correction) pointed to by the solid arrow, the ambient temperature, temperature measurement distance and emissivity parameters are fused to perform dynamic temperature correction; Step S4 (edge-side thermal defect judgment and early warning output) the defect judgment is completed on-site and an early warning signal is output.

[0027] Explanation of reference numerals in the attached figures: In Figure 1: 1—Edge-side infrared image acquisition and preprocessing; 2—Lightweight CNN recognition and automatic ROI locking; 3—Multi-parameter dynamic temperature compensation correction; 4—Edge-side thermal defect judgment and early warning output. Detailed Implementation

[0028] The specific embodiments of the present invention will be further described in detail below with reference to the accompanying drawings. Example 1

[0029] An edge computing-based method for busbar thermal defect image recognition and temperature compensation is used for infrared online monitoring of busbar joints in substations.

[0030] S1. Edge-side Acquisition Preprocessing: The infrared thermal imaging sensor acquires 224×224-sized bus infrared images in real time and sends them to the edge terminal (embedded edge box). The edge terminal performs noise reduction, grayscale equalization, and pixel normalization. The raw images are not sent to the cloud. Specifically, the raw single-channel temperature data output by the infrared thermal imaging sensor is mapped to an RGB three-channel image using pseudo-color encoding (such as an ironbow palette) to meet the network input size requirement of 224×224×3. The temperature measurement distance is obtained by a laser ranging module, and the ambient temperature is collected by an ambient temperature sensor.

[0031] The lightweight convolutional neural network has an input image size of 224×224×3. The backbone network is built on MobileNetV2 (width multiplier is 0.5). The total number of parameters is about 2.8 million, and the total computation is less than 0.4 GFLOPs. The single-frame inference time on the edge computing terminal does not exceed 22ms, which meets the millisecond-level latency requirements for real-time monitoring of substation busbars.

[0032] During the training process of the lightweight convolutional neural network, infrared sample datasets of bus joints and terminals in substations are collected and labeled. The training set, validation set, and test set are divided in a ratio of 7:1.5:1.5. A joint loss function consisting of cross-entropy loss and Focal Loss is used, and the AdamW optimizer is used for training. The initial learning rate is 1e-3, and a cosine annealing strategy is used for learning rate scheduling. The total number of training rounds is 150. After training, the model is deployed to edge terminals.

[0033] S2. Lightweight CNN Network Recognition and ROI Locking: Preprocessed infrared images are fed into an improved MobileNetV2 lightweight network; the network backbone has 9 convolutional layers, with an SE attention module added after the 4th and 7th convolutional layers; the FPN layer extracts feature maps from the 3rd, 5th, 7th, and 9th convolutional layers of the backbone network, performs upsampling and element-wise addition fusion to obtain multi-scale fused feature maps; a detection head is set on the fused feature map at each scale, and the detection head consists of parallel classification and regression branches, which output the target category confidence and bounding box coordinates at each spatial location at that scale; the detection results at each scale are fused by non-maximum suppression to obtain the target category and target box coordinates of the bus terminal; the ROI region is obtained based on the target box, the cabinet and irrelevant cable background are excluded, and only the temperature measurement area of ​​the bus joint is retained.

[0034] Meanwhile, the backbone network output features are further processed by global average pooling (GAP) and fed into an auxiliary classification branch to determine whether bus terminals exist in the current image. This branch only performs presence detection to suppress false alarms in empty backgrounds and does not participate in bounding box regression. It is important to emphasize that bounding box regression must be performed on a multi-scale fused feature map that preserves spatial dimensions, rather than on the one-dimensional features after global average pooling.

[0035] S3. Dynamic Temperature Compensation: The edge terminal reads the current ambient temperature, the distance measurement value from the sensor to the bus terminal, and the inherent emissivity of the bus metal of this model; it takes the average temperature of the top 5% of the pixels with the highest temperature within the ROI coverage area as the original temperature measurement value Traw; it calls the pre-calibrated kt, kd, and ke coefficients, substitutes them into the compensation model to calculate, and outputs the compensated true temperature Treal. The environmental adaptive dynamic correction adopts the following formula: Treal = Traw × ke × (1 + kt × ΔT + kd / d) Where ΔT = T_env − T0, T0 is the standard reference temperature, which is 298.15 K (i.e., 25℃), T_env is the ambient temperature, substituted with thermodynamic temperature (unit: K), d is the temperature measurement distance, in meters, ke is the dimensionless emissivity correction coefficient, kt is the ambient temperature deviation coefficient, in dimension 1 / K, and kd is the temperature measurement distance attenuation coefficient, in dimension meters (m), to ensure that the dimensions of each item in the formula are consistent; the calculated Treal is the thermodynamic temperature, which is converted to degrees Celsius after subtracting 273.15.

