Insulator zero value detection method and system fusing main shaft positioning and line interest pooling

Abstract

The invention provides an insulator zero value detection method and system fusing main shaft positioning and line interest pooling, and belongs to the technical field of insulator detection. According to the method, in combination with an infrared thermal imaging image obtained by unmanned aerial vehicle inspection, main tandem axis positioning (SAL), line interest pooling (LOI-Pooling) and pointer type sequence decoding (PSH) are introduced into a YOLOv8 framework, paired thermal relation coding (PRT) and rhythm consistency (RCH) are assisted, weighted fusion with confidence is carried out to form a comprehensive energy index, and according to the comprehensive energy index, the optimal energy index is obtained. And adopting a window voting mechanism with confidence coefficient smoothness and high and low threshold hysteresis to output a judgment result of whether a zero-value insulator exists, and positioning the specific piece number of the fault insulator and the confidence coefficient of the fault insulator. According to the method, high-precision and strong-robustness detection can be carried out on the zero-value insulator on the premise of not depending on electric parameters.

Description

Technical Field

[0001] This invention relates to the field of insulator testing technology, and in particular to a method and system for detecting zero values ​​in insulators that integrates spindle positioning and line interest pooling. Background Technology

[0002] Insulators, as a crucial component of power transmission lines, primarily function to reliably connect conductors to towers or crossarms and provide electrical insulation between conductors and grounding electrodes. With the rapid development of power grids and the continuous increase in voltage levels, long-distance power transmission across regions and provinces has become commonplace, significantly increasing the number and length of insulator strings and making operational scenarios more complex. The operational status of insulators directly affects the safety and stability of the power system; failure can easily lead to line tripping or even large-scale power outages.

[0003] During long-term operation, insulators are continuously subjected to high voltage and strong electric fields, and are affected by external environmental factors such as wind, sand, rain, snow, ice, and temperature differences, which may lead to insulation degradation, mechanical damage, and electrical failure. Zero-value insulators are the most insidious and pose the most serious threat: when an insulator completely loses its insulating capacity due to damage, puncture, or internal breakdown, it is electrically equivalent to a conductor, no longer sharing voltage, resulting in a severe imbalance in the voltage distribution of the entire string. Other parts experience overvoltage and may experience localized overheating and flashover. If not detected and addressed in time, this can easily trigger a cascading failure. Therefore, zero-value detection is a crucial aspect of condition-based maintenance and operation of transmission lines.

[0004] Existing zero-value detection technologies mainly include manual inspection, live-line testing, and infrared thermal imaging interpretation. Manual inspection relies heavily on personnel experience, is inefficient, and is limited by working at heights and viewing angles. Live-line testing has high accuracy but often requires specialized equipment or working windows, resulting in high on-site implementation costs. Infrared thermal imaging is suitable for long-distance, non-contact inspection and has become the main method in recent years. However, existing detection schemes still have limitations in modeling and data supply when relying solely on infrared images. Specifically, the geometric / thermal mismatch introduced by environmental and viewing angle changes is not adequately compensated; there is a heavy reliance on thresholds and empirical rules, making it difficult to scale to a large number of scenarios; scarce anomalous samples limit the learning of discrimination boundaries; and there is insufficient utilization of tandem structures and inter-chip couplings, lacking mechanism-oriented explanations and constraints. Therefore, how to achieve structured alignment, sequential modeling, and physical prior constraints based on infrared images has become the key to the engineering application of zero-value detection.

[0005] In recent years, deep learning has been applied in infrared image analysis. Convolutional neural networks can automatically extract features and achieve higher accuracy than traditional methods in some scenarios. However, three bottlenecks still exist: First, it is highly dependent on large-scale labeled data, while real zero-value samples are difficult to collect, and the generalization ability is reduced due to inter-domain differences; second, it lacks physical prior constraints and structural cognition, and simple two-dimensional box detection is easily sensitive to shooting angle, scale and occlusion, making it difficult to stably output film number-level conclusions; third, the network features have strong "black box" properties, making it difficult to provide an interpretable basis for the underlying mechanism. Summary of the Invention

[0006] Given the limitations of existing technologies in zero-value insulator detection, this invention provides a method for zero-value insulator detection that integrates main axis positioning and line interest pooling. Based on the YOLOv8 framework, this method introduces main axis positioning (SAL), line interest pooling (LOI-Pooling), and pointer-type sequence decoding (PSH) modules, combined with pairwise thermal relation coding (PRT) and rhythm consistency (RCH). Through the synergistic effect of these modules, and utilizing energy fusion and hysteresis mechanisms (EFP), accurate detection of zero-value insulators is achieved.

