Soft insulating sling defect evaluation method, system and equipment based on image recognition

Abstract

The invention relates to the technical field of electric power safety, and discloses a soft insulation sling defect evaluation method, system and device based on image recognition. The method comprises the following steps: acquiring sensing layer data and high-resolution image data; performing feature extraction and feature fusion according to the perception layer data and the high-resolution image data to generate a fusion feature model; carrying out sling defect detection according to the fusion feature model to generate a preliminary defect identification result; based on a finite element algorithm, crack propagation trajectory simulation is carried out after the sling is stressed, and a crack simulation result is obtained; and generating a crack propagation prediction result based on the preliminary defect identification result and the crack simulation result, and performing safety scoring according to the crack propagation prediction result. According to the method, continuous real-time data is collected for defect analysis, and tiny defect recognition is achieved in combination with feature extraction. And the prediction accuracy is enhanced by predicting and analyzing data. And the data analysis depth is improved by fusing the feature model, so that the risk assessment is more comprehensive, and the timeliness and accuracy of fault prediction are improved.

Description

Technical Field

[0001] The present invention relates to the field of electric power safety technology, and in particular to a method, system and device for evaluating defects of soft insulating slings based on image recognition. Background Art

[0002] The field of power safety technology focuses on protecting and improving the operational safety of power systems, ensuring the stability of power equipment and infrastructure and minimizing their impact on the environment. Among them, defect assessment of soft insulating slings is a technology specifically used to detect and evaluate the safety and integrity of insulating slings used in power lines or related power facilities.

[0003] Existing power safety technologies rely on periodic inspections and basic electrical tests to deal with insulation strap defects, which are inadequate in detecting early or minor structural defects. At the same time, the lack of real-time monitoring capabilities and analysis of complex data makes it difficult to predict the development trend of defects, resulting in delayed preventive measures. For example, cracks that are not identified in time may cause sudden fractures when the power load increases, causing safety accidents and power outages. Therefore, improvements are urgently needed. Summary of the invention

[0004] The main purpose of the present invention is to provide a method, system and equipment for evaluating defects of soft insulating slings based on image recognition, aiming to solve at least one of the above-mentioned technical problems.

[0005] To achieve the above object, the present invention provides a method for evaluating defects of soft insulating slings based on image recognition, comprising:

[0006] Collect perception layer data and high-resolution image data;

[0007] Performing feature extraction and feature fusion according to the perception layer data and the high-resolution image data to generate a fusion feature model;

[0008] Perform sling defect detection according to the fusion feature model to generate a preliminary defect recognition result;

[0009] Based on the finite element algorithm, the crack propagation trajectory of the sling after being stressed is simulated to obtain the crack simulation results;

[0010] A crack propagation prediction result is generated based on the preliminary defect identification result and the crack simulation result, and a safety score is performed according to the crack propagation prediction result.

[0011] In some embodiments, the collecting of perception layer data and high-resolution image data includes:

[0012] Embedding a flexible sensor in a soft insulating sling, and monitoring the mechanical stress and surface state of the sling in real time based on the flexible sensor to collect mechanical stress data and surface state data;

[0013] Using the mechanical stress data and the surface state data as original monitoring data;

[0014] Calculating the weighted accuracy of the original monitoring data;

[0015] Filtering the original monitoring data based on the weighted accuracy, and generating perception layer data according to the filtered original monitoring data;

[0016] High-resolution image data is acquired based on image acquisition equipment.

[0017] In some embodiments, the acquiring of high-resolution image data based on an image acquisition device includes:

[0018] Based on the image acquisition device, an active light source and a dynamic exposure algorithm are used to acquire an initial RGB image, an initial IR image and an initial X-ray image;

[0019] Calculating a comprehensive contrast enhancement index of an image based on the initial RGB image, the initial IR image, and the initial X-ray image;

[0020] The initial RGB image, the initial IR image and the initial X-ray image are subjected to image enhancement processing based on the comprehensive contrast enhancement index of the image to obtain high-resolution image data.

[0021] In some embodiments, the extracting and fusing features according to the perception layer data and the high-resolution image data to generate a fused feature model includes:

[0022] Processing the high-resolution image data through a convolutional neural network to extract image features;

[0023] Analyzing the time series information of the perception layer data through a long short-term memory network to generate time series features;

[0024] The image features and time series features are input into a Transformer network configured with a self-attention mechanism to generate a fusion feature model.

[0025] In some embodiments, processing the high-resolution image data by a convolutional neural network to extract image features includes:

[0026] Inputting the high-resolution image data into a convolutional neural network for filtering and feature learning to obtain initial feature mapping data;

[0027] Calculating a comprehensive activity score of a feature based on the initial feature mapping data;

[0028] The initial feature mapping data is screened and optimized according to the comprehensive activity score to extract image features.

[0029] In some embodiments, analyzing the time series information of the perception layer data by a long short-term memory network to generate time series features includes:

[0030] Loading the perception layer data into the long short-term memory network, adjusting the network weights through forward propagation and back propagation algorithms, and obtaining time-dependent feature mapping data;

[0031] Obtaining refined temporal features based on the time-dependent feature mapping data according to a forget gate and an update gate;

[0032] Analysis and screening are performed based on the refined time series features to obtain time series features.

[0033] In some embodiments, generating a crack propagation prediction result based on the preliminary defect identification result and the crack simulation result, and performing a safety score according to the crack propagation prediction result, includes:

[0034] Combining the preliminary defect identification result with the crack simulation result to obtain a crack extension and defect identification data set;

[0035] Based on the crack propagation and defect identification data set, data analysis is performed through a long short-term memory network to generate a crack propagation prediction result;

[0036] A safety score is performed based on the crack extension prediction result.

[0037] In some embodiments, performing safety scoring according to the crack propagation prediction result includes:

[0038] Obtaining a severity score of a crack, a potential hazard value of a crack, and a total number of crack paths according to the crack propagation prediction result;

[0039] A safety score is calculated based on the severity score of the crack, the potential hazard value of the crack, and the total number of crack paths.

[0040] In addition, to achieve the above purpose, the present invention also proposes a soft insulating sling defect assessment system based on image recognition, comprising:

[0041] Data acquisition module, used to collect perception layer data and high-resolution image data;

[0042] A feature extraction module, used for performing feature extraction and feature fusion according to the perception layer data and the high-resolution image data to generate a fusion feature model;

[0043] A defect detection module, used to perform sling defect detection according to the fusion feature model and generate a preliminary defect recognition result;

[0044] The crack simulation module is used to simulate the crack extension trajectory of the sling after being stressed based on the finite element algorithm to obtain the crack simulation results;

[0045] A safety assessment module is used to generate a crack extension prediction result based on the preliminary defect identification result and the crack simulation result, and to perform a safety score according to the crack extension prediction result.

[0046] In addition, to achieve the above-mentioned purpose, the present invention also proposes an electronic device, which includes: a memory, a processor, and a soft insulating sling defect assessment program based on image recognition stored in the memory and run on the processor, and the soft insulating sling defect assessment program based on image recognition is configured to implement the soft insulating sling defect assessment method based on image recognition as described above.