[0036] The calibration method for the coefficients kt, kd, and ke is as follows: A constant-temperature heating block is selected at the substation site as a standard heat source and placed at five known distances from the tested busbar equipment: 0.5m, 1.0m, 2.0m, 3.0m, and 5.0m. At each distance, the ambient temperature is adjusted to four set values: 10℃, 20℃, 30℃, and 40℃. Each operating condition is measured at least three times, and the original temperature value Traw output by the infrared sensor and the actual temperature Tstd of the standard heat source are recorded. Based on covering at least 20 operating conditions and selecting at least 60 sets of (Traw, Tstd, ΔT, d) data, intermediate variables A = ke, B = ke·kt, and C = ke·kd are introduced. The above formula is transformed into a linear regression model about the characteristic term [Traw, Traw·ΔT, Traw / d]. A, B, and C are solved once using the least squares method. Then, ke = A and kt = ... B / A and kd = C / A are used to restore the three coefficients to minimize the sum of squared residuals; the fitted coefficients are then stored in the local configuration file of the edge terminal.

[0037] S4. Edge-based Local Defect Judgment: Preset temperature rise threshold and relative temperature difference threshold; compare the compensated actual temperature with the threshold. When the temperature rise or temperature difference exceeds the threshold, the edge terminal generates an alarm locally, outputting the device number, location, original temperature, compensated temperature, environmental parameters, and defect level; local audible and visual alarms are possible, only the alarm event and key values ​​are uploaded to the backend, not the original large image. Defect levels are classified into three levels: general defects, severe defects, and critical defects, according to the defect classification requirements specified in the power industry standard DL / T 664 "Application Specification for Infrared Diagnosis of Live Equipment".

[0038] The edge computing terminal is an embedded hardware device deployed at the substation site, such as an NVIDIA Jetson Orin series edge AI computing module or an edge AI computing platform with equivalent computing power.

[0039] Experimental verification To verify the technical effectiveness of this invention, a comparative experiment was conducted at a substation site using a constant-temperature heating block as the standard heat source. The hardware platform consisted of an edge computing terminal deployed at the substation site (e.g., an NVIDIA Jetson Orin series edge AI computing module or an edge AI computing platform with equivalent computing power). The infrared thermal imaging sensor had a resolution of 640×480, an input image size of 224×224×3, a total network parameter count of approximately 2.8 million, and a computational load of less than 0.4 GFLOPs. The calibration phase covered five distances: 0.5m, 1.0m, 2.0m, 3.0m, and 5.0m; and four ambient temperatures: 10℃, 20℃, 30℃, and 40℃. Each group was repeated three times, covering no fewer than 20 working conditions and collecting no fewer than 60 sets of measured data (Traw, Tstd, ΔT, d). In all experiments, the average temperature of the top 5% of pixels with the highest temperature was used as the statistical criterion for temperature measurement. The only difference was whether the temperature measurement area was the ROI coverage area or the entire image.

[0040] To verify the recognition effect of the proposed solution on bus terminal targets, a comparative test of recognition accuracy was conducted. The lightweight CNN network of the proposed solution was compared with existing mainstream network models, and the test results are shown in Table 1. The comparison objects were ① original MobileNetV2 (without SE, without FPN), ② MobileNetV2+SE (without FPN), ③ MobileNetV2+FPN (without SE), and ④ the proposed solution (SE+FPN). All tests were conducted on the same dataset and the same edge hardware, and the results are shown in the table below.

[0041]

[0042] As can be seen from Experiment 1, ④ the present invention is significantly better than ① to ③ in terms of small target recall rate and mAP@0.5, and the inference time is still no more than 22ms, which meets the millisecond-level real-time requirement. This proves that the SE attention module and FPN multi-scale feature fusion are suitable for the specific problem of "complex background + multi-scale terminals".

[0043] To verify the improved temperature measurement accuracy brought about by the multi-parameter dynamic temperature compensation correction step, temperature compensation ablation verification was conducted, and the test results are shown in Table 2. In this experiment, all strategies uniformly adopted a fixed region of interest (ROI) (non-automatically locked ROI) to eliminate the interference of ROI accuracy on the compensation effect. Four compensation strategies were compared: A: fixed emissivity, no compensation; B: emissivity correction only (only ke); C: emissivity + ambient temperature dual factors; D: the present invention, emissivity + ambient temperature + measurement distance three factors. Using the true temperature of the standard heat source Tstd as the benchmark, the mean absolute error (MAE) and root mean square error (RMSE) of the temperature measurement error were statistically analyzed, and the results are shown in the table below.