[0007] The technical solution adopted by the embodiments of the present invention to solve its technical problem is as follows:

[0008] A method for detecting zero values ​​in insulators that integrates spindle positioning and line interest pooling includes:

[0009] Step S1: Acquire infrared thermal imaging images of the insulator string, input them into the YOLOv8 network of the parallel main string shaft positioning head SAL, and predict the axial thermal map, the position of the two end points, and the axial sequence direction from the high voltage side to the grounding side of the insulator string.

[0010] In step S2, the backbone and neck network of YOLOv8 extract multi-scale features from the infrared thermal imaging image, fuse and reduce the dimensions to obtain a fused feature map.

[0011] Step S3: Use the principal axis coordinate system to perform line interest pooling on the fused feature map, align the two-dimensional observations and reduce the dimension to a one-dimensional strip sequence;

[0012] Step S4: Using Paired Thermal Relation Code (PRT), radial inner and outer bands and left and right neighbor sampling are performed on candidate slices on the strip to construct four relative thermal features: difference, ratio, derivative, and axial gradient, generating a relative thermal anomaly score; using Rhythm Consistency Code (RCH), autocorrelation frequency domain analysis is performed on the one-dimensional strip sequence to measure whether the near-equidistant rhythm of the umbrella skirt is disrupted, generating a rhythm consistency index; the relative thermal anomaly score, rhythm consistency index, and one-dimensional strip features are concatenated to obtain the prior-bearing strip sequence.

[0013] Step S5: The prior-blessed strip sequence is input into the pointer-type sequence decoder PSH for encoding. The sequence context is encoded through one-dimensional convolution and recurrent units. Under monotonic constraints, the slice position result is output using pointer decoding and back-projected into the image plane candidate box. The slice position result includes confidence.

[0014] Step S6, Anomaly Detection and Decision Fusion (EFP): The relative thermal anomaly score, rhythm consistency index, and confidence level are weighted and fused into a comprehensive energy index. Based on the comprehensive energy index, a window voting mechanism with confidence level smoothing and high and low threshold hysteresis is adopted to output the determination result of whether there is a zero-value insulator, and to locate the specific piece number of the faulty insulator and its confidence level.

[0015] Preferably, step S1 includes:

[0016] Step S11: An infrared thermal imager mounted on a drone or ground platform is used to inspect the insulator strings of the transmission line, acquire single-frame or short-sequence infrared thermal imaging images, and perform image processing using a gain-bias model.

[0017] Step S12: Connect the master axis positioning head in parallel to the backbone and neck features {P3,P4,P5} of YOLOv8 to regress the master axis parameters and endpoints:

[0018] ;

[0019] In the formula, p(t) represents the coordinates of a point on the principal axis; c = (c x ,c y ) is the main axis center; v = (v x ,v y ) is the main axis unit direction; t is the axis parameter; S is the axis length; ||·||2 is the L2 norm; the comprehensive supervision is:

[0020] ;

[0021] In the formula, L axis is the axis localization loss; Focal(·) is the focus loss; H is the predicted heatmap output from the model axis localization branch and normalized by the Sigmoid function; This is a Gaussian-annotated heatmap generated by projecting the center position of the insulator discs along the principal axis from the annotated sample. These are the predicted and labeled values ​​for both endpoints; The predicted and labeled values ​​of the direction vectors in axis order; "·" represents the dot product; λ e , λ o , λ L is the loss weight; ||·||1 is the L1 norm.