[0047] The present invention provides a defect assessment method for soft insulating slings based on image recognition, including: collecting perception layer data and high-resolution image data; performing feature extraction and feature fusion according to the perception layer data and the high-resolution image data to generate a fusion feature model; performing sling defect detection according to the fusion feature model to generate a preliminary defect recognition result; simulating the crack extension trajectory of the sling after being stressed based on a finite element algorithm to obtain a crack simulation result; generating a crack extension prediction result based on the preliminary defect recognition result and the crack simulation result, and performing a safety score according to the crack extension prediction result. In the present invention, continuous and real-time perception layer data and high-resolution image data are collected for defect analysis, and the deep feature extraction of a convolutional neural network is combined to realize the recognition of tiny defects. By predicting and analyzing data, the prediction accuracy of future crack paths is enhanced. The depth of data analysis is further improved by fusion feature models, making risk assessment more comprehensive, thereby improving the timeliness and accuracy of fault prediction, and providing strong technical support for the safety management of power facilities. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 A schematic diagram of the structure of an electronic device in a hardware operating environment involved in an embodiment of the present invention;

[0049] Figure 2 It is a flow chart of an embodiment of a method for evaluating defects of soft insulating slings based on image recognition according to the present invention;

[0050] Figure 3 A schematic diagram of the steps involved in the embodiment of the present invention;

[0051] Figure 4 The present invention is a structural block diagram of an embodiment of a soft insulating sling defect assessment system based on image recognition.

[0052] The realization of the purpose, functional features and advantages of the present invention will be further explained in conjunction with embodiments and with reference to the accompanying drawings. DETAILED DESCRIPTION

[0053] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0054] It should be noted that all directional indications in the embodiments of the present invention (such as up, down, left, right, front, back, etc.) are only used to explain the relative position relationship, movement status, etc. between the components under a certain specific posture (as shown in the accompanying drawings). If the specific posture changes, the directional indication will also change accordingly.

[0055] In addition, the descriptions of "first", "second", etc. in the present invention are only used for descriptive purposes and cannot be understood as indicating or implying their relative importance or implicitly indicating the number of the indicated technical features. Therefore, the features defined as "first" and "second" may explicitly or implicitly include at least one of the features. In addition, the technical solutions between the various embodiments can be combined with each other, but it must be based on the ability of ordinary technicians in the field to implement. When the combination of technical solutions is contradictory or cannot be implemented, it should be deemed that such combination of technical solutions does not exist and is not within the scope of protection required by the present invention. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0056] Reference Figure 1 , Figure 1 The figure is a schematic diagram of the structure of an electronic device of the hardware operating environment involved in the embodiment of the present invention.

[0057] like Figure 1As shown, the electronic device may include: a processor 1001, such as a central processing unit (CPU), a communication bus 1002, a user interface 1003, a network interface 1004, and a memory 1005. Among them, the communication bus 1002 is used to realize the connection and communication between these components. The user interface 1003 may include a display screen (Display), an input unit such as a keyboard (Keyboard), and the optional user interface 1003 may also include a standard wired interface and a wireless interface. The network interface 1004 may optionally include a standard wired interface and a wireless interface (such as a wireless fidelity (Wireless-Fidelity, Wi-Fi) interface). The memory 1005 may be a high-speed random access memory (Random Access Memory, RAM memory) or a stable non-volatile memory (Non-Volatile Memory, NVM), such as a disk memory. The memory 1005 may also be a storage device independent of the aforementioned processor 1001.

[0058] Those skilled in the art will understand that Figure 1 The structure shown in the figure does not constitute a limitation on the electronic device, and may include more or less components than shown in the figure, or combine certain components, or arrange the components differently.

[0059] like Figure 1 As shown, the memory 1005 as a storage medium may include an operating system, a network communication module, a user interface module, and a soft insulating suspender defect assessment program based on image recognition.

[0060] exist Figure 1 In the electronic device shown, the network interface 1004 is mainly used for data communication with a network server; the user interface 1003 is mainly used for data interaction with a user; the processor 1001 and the memory 1005 in the electronic device of the present invention can be set in the electronic device, and the electronic device calls the soft insulating suspender defect assessment program based on image recognition stored in the memory 1005 through the processor 1001, and executes the soft insulating suspender defect assessment method based on image recognition provided by an embodiment of the present invention.

[0061] The present invention provides a method, system and equipment for evaluating defects of soft insulating slings based on image recognition.

[0062] The embodiment of the present invention provides a method for evaluating defects of soft insulating slings based on image recognition, referring to Figure 2 , Figure 2 The figure is a flow chart of an embodiment of a method for evaluating defects of soft insulating slings based on image recognition according to the present invention.

[0063] like Figure 2 As shown, the soft insulating sling defect assessment method based on image recognition includes:

[0064] Step S100: collecting perception layer data and high-resolution image data;

[0065] Step S200: performing feature extraction and feature fusion according to the perception layer data and the high-resolution image data to generate a fusion feature model;

[0066] Step S300: performing sling defect detection according to the fusion feature model to generate a preliminary defect recognition result;

[0067] Step S400: simulating the crack propagation trajectory of the sling after being stressed based on a finite element algorithm to obtain a crack simulation result;

[0068] Step S500: generating a crack propagation prediction result based on the preliminary defect identification result and the crack simulation result, and performing a safety score according to the crack propagation prediction result.

[0069] It should be noted that the execution subject in this embodiment may be an electronic device, which may be a computer device with data processing functions, or other devices that can achieve the same or similar functions. This embodiment does not limit this. In this embodiment, a computer device is used as an example for explanation.

[0070] In one embodiment, the sensing layer data and the high-resolution image data are collected, including: embedding a flexible sensor in a soft insulating sling, monitoring the mechanical stress and the surface state of the sling in real time based on the flexible sensor to collect the mechanical stress data and the surface state data; using the mechanical stress data and the surface state data as original monitoring data; calculating the weighted accuracy of the original monitoring data; filtering the original monitoring data based on the weighted accuracy, and generating the sensing layer data according to the filtered original monitoring data; and acquiring the high-resolution image data based on the image acquisition device.

[0071] Specifically, Figure 3 As shown, flexible sensors are embedded in the slings to collect mechanical stress data and surface state data to generate perception layer data; through image acquisition equipment such as industrial cameras and X-ray machines, active light sources and dynamic exposure algorithms are used to collect RGB, IR images and X-ray images to generate high-resolution image data.

[0072] Exemplarily, the perception layer data is obtained by embedding a flexible sensor in the sling, monitoring the mechanical stress and surface state of the sling in real time, collecting the mechanical stress data and the surface state data, and obtaining the original monitoring data; based on the original monitoring data, calculating the weighted accuracy of the data, the calculation formula is:

[0073]

[0074] Among them, x i is the value of the ith data point, w i is the weight of the ith data point, n is the total number of data points, μ is the average value of the data points, and R represents the weighted accuracy.

[0075] Exemplarily, based on the weighted accuracy, key mechanical stress and surface state data are screened to form screened monitoring data and obtain perception layer data.