[0044]

[0045] As seen in Experiment 2, as the number of influencing factors included in the compensation model increased from zero to three, the temperature measurement MAE decreased monotonically from 4.8℃ to approximately 0.8–1.0℃. This demonstrates that emissivity, ambient temperature, and measurement distance all have a substantial impact on the on-site temperature measurement results, and unified calibration within the same model is necessary to achieve optimal results. Furthermore, since the model has a unique closed-form solution after variable substitution, the coefficients obtained by repeated fitting of the same batch of calibration data are consistent, and the calibration results are stable and reproducible.

[0046] To verify the comprehensive performance gain brought about by combining automatic ROI locking and multi-parameter dynamic temperature compensation, closed-loop ablation verification was conducted. It should be noted that Experiment 2 has already separately verified the factor completeness of the compensation model; this Experiment 3 is used to examine the differences in error sources between the ROI locking and temperature compensation stages. An ablation and cross-combination design was adopted, with all three configurations using the same temperature sampling statistical caliber: the average temperature of the top 5% of pixels with the highest temperature. The only difference was the sampling area: the first and third groups sampled the ROI coverage area, while the second group did not perform ROI locking and directly sampled the temperature from the entire infrared image, simulating existing technical solutions without spatial locking of the temperature measurement area. The qualitative comparison results of temperature measurement error, defect detection rate, and false alarm rate under different combinations are shown in the table below.

[0047]

[0048] Experiment 3 shows that: First, under the same condition of applying complete three-factor compensation, simply canceling ROI locking and instead taking temperature readings from the entire infrared image significantly increases the temperature measurement error, making it inferior to the "precise ROI + no compensation" combination. This is because when the temperature measurement area is not locked, radiation from the cabinet, cables, and environmental obstructions mixes into the temperature measurement area, causing an object-specific deviation in the original temperature value (Traw). That is, the measured object is no longer the bus terminal itself. This type of deviation is considered "measuring the wrong object," and its magnitude is greater than the parametric deviation, and cannot be eliminated by any temperature compensation model. Second, under the condition of precise ROI, the temperature measurement error after canceling compensation significantly increases compared to the present invention, indicating that the parametric deviations introduced by ambient temperature, measurement distance, and emissivity still remain and cannot be replaced by ROI locking. These results show that the two types of error sources are different in nature and cannot be substituted for each other; only when both are present can optimal temperature measurement accuracy be obtained, which corroborates the fifth beneficial effect mentioned above. Example 2

[0049] The difference between this embodiment and Embodiment 1 is that the present invention further provides a cloud-based collaborative iterative system, including a cloud-based model training server and an edge computing terminal; the edge computing terminal sends back some sample recognition results (including difficult samples and low-confidence samples) and corresponding annotation information to the cloud-based model training server; the cloud-based model training server performs fine-tuning on the lightweight model based on the returned difficult samples; after fine-tuning, the updated lightweight model is sent to the edge computing terminal through an encrypted channel; the edge computing terminal loads the updated model to replace the old model, thereby realizing online iterative upgrade of the model.

[0050] The difficult samples are those whose confidence level identified by the edge computing terminal is below a preset threshold (e.g., 0.6), or those locally marked as abnormal by the edge computing terminal but whose defect level cannot be determined. The cloud-based model training server also maintains a sample pool, performing deduplication and active learning filtering on the returned samples, including only the samples with the highest information gain in the fine-tuning training set to reduce the computational overhead of cloud training. It should be noted that in this embodiment, only the identification results and annotation information of the difficult samples are returned to the cloud; the original infrared images are not uploaded to the cloud.

[0051] The above embodiments are only used to illustrate the technical solution of the present invention and are not intended to limit it; those skilled in the art can make equivalent modifications based on the principle of the present invention, all of which fall within the protection scope of the present invention.