[0022] Preferably, step S2 includes:

[0023] Step S21: Extract multi-scale features of the infrared thermal imaging image within the axial window along the axial sequence and perform normalization processing, including temperature statistical features, temperature difference features, and texture statistical features:

[0024] Temperature statistical characteristics include average temperature, peak temperature, minimum temperature, and temperature standard deviation. ;

[0025] In the formula: T k,t The temperature value of the t-th pixel in the region of the k-th insulator; n i Let be the total number of pixels in the region of the k-th insulator. Let be the average temperature of the region of the k-th insulator; These represent the peak and minimum temperature values ​​in the region of the k-th insulator. The degree of dispersion of the temperature distribution in the region of the k-th insulator;

[0026] Temperature difference characteristics include temperature difference between adjacent sections and peak temperature difference: ;

[0027] In the formula: △T k,k+1 The average temperature difference between two adjacent sections; The difference between the peak temperatures of two adjacent plates; The average temperature of adjacent plates; , The peak temperature of adjacent plates;

[0028] Texture statistical features include contrast, energy, entropy, and correlation. ;

[0029] In the formula: P(gi,gj) is the co-occurrence probability of pixel gray values ​​gi and gj; μ gi μ gj σ is the mean of the grayscale distribution; gi σ gj is the standard deviation of grayscale distribution; Con is the contrast; Energy is the energy value reflecting the uniformity of texture distribution; Entropy is the entropy representing texture complexity; Cor is the correlation.

[0030] All extracted features were normalized using the Z-score normalization method;

[0031] Step S22: Upsample, align, and channel-wise reduce the multi-scale features to obtain the fused feature map F. fuse : ;

[0032] In the formula: F (l) W represents the feature of the l-th layer, where l ∈ [1, L], and L is the number of feature layers involved in the fusion. (l) To align convolutions; This is the upsampling factor; For channel splicing; Ψ represents first-order dimensionality reduction; The convolution is performed; the temperature statistical features, temperature difference features, and texture statistical features are used as auxiliary feature channel inputs, and are combined with YOLOv8's multi-scale features F. (l) Channel splicing or weighted fusion is performed to obtain a joint feature map.

[0033] Preferably, step S3 includes:

[0034] Step S31, based on the parameters of the principal axis, in axis-normal coordinates, fuse the feature map F fuse Construct differentiable sampling bands along the principal axis and radial direction, specifying the sampling location along the axis and with lateral offset:

[0035] ;

[0036] In the formula: c(s) is the point at arc length s on the axis; s is the arc length position along the principal axis; P0 is the axis starting point; q(s,r) is the sampling point at s with a normal offset of r; r is the normal displacement; n is the normal unit vector; S is the axis length; R is the set of discrete normal radii;

[0037] Step S32, after determining the sampling point, from F fuse Read out the corresponding features of the sampling points and perform weighted summation within the band to obtain the one-dimensional representation z(s) of each s: ; ;

[0038] In the formula: Let be the vector read from the fused features at position q(s,r); BL-sample is the bilinear interpolation sampling function; z(s) is the one-dimensional strip feature; W r The learnable weights are the values ​​corresponding to the radius r; exp(·) is the exponential function; α r (s) represents the weights after Softmax normalization.

[0039] Preferably, step S4 includes:

[0040] Step S41: Use Paired Thermal Relationship (PRT) to analyze the relative thermal anomalies of candidate chip k with the inner / outer zone and neighboring chips to obtain the thermal anomaly score A. PRT (s k ): ;

[0041] In the formula: s k R represents the axial position of the center of the k-th candidate slice; in R out Let R be the set of inner and outer zone radii; in |、|R out | represents the number of elements in the set; I(q) represents the normalized temperature at position q. It is the k-th candidate slice along the principal axis, within the inner and outer zone radius set R. in R out All sampling points; μ in μ out The average temperature is the temperature of the inner and outer zones; ε is a very small positive number. For in s k The derivative along the axis at δ; s Adjacent slice spacing; φ(s) k ) represents the relative thermal feature vector; w represents the linear scoring parameter; σ(·) represents the Sigmoid function; the four relative thermal features—difference, ratio, derivative, and axial gradient—are μ, respectively. in -μ out , , μ in (s k )-μ in (s k ±δ s );

[0042] Step S42: Perform autocorrelation / frequency domain analysis on the one-dimensional strip sequence using rhythm consistency RCH to obtain the rhythm consistency index sequence C. RCH : ;

[0043] In the formula: x(s) is the one-dimensional strip sequence response; R(τ) represents the mean; T represents the sequence length; R(τ) represents the autocorrelation. The desired pitch; max τ To maximize τ; ε is a minimum positive number; T-τ is the effective number of samples calculated by autocorrelation; This corresponds to the expected pitch autocorrelation value;

[0044] Step S43, assign the relative thermal anomaly score A PRT (s k ), rhythm consistency index sequence C RCH By splicing the data to one-dimensional strip features, a priori-blessed strip sequence is obtained.