[0076] In one example, a handheld marking tool can be used to specify the embedding point on the surface of the sling, a micro-knife head can be used to cut the outer fiber of the sling, and the flexible sensor can be marked one-to-one according to the identification card corresponding to the number and inserted into a specific position. The sensor can be fixed in the cross section by bonding, and its electrical lead can be connected to the reserved interface of the data acquisition unit. The acquisition unit is set to a fixed polling interval and the mechanical stress and surface state data at the corresponding moment can be recorded. In order to clarify the classification standard of the data value, the conventional range of mechanical stress is statistically recorded through the records of the sling load test, so as to obtain the upper and lower limits of the interval division, and to define that if the data exceeds the classification threshold, it is recorded as a high stress grade, otherwise it is recorded as a low stress grade. At the same time, the surface state is quantified, and the surface roughness and protrusion features are quantified into numerical values ​​by comparing with the standard physical scale sample, and recorded in the corresponding data file. At the same time, the timestamp and position index of each record are obtained, and finally all the obtained mechanical stress and surface state values ​​are collected to form the original monitoring data.

[0077] In this embodiment, the calculation formula for calculating the weighted accuracy rate is beneficial in that the weight parameter w is introduced i Differentiated processing of each data point allows the credibility of the data source and measurement deviation to be fully considered when calculating the weighted accuracy, thereby making a more reasonable accuracy assessment of the monitoring data overall.

[0078] In one example, get x i Parameters: Extract 5 consecutive data points from the original monitoring data, read the mechanical stress values ​​from the flexible sensor output one by one, and record the obtained mechanical stress values ​​as x1 = 13.2MPa, x2 = 14.1MPa, x3 = 12.8MPa, x4 = 13.7MPa, x5 = 13.4MPa. Get n parameter: Directly count the total number of data points, where n = 5. Get μ parameter: Take the average of the above 5 stress data points Get w iParameters: By statistically analyzing the measurement stability of each sensor in the previous sensor calibration data, the corresponding weight value is selected according to the discrete degree of multiple measurement results under the same load. For example, a larger weight is given to the sensor with a more stable output. By comparing the previous data, w1=0.9, w2=0.85, w3=0.8, w4=0.88, and w5=0.92 are determined. The weight value is obtained by linear transformation of the coefficient of variation of the historical measurement data of the sensor.

[0079] Calculation process: First calculate the square of the difference between each data point and the mean:

[0080] |x1-μ| 2 =|13.2-13.44| 2 =(-0.24) 2 =0.0576;

[0081] |x2-μ| 2 =|14.1-13.44| 2 =(0.66) 2 =0.4356;

[0082] |x3-μ| 2 =|12.8-13.44| 2 =(-0.64) 2 =0.4096;

[0083] |x4-μ| 2 =|13.7-13.44| 2 =(0.26) 2 =0.0676;

[0084] |x5-μ| 2 =|13.4-13.44| 2 =(-0.04) 2 =0.0016;

[0085] Multiply the above results by the weights and sum them up:

[0086] Calculation: =0.05184+0.37026+0.32768+0.059488+0.001472=0.81074

[0087] Divide by the total number of data points n = 5: 0.81074 / 5 = 0.162148

[0088] Finally, take the square root of the result:

[0089] The result shows that the weighted accuracy is 0.40267. For the current step, this value is used to evaluate the overall accuracy of the selected data. A value closer to 0 means that the data points are relatively concentrated, indicating that the sling mechanical stress measurement is closer to the average value range, while an increase in the value indicates that there is a large deviation between the points. This result can be further used in the subsequent key data screening step. When the weighted accuracy is less than 0.5, the corresponding data can be regarded as a relatively concentrated distribution, which facilitates the determination of key mechanical stress and surface state data.

[0090] In one example, based on the weighted accuracy, key mechanical stress and surface state data are screened to form screened monitoring data and obtain perception layer data: the recorded mechanical stress and surface state data are loaded into the comparison module through the obtained weighted accuracy parameters, and the data list is sorted according to the degree of distance from the overall data mean according to the weighted accuracy value just calculated (for example, 0.40267). In order to set the data screening limit value, reference values ​​are extracted from the monitoring data under the same environment in the past, and the corresponding distribution intervals are statistically calculated, from which the weighted accuracy is less than the threshold value. The sample with a value of 0.5 is taken as the concentrated distribution data, and the change of mechanical stress parameter in the high and low stress range is taken as the judgment criterion. The boundary value of high stress and low stress is determined according to the data statistics from previous tests. The surface state parameters are subjected to the same quantitative processing. The height of surface micro-protrusions measured by a referenced standard sample is used as the benchmark. The surface data is converted into a continuous numerical sequence and corresponded to the mechanical stress sequence. By comparing these concentrated distribution data point by point and eliminating points with excessive deviation, the remaining data list converges to the final screened monitoring data, and the data set is marked as the perception layer data.

[0091] In one embodiment, high-resolution image data is acquired based on an image acquisition device, including: acquiring an initial RGB image, an initial IR image, and an initial X-ray image based on the image acquisition device using an active light source and a dynamic exposure algorithm; calculating a comprehensive contrast enhancement index of the image based on the initial RGB image, the initial IR image, and the initial X-ray image; and performing image enhancement processing on the initial RGB image, the initial IR image, and the initial X-ray image based on the comprehensive contrast enhancement index of the image to obtain high-resolution image data.

[0092] Exemplarily, high-resolution image data is obtained: an industrial camera and an X-ray machine are set up, and an active light source and a dynamic exposure algorithm are used to obtain preliminary RGB, IR images and X-ray images (initial RGB images, initial IR images and initial X-ray images); based on the preliminary RGB, IR images and X-ray images, the comprehensive contrast enhancement index of the image is calculated, and the calculation formula is:

[0093]

[0094] Among them, C RGB Represents the contrast of RGB image, C IR represents the contrast of the IR image, C X Indicates the contrast of X-ray images.

[0095] Exemplarily, based on the comprehensive contrast enhancement index of the image, image enhancement is applied to the preliminary RGB, IR images and X-ray images to obtain high-resolution image data.

[0096] In one example, the shutter time of the industrial camera and the exposure intensity of the X-ray machine are set step by step with reference to the operating instructions provided by the manufacturer. According to the execution process, the RGB images captured by the industrial camera under different lighting conditions need to be recorded and stored in an image matrix format. At the same time, the X-ray images generated by the X-ray machine under the selected tube voltage and tube current need to be recorded and stored in the same format. In order to clarify the illumination range of the active light source and the control standard of dynamic exposure, the light source power and spectrum distribution data are obtained by referring to the equipment technical manual, and the exposure parameters are set in grades and recorded in the corresponding supporting parameter files. Each exposure parameter is called in turn to capture RGB, IR images and X-ray images. These images are associated with the corresponding exposure grades through number identification to form an index table. The file path of each set of RGB, IR image and X-ray image data recorded in the index table is parsed. The pre-recorded exposure parameter value and light source output intensity are queried through the number. The corresponding parameter value is used as a reference condition to gradually iterate the shooting process and record the generated preliminary image file set, and finally the preliminary RGB, IR image and X-ray image data set is output.

[0097] In this embodiment, the calculation formula for the comprehensive contrast enhancement index of the above-mentioned image is beneficial in that the comprehensive contrast enhancement index is obtained by weighted fusion of the contrast of RGB, IR and X-ray images, which helps to select more suitable enhancement parameters in subsequent image enhancement processing.