Claims

1. An edge computing-based bus thermal defect image recognition and temperature compensation method, characterized in that, The process includes the following steps: S1. Acquire an infrared thermal image of the busbar and perform preprocessing on the infrared thermal image locally on the edge computing terminal to obtain a preprocessed image; all preprocessing is performed locally on the edge computing terminal, and the original infrared image is not uploaded to the cloud. S2. Input the preprocessed image into a lightweight convolutional neural network. The network outputs the target category and bounding box coordinates of the bus terminal, and generates a region of interest (ROI) based on the bounding box coordinates. The lightweight convolutional neural network embeds an attention module in its backbone network and combines multi-scale feature fusion processing. S3. Fuse the ambient temperature, temperature measurement distance, and inherent emissivity of the bus equipment to construct a linear empirical correction model. Use the original temperature measurement value within the ROI region as input, and output the true temperature of the bus after compensation by the linear empirical correction model. S4. Compare the compensated busbar temperature with the preset safe temperature rise threshold and temperature difference threshold to determine whether there is a thermal defect in the busbar, and output an early warning signal when a thermal defect is determined to exist.

2. The method of claim 1, wherein, In step S2, the lightweight convolutional neural network adopts a lightweight backbone network improved based on MobileNetV2. The backbone network contains 9 convolutional layers, with SE attention modules added after the 4th and 7th convolutional layers.

3. The method of claim 2, wherein, In step S2, the lightweight convolutional neural network has an FPN multi-scale feature fusion layer below the backbone network. The FPN multi-scale feature fusion layer extracts multi-scale feature maps from layers 3, 5, 7, and 9 of the backbone network and performs feature fusion by upsampling and element-wise addition to obtain a multi-scale fused feature map. A detection head is set at each scale of the multi-scale fused feature map. The detection head includes a classification branch and a regression branch set in parallel, which output the target category confidence and bounding box coordinates of each spatial location at that scale. The detection results output at each scale are fused by non-maximum suppression to obtain the target category and bounding box coordinates of the bus terminal. The output features of the backbone network are further fed into an auxiliary classification branch after global average pooling to determine whether the bus terminal exists in the image. The auxiliary classification branch does not participate in bounding box regression.

4. The method of claim 1, wherein, In step S2, the input image size of the lightweight convolutional neural network is 224×224×3; Actual measurements show that, with a backbone network width multiplier of 0.5, the total number of network parameters is less than 3 million, the total computational load is less than 0.4 GFLOPs, and the single-frame inference time on the edge computing terminal does not exceed 22ms.

5. The method of claim 1, wherein, In step S3, the raw temperature value Traw is extracted by taking the average temperature of the top 5% of pixels with the highest temperature in the ROI coverage area, so as to suppress the influence of isolated noise and transient interference in the infrared temperature matrix in the ROI coverage area on the raw temperature value.

6. The method of claim 1, wherein, In step S3, the calculation formula of the linear empirical correction model is: Treal = Traw × ke × (1 + kt × ΔT + kd / d); where Treal is the compensated true temperature, Traw is the original temperature measurement value in the ROI area, and both Traw and T_env are substituted into the calculation with thermodynamic temperature (unit: K). The obtained Treal is converted to Celsius output after subtracting 273.15; ke is the dimensionless emissivity correction coefficient, kt is the ambient temperature deviation coefficient with dimension 1 / K, and kd is the temperature measurement distance attenuation coefficient with dimension meter (m); ΔT = T_env − T0, T0 is the standard reference temperature with a value of 298.15 K (i.e., 25℃), T_env is the ambient temperature, and d is the temperature measurement distance with unit meter; the kt, kd, and ke are obtained by fitting the data in one step from the field calibration test, based on covering no less than 20 working conditions and collecting no less than 60 samples. The data set (Traw, Tstd, ΔT, d) was obtained using the least squares method.

7. The method of claim 6, wherein, The on-site calibration test includes: placing a standard heat source at five known distances of 0.5m, 1.0m, 2.0m, 3.0m, and 5.0m from the busbar equipment under test; adjusting the ambient temperature to four set values ​​of 10℃, 20℃, 30℃, and 40℃ at each distance, and recording the original temperature value Traw and the actual temperature Tstd of the standard heat source within the ROI area under each operating condition; repeating the measurement at least 3 times for each operating condition, covering at least 20 operating conditions and collecting at least 60 sets of (Traw, Tstd, ΔT, d) data; introducing intermediate variables A = ke, B = ke·kt, and C = ke·kd, transforming the compensation formula into a linear regression model about the characteristic term [Traw, Traw·ΔT, Traw / d], solving A, B, and C at once using the least squares method, and then using ke = A, kt = B / A, and kd = C / A. The coefficients kt, kd, and ke are restored to minimize the sum of squared residuals Σ(Tstd − Traw × ke × (1 + kt × ΔT + kd / d))². The obtained coefficients are written into the local configuration file of the edge computing terminal for direct use during real-time inference.