[0045] Preferably, step S5 includes:

[0046] Step S51: Input the prior-bearing stripe sequence into the pointer-type sequence decoder PSH, and encode the sequence context through one-dimensional convolution and recurrent unit to obtain the encoded sequence H. enc :

[0047] H enc =BiGRU(ConvlD(Z||A PRT (s)||C RCH ))

[0048] In the formula: Z is the prior-bearing strip sequence; A PRT (s) is a paired thermal relation score sequence, A PRT (s)={A PRT (s k )};C RCH The sequence represents the rhythm consistency index; BiGRU represents a bidirectional gated cycle.

[0049] Step S52: Use pointer attention and monotonic masking to score and obtain the score e. k,s : ;

[0050] In the formula: exp(·) is the exponential function; s is the position of the candidate point on the principal axis; h k-1 The previous state; W q W k These are the query and key mapping matrices, respectively; v is a vector; tanh is the hyperbolic tangent; M k(s) For monotonic masks; π k(s) s' is the Softmax probability; T is the total length of the encoded sequence; s' is the index of the Softmax normalized summation.

[0051] Step S53: Process the image using position, width, confidence level, and reprojection information, and backproject the image into a candidate bounding box based on the image position result. ;

[0052] In the formula: Centered on the axis; Width of axis; For confidence; u, g, b u b g The header parameter is ReLU; ReLU is the activation function. is the center of the axial coordinate system; n is the normal unit vector; Proj is the projection function; Candidate bounding boxes for the image plane;

[0053] Step S54: During model training, sequence supervision is used in conjunction with the native YOLOv8 detection head for joint optimization. ; ;

[0054] In the formula, Let λ be the true position, true width, and confidence label of the k-th slice, respectively; ctr , λ w , λ conf , λ mono For loss weights; BCE is the binary cross-entropy; δ is the monotonic margin; max(0,·) is the ReLU form of monotonic constraint; L yolo The native detection loss for YOLOv8 includes bounding box regression loss, object confidence loss, and classification loss; L PRT For paired thermal relationship losses, used to constrain inter-plate thermal symmetry characteristics; L RCH Rhythm consistency loss is used to measure the periodicity and regularity of strip sequences; L joint The total loss is λ. prt , λ rch For the two prior loss weights;

[0055] Monotonic pointer decoding is used to achieve segment number-level structured output, and joint training with two-dimensional detection ensures sequence geometric consistency, thereby obtaining segment number-level detection results with continuous positions and ordered sequences.

[0056] Preferably, step S6 includes:

[0057] Step S61: The relative thermal anomaly score, rhythm consistency index, and detection confidence are weighted and fused to construct a comprehensive energy index E. k : ; ;

[0058] In the formula: E k λ represents the original fusion energy of the k-th fragment. a , λ c , λ p Weights for the three-path fusion; To test the confidence level; γ is the energy after EMA smoothing; γ is the EMA coefficient;

[0059] Step S62: Using hysteresis, window voting, and probability mapping, output the determination result of whether a zero-value insulator exists, as well as the final determination of the zero-value insulator and the piece number-level confidence level. ; ;

[0060] In the formula: T high T lowy represents the hysteresis threshold; 1[·] represents the indicator function; W represents the window length; k To determine whether the k-th slice is an anomaly; Score k α represents the confidence level of the final film number; α and β are the linear mapping parameters between energy and probability.

[0061] An insulator zero-value detection system integrating spindle positioning and line interest pooling, for implementing the aforementioned method, includes:

[0062] The acquisition module is used to acquire infrared thermal imaging images of the insulator string;

[0063] The prediction module uses the YOLOv8 network of the parallel main string shaft positioning head SAL to predict the axial thermal map of the insulator string, the positions of the two endpoints, and the axial sequence direction from the high voltage side to the grounding side.

[0064] The extraction module is used to extract multi-scale features from infrared thermal imaging images from the backbone and neck network of YOLOv8, perform fusion and dimensionality reduction, and obtain a fused feature map.

[0065] The line interest pooling module is used to perform line interest pooling on the fused feature map using the principal axis coordinate system, aligning the two-dimensional observations and reducing them to a one-dimensional strip sequence.