[0098] In one example, get C RGB Parameters: By obtaining the RGB image, record the distribution range of the image brightness histogram, and calculate the contrast to obtain C RGB =32.5; Get C IR Parameters: Through the obtained IR image, the brightness difference of the pixel matrix is ​​counted, and the ratio of the maximum brightness to the minimum brightness difference and the average brightness is taken and recorded as C IR =28.4; Get C X Parameters: The contrast ratio is the ratio of the maximum grayscale difference and the minimum grayscale difference of the pixel value divided by the average grayscale value in the obtained X-ray image, recorded as C X =30.2.

[0099] Calculation process: Substitute the parameter value into the calculation formula of the comprehensive contrast enhancement index of the image:

[0100]

[0101] First calculate the sum of squares:

[0102] (32.5) 2 =1056.25

[0103] (28.4) 2 =806.56

[0104] (30.2) 2 =912.04

[0105] Sum: 1056.25+806.56+912.04=2774.85

[0106] Divide by 3: 2774.85 / 3 = 924.95

[0107] Square root:

[0108] The results show that the comprehensive contrast enhancement index is 30.41. When the value is greater than 20, it shows that the contrast level of the multi-source image is in a better state, which is beneficial to the subsequent enhancement process. When the value is close to 30 or above, it means that the overall contrast of RGB, IR, and X-ray images is high, and the exposure compensation can be appropriately reduced in the subsequent enhancement process. When the value is less than 10, it means that the contrast is low, and the corresponding adjustment strength should be increased in the enhancement step.

[0109] In one example, based on the comprehensive contrast enhancement index of the image, image enhancement is applied to preliminary RGB, IR and X-ray images to obtain high-resolution image data: the preliminary RGB, IR and X-ray images obtained in the previous step are input through the obtained comprehensive contrast enhancement index parameter set, the recorded exposure parameter index and contrast reference value list are called, the preset contrast enhancement standard value is retrieved from the list, the standard value is compared with the comprehensive contrast enhancement index of the current image combination, and the exposure offset recorded in the light source power file is superimposed step by step according to the difference between the standard value given in the list and the actual contrast value, and each image channel is processed independently, and the pixel row and column indexes that need to improve the brightness and the pixel area coordinates of the enhanced edge are determined by comparing the spectral response record of the IR image, and the brightness parameters of the RGB image and the channel gain values ​​of the X-ray image are modified by the specified increments and the updated image matrix file is recorded again to generate high-resolution image data.

[0110] In this embodiment, a flexible sensor is embedded in the sling to collect mechanical stress data and surface state data to generate perception layer data. The flexible sensor is introduced to collect the mechanical stress data and surface state of the sling, providing continuous real-time data for defect analysis. The high-resolution images collected by industrial cameras and X-ray machines are combined with deep feature extraction of convolutional neural networks to realize the identification of tiny defects.

[0111] In one embodiment, feature extraction and feature fusion are performed based on the perception layer data and the high-resolution image data to generate a fusion feature model, including: processing the high-resolution image data through a convolutional neural network to extract image features; analyzing the time series information of the perception layer data through a long short-term memory network to generate timing features; and inputting the image features and timing features into a Transformer network configured with a self-attention mechanism to generate a fusion feature model.

[0112] Specifically, Figure 3 As shown in the figure, the high-resolution image data is processed by a convolutional neural network to extract image features and generate image features; the perception layer data is analyzed through a long short-term memory network to analyze the time series information and generate time series features; the image features and time series features are fused, and the global learning of the features is enhanced using the Transformer network to generate a fusion feature model.

[0113] In one embodiment, the high-resolution image data is processed by a convolutional neural network to extract image features, including: inputting the high-resolution image data into a convolutional neural network for filtering and feature learning to obtain initial feature mapping data; calculating a comprehensive activity score of the feature based on the initial feature mapping data; and screening and optimizing the initial feature mapping data according to the comprehensive activity score to extract image features.

[0114] Exemplarily, image features are obtained: high-resolution image data is input into a convolutional neural network, and the convolutional neural network obtains preliminary feature mapping data (initial feature mapping data) through filtering and feature learning; based on the preliminary feature mapping data, the comprehensive activity score of the feature is calculated, and the improved calculation formula is:

[0115]

[0116] Among them, f k represents the activation value of the kth feature map, K is the total number of feature maps, e is the base of the natural logarithm, and F is the comprehensive activity score.

[0117] Exemplarily, based on the calculated comprehensive activity score, the features of the initial feature mapping data are screened and optimized to extract image features.

[0118] In one example, high-resolution image data is input into a convolutional neural network. Next, the received image data is channel-decomposed, and the data is formed into a matrix sequence according to the RGB channels. The convolution kernel size information such as 3×3 and the corresponding step size of 1 are read from the device calibration file. These parameters are recorded in the calibration data provided by the device manufacturer. The convolution kernel is moved row by row and column by column to calculate the output matrix data value by performing matrix multiplication and numerical accumulation and summation processing on each pixel matrix element. Each output matrix data value corresponds to the weighted comprehensive result of the local pixel area of ​​the input matrix. In order to reduce the amount of data, a downsampling ratio such as 2:1 is selected from the device parameter table. According to the ratio, several row and column positions are skipped to obtain data. The downsampling output matrix is ​​recorded in the file with row and column indexes, and the corresponding convolution kernel number is marked. The above steps are repeated for different convolution kernels, and the corresponding result matrix data values ​​are recorded respectively. Channel merging operation is performed on all output matrices. The merged matrix data is numbered and recorded by querying the corresponding mapping index in the configuration file. The matrices corresponding to all numbers are recorded in the same data file as a preliminary set of feature mapping values. After completing the above steps, all feature mapping matrix data are integrated and output to obtain preliminary feature mapping data (initial feature mapping data).

[0119] In this embodiment, the calculation formula for the improved comprehensive activity score of the feature is beneficial in that a logical function is introduced to transform the feature map activation value, the activities of multiple feature maps are fused in a product manner, and the Kth root is taken, thereby forming a quantitative score for the overall activity of the feature.

[0120] In one example, f is obtained k Parameters: Select the kth feature map from the obtained initial feature map data, add up all its pixel values ​​and divide by the total number of pixels to obtain the average activation value f k For example, by analyzing the data file corresponding to the feature map, the map contains 1000 pixels, and the activation value range is approximately between 0 and 2. The average of the map is obtained to obtain f1=0.8, f2=1.2, f3=0.5, f4=1.5, and f5=0.9. Get the K parameter: The total number of statistical feature maps K=5.

[0121] Calculation process: Substitute the above parameter values:

[0122]

[0123] Compute the exponential term:

[0124]

[0125] Multiply all the values ​​together:

[0126] 0.6900×0.7684×0.6228×0.8179×0.7116=0.1923

[0127] Find the fifth root again:

[0128] F=0.1923 1 / 5 ≈0.753

[0129] The results show that the comprehensive activity score is about 0.753. When the F value is close to 1, it means that the feature maps are generally strongly activated. When the F value is lower than 0.3, it means that the overall activation level is low. When the F value is between 0.5-0.8, it means that the overall activation of the features is moderate. Subsequent screening and optimization can be carried out on the feature maps within this range.