8. The method of claim 1, wherein, In step S2, the training process of the lightweight convolutional neural network is as follows: collect infrared sample datasets of bus joints and terminals in substations and complete the annotation; divide the annotated samples into training set, validation set and test set; adopt a supervised learning strategy and train the network based on the joint loss function, which is used to fuse classification loss and regression loss. The model is updated using an adaptive optimizer and the learning rate is scheduled using a learning rate scheduling strategy. After training, the model is deployed to the edge computing terminal.

9. The method according to claim 1, characterized in that, All processing steps S1 to S4 are completed locally on the edge computing terminal. The edge computing terminal does not upload the original infrared image to the cloud, but only selectively uploads the recognition result, the compensated temperature, and alarm events.

10. The edge computing-based bus thermal defect image recognition and temperature compensation method according to any one of claims 1 to 9, characterized in that, The thermal defect levels are classified into general defects, severe defects, and critical defects. General defects are when the equipment temperature rises to the attention threshold, and the equipment can continue to operate, but the trend needs to be monitored. Severe defects are when the equipment temperature rises to the alarm threshold, and maintenance should be arranged as appropriate. Critical defects are when the equipment temperature rises to the emergency threshold, and immediate action is required.

11. The method according to any one of claims 1 to 10, characterized in that, It also includes a cloud-based collaborative iteration step: the edge computing terminal will send back the identification confidence level of difficult samples below a preset threshold and their corresponding annotation information to the cloud-based model training server. The difficult samples are those with an identification confidence level below 0.6, or those locally marked as abnormal by the edge computing terminal but whose defect level cannot be determined. Based on the returned difficult samples and annotation information, the cloud-based model training server will fine-tune the lightweight convolutional neural network, and after the fine-tuning is completed, it will send the updated model to the edge computing terminal via an encrypted communication link in a full or differential manner. The edge computing terminal loads the updated model to replace the old model, realizing online iterative upgrades of the model; wherein, the encrypted communication link is an encrypted communication link based on the TLS protocol, and performs two-way identity authentication between the cloud model training server and the edge computing terminal based on a pre-set certificate or key.

12. The method of claim 11, wherein, The cloud-based model training server also maintains a sample pool, performs deduplication and active learning filtering on the returned samples, and only includes the samples with the greatest information gain into the fine-tuning training set. The content transmitted back from the edge computing terminal to the cloud model training server does not include the original infrared image.

13. A busbar thermal defect monitoring system for implementing the method of any one of claims 1 to 12, characterized in that, include: Infrared image sensor, used to acquire infrared thermal images of the busbar; ambient temperature sensor, used to acquire ambient temperature data. The ranging module is used to obtain the temperature measurement distance; An edge computing terminal is connected to the infrared image sensor, the ambient temperature sensor, and the ranging module; the edge computing terminal has a built-in lightweight CNN network inference module, ROI locking module, temperature compensation calculation module, and defect judgment and alarm output module, and is configured to perform the steps of the method described in any one of claims 1 to 12.

14. The busbar thermal defect monitoring system of claim 13, wherein, It also includes a cloud-based model training server, which is connected to the edge computing terminal via an encrypted communication link. The encrypted communication link is based on the TLS protocol and performs two-way authentication between the cloud-based model training server and the edge computing terminal based on a pre-set certificate or key. The edge computing terminal is also used to send back difficult samples with a confidence level below a preset threshold and their corresponding annotation information to the cloud model training server. The difficult samples are samples with a confidence level below 0.6, or samples locally marked as abnormal but whose defect level cannot be determined. The cloud model training server is used to receive the difficult samples and annotation information, fine-tune the lightweight convolutional neural network, and after fine-tuning, send the updated model to the edge computing terminal through the encrypted communication link in a full or differential manner. The edge computing terminal loads the updated model to replace the old model, realizing online iterative upgrades of the model.

15. The busbar thermal defect monitoring system of claim 14, wherein, The cloud-based model training server also maintains a sample pool, performs deduplication and active learning filtering on the returned samples, and only includes the samples with the greatest information gain into the fine-tuning training set. The content transmitted back from the edge computing terminal to the cloud model training server does not include the original infrared image.

16. A computer-readable storage medium having stored thereon a computer program, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1 to 12.

Citation Information

Patent Citations

  • Data migration method and electronic equipment

    CN114356892A

  • Power big data application method for smart power grid

    CN116776543A