[0130] In one example, based on the calculated comprehensive activity score, the features of the initial feature mapping data are screened and optimized to extract image features: based on the comprehensive activity score, the activation data of each feature map is then arranged in order to query the feature identification number from the data record file, these feature numbers are matched with the previously counted activation value list, and the grading standard for feature selection is consulted from the relevant parameter table. If it is necessary to determine the screening standard value such as 0.6, the typical activity cutoff value range recorded in the expert experience statistical data file is between 0.4 and 0.7, so 0.6 is selected as the parameter threshold for feature screening, and features with activation values ​​greater than 0.6 are recorded in a specific list, and features less than 0.6 are recorded in another list, so that similar feature comparisons can be performed on the features in each list later. During the comparison process, the row and column positions in the feature mapping matrix are scanned row by row, and the values ​​at these positions are compared in turn. The group with larger values ​​is marked in the record table, and these marked values ​​are associated with the feature optimization file through index matching. After performing the above operations, the final extracted image features will be obtained.

[0131] In one embodiment, the time series information of the perception layer data is analyzed through a long short-term memory network to generate time series features, including: loading the perception layer data into the long short-term memory network, adjusting the network weights through forward propagation and back propagation algorithms to obtain time-dependent feature mapping data; obtaining refined time series features based on the time-dependent feature mapping data according to a forget gate and an update gate; and analyzing and screening based on the refined time series features to obtain time series features.

[0132] Exemplarily, the time series features are obtained: the perception layer data is loaded into the long short-term memory network, the network weights are adjusted through the forward propagation and back propagation algorithms, and the time-dependent feature mapping data is obtained; based on the time-dependent feature mapping data, the forgetting gate is used to decide whether to discard or retain the information, and the update gate is used to adjust the degree of updating of the information to obtain refined time series features; based on the refined time series features, the time series features with predictive value and representativeness are analyzed and screened out.

[0133] In one example, the perception layer data is input into the long short-term memory network structure. In order to orderly process the values ​​at each time point in the input data, the learning rate and the initial value records of the weights are first extracted from the network parameter file, and the learning rate is set in the range of 0.001-0.01. Then, the corresponding data row and column values ​​are input at each time step, and passed to the network unit row by row for calculation. The output error value is compared with the expected value item by item. If the error deviation is greater than the reference value of about 0.05-0.1 calculated from historical experience, the error backtracking process is called, and the values ​​of the corresponding positions in the weight matrix are updated one by one by reading the error gradient parameter table. The update results are recorded in the weight update log file again, and the process is repeated until the error is lower than the set range. The final weight state is saved in the parameter record file, and finally the time-dependent feature mapping data is obtained.

[0134] In one example, the generated time-dependent feature mapping data is used as the input source, the forget gate threshold is set in the range of 0.3-0.4, and the update gate threshold is set in the range of 0.5-0.6, which is determined by multiple test results recorded in the historical verification file. The feature mapping value at the current time point is compared with the gating parameter, and the numerical value corresponding to each feature is compared in turn. If the difference between the numerical value and the corresponding threshold of the forget gate is less than the reference value of 0.05, the row and column position of the feature is marked in the record table and abandoned, otherwise the feature is retained, and then the update gate is compared in the same category. If the feature value deviates from the update gate interval, the corresponding value is modified in the update record table, the row and column write operation of the feature value is completed, and the updated feature information and its row and column position index are marked in a specific document. Finally, the matrix file after the processing is completed is recorded to obtain refined time series features.

[0135] In one example, using the refined time series feature set that has been obtained, the feature filtering standard range can be set between 0.4 and 0.7. This range is statistically obtained through multiple historical data comparison processes. Each feature value is traversed in rows and columns, and each feature value is compared with the interval between 0.4 and 0.7. If the feature value falls within this interval, it is marked as a potentially useful feature in the feature list. If the feature value is lower than 0.4 or higher than 0.7, the feature is recorded in another list for subsequent processing. The historical record table corresponding to the row and column position of the feature value is compared item by item through the query index file, and the feature values ​​that meet the interval conditions are copied to a new list file in a row-by-row scanning manner, and the list file is recorded in the parameter comparison table with the row and column coordinate information. After completing the entire feature determination process, a time series feature with predictive value and representativeness is obtained.

[0136] It is understandable that LSTM (Long Short-Term Memory Network) can effectively perform feature mapping and optimization on time series data through its unique gating mechanism, especially forget gate and update gate. In one example, time series feature mapping: LSTM maps the time-dependent features in the sequence to the hidden state by recursively processing each time step in the sequence data. This process involves combining the input data of the current time step with the hidden state of the previous moment to capture the time series features in the sequence. The role of the forget gate: The forget gate is the first gating mechanism in LSTM. Its role is to determine which information needs to be discarded from the memory unit and which information needs to be retained. The output of the forget gate is a value between 0 and 1, where 1 means that the information is completely retained and 0 means that it is completely discarded. The calculation process of the forget gate is as follows: The hidden state of the previous moment and the current input data are processed using the Sigmoid activation function to generate an output vector of the forget gate. By multiplying the forget gate output vector with the current state of the memory unit, an updated memory unit state is obtained, in which some information may be discarded. The role of the update gate: The update gate is the second gating mechanism in LSTM, which determines the degree of influence of the current input information on the memory unit. The update gate also uses the Sigmoid activation function to generate a value between 0 and 1 to control the degree of information update. The calculation process of the update gate is as follows: Use the Sigmoid activation function to process the hidden state of the previous moment and the current input data to generate an output vector of the update gate. Process the current input data through a tanh layer to generate a new candidate value. Multiply the output vector of the update gate with this candidate value to obtain a new updated memory unit state.

[0137] In this embodiment, through the action of forget gate and update gate, LSTM can effectively adjust and refine the time series features. The information retained in the memory unit is screened and only contains long-term dependency information that is useful for the current task. This information processing method enables LSTM to capture important features in the sequence while ignoring unimportant interference information. Through the dynamic adjustment of these gating mechanisms, LSTM can output a refined time series feature at each time step. These features can better reflect the long-term dependencies in the sequence data, thereby improving model performance.

[0138] Specifically, Figure 3 As shown in the figure, the image features and time series features are fused, and the Transformer network is used to strengthen the global learning of features to generate a fusion feature model. For example, the image features and time series features are input into the Transformer network configured with a self-attention mechanism. The Transformer network is processed through a multi-head attention mechanism, which allows the influence of other features to be considered when processing each feature, and a fusion feature model is obtained. Based on the fusion feature model, anomalies or patterns in the data are analyzed to identify potential sling defects, and preliminary defect recognition results are obtained.

[0139] In one example, image features and time series features are combined and imported into a Transformer model with a self-attention mechanism. First, feature normalization is performed to ensure that the input data is within a reasonable numerical range, such as scaling all feature values ​​to between 0 and 1. This can avoid deviations caused by excessive values ​​of certain features during model training. Next, the model processes multiple feature combinations in parallel through a multi-head attention layer and automatically adjusts the weights between features to enhance the model's understanding of the mutual influence between different features. Each feature is not only evaluated individually, but also compared and correlated with other features to fully capture the dynamic relationship between features. In this way, the model is able to generate a comprehensive feature representation that combines key information from image and time series data to form a comprehensive view of the sling state.

[0140] It can be understood that fusing image features with temporal features and using the Transformer network to strengthen the global learning of features can generate a powerful fusion feature model: the representation of image features and temporal features is represented as a vector or matrix for input into the Transformer network. Before inputting the features into the Transformer, the image features and temporal features can be fused in the following ways: directly concatenating the image feature vector and the temporal feature vector or using the outer product or inner product to interact the two features. The fused features are input into the Transformer encoder, which contains multiple self-attention layers and feedforward networks, which can capture the global dependencies between features. Among them, the self-attention mechanism in the Transformer allows the model to consider the information of all other features when processing each feature, thereby capturing the global dependencies; in order to retain the temporal information, position encoding can be added to the feature vector so that the model can consider the position information of the feature in the sequence. The Transformer network can process features of different scales by stacking multiple encoder layers. Each layer can focus on feature dependencies in different ranges, thereby achieving multi-scale feature fusion. After being processed by the Transformer encoder, the fused features will contain global dependencies and contextual information, which can be used for subsequent downstream tasks such as classification, regression, or generative models. When it is necessary to generate image or sequence outputs, a Transformer decoder can be added to generate the final fused features or prediction results. Through the above steps, the Transformer network can strengthen the global learning of features and generate a fused feature model that combines image features and time series features.

[0141] In one embodiment, sling defect detection is performed based on the fusion feature model to generate a preliminary defect recognition result.

[0142] Specifically, based on the fusion feature model, anomalies or patterns in the data are analyzed to identify potential sling defects and obtain preliminary defect identification results.

[0143] In one example, defect detection is performed based on a fusion feature model: first, safety thresholds for defect detection are set, such as setting abnormal amplitudes of vibration data outside 0.5 standard deviations, and temperatures exceeding 10% of normal operating temperatures as potential risk signals; the fusion feature model uses these safety thresholds as judgment criteria to identify data points that exceed these thresholds, which may indicate potential defects; cluster analysis is performed on these abnormal data points to determine the type and severity of the defects; in addition, data points that exceed these thresholds also use historical defect data as a reference to identify various defect types, such as cracks or wear, by comparing the current pattern with historical known defect patterns; ultimately, preliminary defect identification results are obtained.

[0144] In one embodiment, a crack extension prediction result is generated based on the preliminary defect identification result and the crack simulation result, and a safety score is performed according to the crack extension prediction result, including: combining the preliminary defect identification result and the crack simulation result to obtain a crack extension and defect identification data set; performing data analysis through a long short-term memory network based on the crack extension and defect identification data set to generate a crack extension prediction result; and performing a safety score according to the crack extension prediction result.

[0145] Specifically, Figure 3 As shown in the figure, the finite element method is used to simulate the crack propagation trajectory of the sling after being subjected to stress. Based on the preliminary defect identification results and crack simulation results, the LSTM network is used to predict the future defect path and generate crack propagation prediction results. Combined with the crack propagation prediction results, a safety score is performed to obtain a comprehensive safety score.

[0146] Exemplarily, the crack extension prediction results are obtained: finite element analysis is applied to simulate the stress of the sling, and the geometric and material models of the sling are created to simulate the forces to which the sling is subjected during use, and the mechanical response is calculated to obtain a preliminary trajectory model of crack extension; based on the preliminary trajectory model of crack extension, it is combined with the preliminary defect identification results to obtain a comprehensive crack extension and defect identification data set; based on the comprehensive crack extension and defect identification data set, a long short-term memory network is used for analysis to generate a crack extension prediction result.

[0147] In this embodiment, based on the preliminary defect identification results, mathematical simulation of crack propagation is applied to predict the change and propagation path of the crack after being subjected to stress, and the crack propagation prediction result is generated.

[0148] In one example, the geometry of a sling is refined, which involves measuring its specific dimensions such as length, width, and thickness, and recording the specific type of material and its mechanical properties, such as elastic modulus, tensile strength, and density. This data is extracted from material technical data sheets or obtained through standard material testing. Using these precise physical and material parameters, a digital 3D model of the sling is constructed. Meshing is performed and element sizes are set. Next, loading scenarios are defined, such as applied forces and fixing conditions, to simulate the maximum loads that the sling may be subjected to in actual operation. Simulations are run to calculate the stress and strain distribution of the sling under various conditions, with a special focus on potential high stress areas, which are often where cracks initiate.

[0149] In one example, after obtaining a preliminary crack trajectory model (preliminary trajectory model of crack extension), the structural stress analysis results are compared and fused with the surface defect image, and data integration is used to process these two types of data to optimize their relevance and complementarity. To this end, a data fusion framework is set up, which includes quantitative analysis of the weight and influence of each data to ensure that the model not only reflects the results of a single data source, but also provides a comprehensive view that captures the interaction between possible cracks and surface defects. Ultimately, a comprehensive crack and defect dataset (comprehensive crack extension and defect identification dataset) is formed.

[0150] In one example, using a comprehensive dataset (comprehensive crack propagation and defect identification dataset), future prediction of crack paths is performed: by configuring a long short-term memory network, inputting time series data of cracks and defects, the network learns the patterns of historical crack behavior through its internal structure. Network parameters such as learning rate and number of iterations are set to optimize the model's predictive ability. Network training includes multiple cycles of forward propagation and backpropagation, through which the network adjusts its weights to minimize the prediction error and capture the dynamics of crack propagation. After training is completed, the model is able to predict possible future crack propagation paths.

[0151] In one embodiment, a safety score is performed based on the crack extension prediction result, including: obtaining a crack severity score, a crack potential hazard value, and the total number of crack paths based on the crack extension prediction result; and calculating a safety score based on the crack severity score, the crack potential hazard value, and the total number of crack paths.

[0152] Specifically, a comprehensive safety score is obtained: based on the crack propagation prediction results, a comprehensive safety score (safety score) is calculated, and the calculation formula is:

[0153]

[0154] Among them, s i represents the severity score of the i-th crack, p i is the corresponding potential danger, m is the total number of crack paths, and ZR is the comprehensive safety score. According to the crack extension prediction results, the estimated risk of the sling is quantitatively calculated to generate a comprehensive safety score. Based on the comprehensive safety score, the overall safety of the sling is evaluated, and a maintenance or replacement plan can be formulated according to the comprehensive safety score.

[0155] In this embodiment, the calculation formula for the comprehensive safety score is beneficial in that it quantifies the comprehensive risk of each crack in the space into a comprehensive safety score in a geometric manner by simultaneously considering the crack severity score and the potential hazard score, thereby selecting the most dangerous crack path among multiple cracks, providing a quantitative basis for subsequent decision-making.

[0156] In one example, get s i Parameters: Based on the obtained comprehensive crack propagation and defect identification data set, the crack depth and surface damage ratio are converted into severity scores by statistically analyzing objective measurement data such as crack depth value (unit: mm) and surface damage ratio (unit: %) from the data files recording crack depth, surface damage degree and stress concentration area. The widely used benchmarks in reality are selected: for crack depths ranging from 1 mm to 10 mm, the greater the depth, the higher the severity. The depth value is divided by 10 to obtain a depth score of 0.1 to 1.0; for surface damage ratios ranging from 0% to 100%, the damage ratio is divided by 100 to obtain a damage score of 0 to 1.0. The two scores are then weighted averaged (the weights are dynamically adjusted according to the variation range of depth and surface damage in the actual experimental statistical data of previous times, such as the depth weight is 0.6 and the surface damage weight is 0.4. The weight value is obtained by fitting a large amount of historical comparative test data) to obtain the severity score s. i .

[0157] In one example, obtaining p i Parameters: Based on the comprehensive crack propagation and defect identification data set obtained, the crack length (unit: mm), crack position (the position score close to the edge or middle of the sling is quantified from 0 to 1.0) and propagation rate (unit: mm / day) are statistically analyzed. The length is divided by a certain reference value, such as 50 mm, to obtain a length score ranging from 0 to 1.0. The crack position is recorded as a position score according to the relative distance from the edge of the sling (0 represents the middle of the sling, and 1 represents close to the edge). The propagation rate is divided by the maximum observed rate of 20 mm / day to obtain an extension score of 0 to 1.0. The three scores are weighted according to the degree of influence on the failure of the sling in historical statistics, such as a length weight of 0.5, a position weight of 0.3, and an extension rate weight of 0.2 (all weights can be determined by fitting multiple actual load test data). The weighted average is used to obtain p i .

[0158] In one example, the m parameter is obtained: a total of 3 crack paths are recorded, then m=3.

[0159] Calculation process: Assuming there are 3 crack paths (m=3), the above method is used to obtain:

[0160] The severity score of the first crack is s1 = 0.75, and the potential hazard score is p1 = 0.90;

[0161] The severity score of the second crack is s2 = 0.65, and the potential hazard score is p2 = 0.80;

[0162] The severity score of the third crack is s3 = 0.85, and the potential hazard score is p3 = 0.70;

[0163] Substitute the values ​​into the formula for the comprehensive safety score:

[0164] For i=1:

[0165] For i=2:

[0166] For i=3:

[0167] Take the maximum value: ZR = max(1.172, 1.031, 1.101) = 1.172

[0168] The results show that the comprehensive safety score is 1.172. When the value is higher than 1.0, it indicates the existence of a high-risk crack path. A value between 0.5 and 1.0 indicates a medium risk, and a value less than 0.5 indicates a low risk. For the current step, the value of 1.172 is higher than 1.0, indicating that the comprehensive risk of the crack path is relatively large, and maintenance or replacement of the path location can be given priority in subsequent decision-making.

[0169] In one example, based on the comprehensive safety score, the score result is loaded into the parameter reference file, and the association table between the safety score and the maintenance plan is extracted from the historical data record. By browsing the position index of the row and column records of the table, the numerical range of the comprehensive safety score is found. In order to clarify the interval division, the score range is divided into several levels from 0 to 2. For example, 0.0-0.5 is used as a low safety risk interval, 0.5-1.0 is used as a medium safety risk interval, 1.0-1.5 is used as a high risk interval, and 1.5-2.0 is used as a high risk interval. These interval boundaries are determined based on a large amount of historical experimental comparison data and expert experience, and the corresponding reference values ​​are listed in the parameter file. The current comprehensive safety score is compared with these intervals. According to the above, if the current score belongs to the high or high risk range, the corresponding row and column position will be marked in the record file, such as the mark symbol "R", if it belongs to the low or medium risk range, it will be marked with the symbol "G". After scanning the rows and columns of the mark table, the maintenance plan corresponding to the record row marked with "R" will be called out from the maintenance strategy document. The maintenance strategy document is composed of the sling maintenance operation records accumulated over a long period of time, which stipulates that the corresponding score level needs to replace the sling or take local reinforcement measures for the crack path position. For the row marked with "G", lower intensity inspections and local micro-correction measures will be called. These measure options are extracted from the maintenance strategy document and recorded in the output file. Finally, the maintenance or replacement plan is determined according to the corresponding grade score mark.

[0170] It should be noted that in this embodiment, flexible sensors are introduced to collect the mechanical stress data and surface state of the sling to provide continuous real-time data for defect analysis. High-resolution images collected by industrial cameras and X-ray machines are combined with deep feature extraction of convolutional neural networks to achieve the identification of tiny defects. In addition, long short-term memory networks are used to analyze time series data to enhance the accuracy of prediction of future crack paths. The global feature learning of the Transformer network further enhances the depth of data analysis and makes risk assessment more comprehensive. This method improves the timeliness and accuracy of fault prediction and provides strong technical support for the safety management of power facilities.

[0171] This embodiment provides a defect assessment method for soft insulating slings based on image recognition, including: collecting perception layer data and high-resolution image data; performing feature extraction and feature fusion based on the perception layer data and high-resolution image data to generate a fusion feature model; performing sling defect detection based on the fusion feature model to generate a preliminary defect recognition result; simulating the crack extension trajectory of the sling after being stressed based on the finite element algorithm to obtain a crack simulation result; generating a crack extension prediction result based on the preliminary defect recognition result and the crack simulation result, and performing a safety score based on the crack extension prediction result. In this embodiment, continuous and real-time perception layer data and high-resolution image data are collected for defect analysis, and deep feature extraction of convolutional neural networks is combined to achieve the recognition of minor defects. By predicting and analyzing data, the accuracy of predicting future crack paths is enhanced. The depth of data analysis is further improved by fusion feature models, making risk assessment more comprehensive, thereby improving the timeliness and accuracy of fault prediction, and providing strong technical support for the safety management of power facilities.

[0172] In addition, an embodiment of the present invention further proposes a storage medium, on which a soft insulating sling defect assessment program based on image recognition is stored. When the soft insulating sling defect assessment program based on image recognition is executed by a processor, the steps of the soft insulating sling defect assessment method based on image recognition as described above are implemented.

[0173] Reference Figure 4 , Figure 4 The present invention is a structural block diagram of an embodiment of a soft insulating sling defect assessment system based on image recognition.

[0174] like Figure 4 As shown, the soft insulating sling defect assessment system based on image recognition includes:

[0175] A data acquisition module 10, used to acquire perception layer data and high-resolution image data;

[0176] A feature extraction module 20, used for performing feature extraction and feature fusion according to the perception layer data and the high-resolution image data to generate a fusion feature model;

[0177] A defect detection module 30, used to perform sling defect detection according to the fusion feature model and generate a preliminary defect recognition result;

[0178] The crack simulation module 40 is used to simulate the crack extension trajectory of the sling after being stressed based on the finite element algorithm to obtain the crack simulation result;

[0179] The safety assessment module 50 is used to generate a crack propagation prediction result based on the preliminary defect identification result and the crack simulation result, and to perform a safety score according to the crack propagation prediction result.

[0180] Specifically, the data acquisition module 10 collects mechanical stress data and surface state data by embedding flexible sensors in the sling, and cooperates with industrial cameras and X-ray machines to adjust the light source and exposure, collect RGB images, infrared images and X-ray images, and generate perception layer data and high-resolution image data.

[0181] Specifically, the feature extraction module 20 performs feature point marking and extraction on the high-resolution image data to generate image features, and generates time series features by calculating the time series differences in the perception layer data, compares and matches the image features with the time series features, and generates a feature fusion model.

[0182] Specifically, the defect detection module 30 may include a feature enhancement module, which performs feature enhancement processing based on a feature fusion model, generates an enhanced feature model, locates potential defects of the sling, and generates a preliminary defect identification result.

[0183] Specifically, the crack simulation module 40 may include a simulation and prediction module, which receives preliminary defect identification results, applies mathematical simulation of crack extension, predicts changes and extension paths of cracks after being subjected to stress, and generates crack extension prediction results.

[0184] Specifically, the safety assessment module 50 performs quantitative calculations on the estimated risks of the sling according to the crack extension prediction results, and generates a comprehensive safety score.

[0185] It should be noted that by introducing flexible sensors to collect the mechanical stress data and surface state of the sling, continuous real-time data is provided for defect analysis. High-resolution images collected by industrial cameras and X-ray machines are combined with deep feature extraction of convolutional neural networks to realize the identification of tiny defects. In addition, long short-term memory networks are used to analyze time series data to enhance the accuracy of prediction of future crack paths. The global feature learning of the Transformer network further enhances the depth of data analysis and makes risk assessment more comprehensive. The above methods improve the timeliness and accuracy of fault prediction and provide strong technical support for the safe management of power facilities.

[0186] This embodiment provides a soft insulating sling defect assessment system based on image recognition. In this embodiment, continuous and real-time perception layer data and high-resolution image data are collected for defect analysis, and deep feature extraction of convolutional neural networks is combined to achieve the recognition of tiny defects. By predicting and analyzing data, the accuracy of prediction of future crack paths is enhanced. By fusing feature models, the depth of data analysis is further improved, making risk assessment more comprehensive, thereby improving the timeliness and accuracy of fault prediction, and providing strong technical support for the safety management of power facilities.

[0187] It should be noted that the technical details that are not described in detail in the embodiment of the soft insulating sling defect assessment system based on image recognition can be found in the soft insulating sling defect assessment method based on image recognition as described above provided in any embodiment of the present invention, and will not be repeated here.

[0188] It should be understood that the above is only an example and does not constitute any limitation on the technical solution of the present invention. In specific applications, technicians in this field can make settings as needed, and the present invention does not limit this.

[0189] It should be noted that the workflow described above is merely illustrative and does not limit the scope of protection of the present invention. In practical applications, technicians in this field can select part or all of them according to actual needs to achieve the purpose of the present embodiment, and no limitation is made here.

[0190] In addition, it should be noted that, in this article, the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, article or system including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or system. In the absence of further restrictions, an element defined by the sentence "comprises a ..." does not exclude the existence of other identical elements in the process, method, article or system including the element.

[0191] The serial numbers of the above embodiments of the present invention are only for description and do not represent the advantages or disadvantages of the embodiments.

[0192] Through the description of the above implementation methods, those skilled in the art can clearly understand that the above-mentioned embodiment methods can be implemented by means of software plus a necessary general hardware platform, and of course by hardware, but in many cases the former is a better implementation method. Based on such an understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product, which is stored in a storage medium (such as a read-only memory (ROM) / RAM, a magnetic disk, or an optical disk), and includes a number of instructions for a terminal device (which can be a mobile phone, a computer, a server, or a network device, etc.) to execute the methods described in each embodiment of the present invention.

[0193] The above are only preferred embodiments of the present invention, and are not intended to limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made using the contents of the present invention specification and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.

Claims

1. A soft insulating sling defect assessment method based on image recognition, characterized in that: include: Collect perception layer data and high-resolution image data; Performing feature extraction and feature fusion according to the perception layer data and the high-resolution image data to generate a fusion feature model; Perform sling defect detection according to the fusion feature model to generate a preliminary defect recognition result; Based on the finite element algorithm, the crack propagation trajectory of the sling after being stressed is simulated to obtain the crack simulation results; A crack propagation prediction result is generated based on the preliminary defect identification result and the crack simulation result, and a safety score is performed according to the crack propagation prediction result.

2. The method according to claim 1, characterized in that The collecting of perception layer data and high-resolution image data includes: Embedding a flexible sensor in a soft insulating sling, and monitoring the mechanical stress and surface state of the sling in real time based on the flexible sensor to collect mechanical stress data and surface state data; Using the mechanical stress data and the surface state data as original monitoring data; Calculating the weighted accuracy of the original monitoring data; Filtering the original monitoring data based on the weighted accuracy, and generating perception layer data according to the filtered original monitoring data; High-resolution image data is acquired based on image acquisition equipment.

3. The method according to claim 2, characterized in that The method of acquiring high-resolution image data based on an image acquisition device includes: Based on the image acquisition device, an active light source and a dynamic exposure algorithm are used to acquire an initial RGB image, an initial IR image and an initial X-ray image; Calculating a comprehensive contrast enhancement index of an image based on the initial RGB image, the initial IR image, and the initial X-ray image; The initial RGB image, the initial IR image and the initial X-ray image are subjected to image enhancement processing based on the comprehensive contrast enhancement index of the image to obtain high-resolution image data.

4. The method according to claim 1, characterized in that The extracting and fusing features according to the perception layer data and the high-resolution image data to generate a fusion feature model includes: Processing the high-resolution image data through a convolutional neural network to extract image features; Analyzing the time series information of the perception layer data through a long short-term memory network to generate time series features; The image features and time series features are input into a Transformer network configured with a self-attention mechanism to generate a fusion feature model.

5. The method according to claim 4, characterized in that The step of processing the high-resolution image data by a convolutional neural network to extract image features includes: Inputting the high-resolution image data into a convolutional neural network for filtering and feature learning to obtain initial feature mapping data; Calculating a comprehensive activity score of a feature based on the initial feature mapping data; The initial feature mapping data is screened and optimized according to the comprehensive activity score to extract image features.

6. The method according to claim 4, characterized in that The analyzing the time series information of the perception layer data by the long short-term memory network to generate time series features includes: Loading the perception layer data into the long short-term memory network, adjusting the network weights through forward propagation and back propagation algorithms, and obtaining time-dependent feature mapping data; Obtaining refined temporal features based on the time-dependent feature mapping data according to a forget gate and an update gate; Analysis and screening are performed based on the refined time series features to obtain time series features.

7. The method according to any one of claims 1 to 6, characterized in that The generating a crack propagation prediction result based on the preliminary defect identification result and the crack simulation result, and performing a safety score according to the crack propagation prediction result, comprises: Combining the preliminary defect identification result with the crack simulation result to obtain a crack extension and defect identification data set; Based on the crack propagation and defect identification data set, data analysis is performed through a long short-term memory network to generate a crack propagation prediction result; A safety score is performed based on the crack extension prediction result.

8. The method according to claim 7, characterized in that The performing of safety scoring according to the crack extension prediction result comprises: Obtaining a severity score of a crack, a potential hazard value of a crack, and a total number of crack paths according to the crack propagation prediction result; A safety score is calculated based on the severity score of the crack, the potential hazard value of the crack, and the total number of crack paths.

9. A soft insulating sling defect assessment system based on image recognition, characterized in that: include: Data acquisition module, used to collect perception layer data and high-resolution image data; A feature extraction module, used for performing feature extraction and feature fusion according to the perception layer data and the high-resolution image data to generate a fusion feature model; A defect detection module, used to perform sling defect detection according to the fusion feature model and generate a preliminary defect recognition result; The crack simulation module is used to simulate the crack extension trajectory of the sling after being stressed based on the finite element algorithm to obtain the crack simulation results; A safety assessment module is used to generate a crack extension prediction result based on the preliminary defect identification result and the crack simulation result, and to perform a safety score according to the crack extension prediction result.

10. An electronic device, characterized in that: The electronic device includes: a memory, a processor, and a soft insulating sling defect assessment program based on image recognition stored in the memory and executable on the processor, wherein the soft insulating sling defect assessment program based on image recognition is configured to implement the soft insulating sling defect assessment method based on image recognition as described in any one of claims 1 to 8.

Images