[0066] The prior module is used to sample candidate slices radially within the inner and outer bands and left and right neighborhoods on the strip using Paired Thermal Relation Code (PRT), constructing four relative thermal features: difference, ratio, derivative, and axial gradient, and generating a relative thermal anomaly score. Rhythm Consistency (RCH) is used to perform autocorrelation frequency domain analysis on the one-dimensional strip sequence to measure whether the near-equidistant rhythm of the umbrella skirt is disrupted, generating a rhythm consistency index. The relative thermal anomaly score, rhythm consistency index, and one-dimensional strip features are concatenated to obtain the prior-enhanced strip sequence.

[0067] The slice output module is used to encode the prior-enhanced strip sequence into the pointer-type sequence decoder PSH, encode the sequence context through one-dimensional convolution and recurrent units, and output the slice result using pointer decoding under monotonic constraints. The back projection is used to generate candidate boxes for the image plane. The slice result includes confidence.

[0068] The judgment module is used for anomaly detection and decision fusion (EFP): it weights and fuses relative thermal anomaly scores, rhythm consistency indicators and confidence levels into a comprehensive energy index. Based on the comprehensive energy index, it adopts a window voting mechanism with confidence level smoothing and high and low threshold hysteresis to output the judgment result of whether there is a zero-value insulator, and locates the specific segment number of the faulty insulator and its confidence level.

[0069] As can be seen from the above technical solution, the insulator zero-value detection method and system integrating main axis positioning and line interest pooling provided by the embodiments of the present invention are used for intelligent detection and positioning of zero-value insulators in transmission lines. This method combines infrared thermal imaging images obtained from UAV inspections, introduces main axis positioning (SAL), line interest pooling (LOI-Pooling), and pointer-type sequence decoding (PSH) into the YOLOv8 framework, and supplements it with pairwise thermal relation coding (PRT) and rhythm consistency (RCH), outputting the zero-value chip number and confidence level through energy fusion and hysteresis (EFP). Specifically, the YOLOv8 backbone and neck network extracts multi-scale features from infrared thermal imaging images, performs fusion and dimensionality reduction to obtain a fused feature map; line interest pooling is applied to the fused feature map using the principal axis coordinate system to align the two-dimensional observations and reduce them to a one-dimensional strip sequence; pairwise thermal relation coding (PRT) is used to sample candidate strips radially in the inner and outer bands and left and right neighbors on the strips, constructing four relative thermal features: difference, ratio, derivative, and axial gradient, generating a relative thermal anomaly score; autocorrelation frequency domain analysis is performed on the one-dimensional strip sequence using rhythm consistency (RCH) to measure whether the near-isotral rhythm of the umbrella skirt is disrupted, generating a rhythm consistency index; the relative thermal anomaly score and rhythm... The consistency index is concatenated with one-dimensional strip features to obtain a priori-bearing strip sequence. This priori-bearing strip sequence is then input into a pointer-type sequence decoder (PSH) for encoding. The sequence context is encoded through one-dimensional convolution and recurrent units, and the slice position result is output using pointer decoding under monotonic constraints. This result is then back-projected into a candidate bounding box in the image plane. Anomaly detection and decision fusion (EFP) involves weighting and fusing relative thermal anomaly scores, rhythm consistency indices, and confidence levels into a comprehensive energy index. Based on this comprehensive energy index, a window voting mechanism using confidence level smoothing and high / low threshold hysteresis is employed to output the determination of whether a zero-value insulator exists, and to locate the specific slice number of the faulty insulator and its confidence level. This invention enables high-precision and robust detection of zero-value insulators without relying on electrical parameters. Attached Figure Description

[0070] Figure 1 A flowchart for an insulator zero-value detection method that integrates spindle positioning and line interest pooling.

[0071] Figure 2 This is a schematic diagram of infrared image detection and feature processing for insulators.

[0072] Figure 3 Clipping the ROI.

[0073] Figure 4 This is a piecewise linear graph showing the statistics of insulator segments and the differences between adjacent segments.

[0074] Figure 5 This is a schematic diagram of a one-dimensional curve of the LOI stripe.

[0075] Figure 6 This is a plot of the E_EMA distribution of the slice's overall energy.

[0076] Figure 7 This is a comparison chart of temperature / priority / fusion curves. Detailed Implementation

[0077] The technical solution and effects of the present invention will be further described in detail below with reference to the accompanying drawings.

[0078] This invention proposes a method and system for detecting zero-value insulators that integrates main axis positioning and line interest pooling, for intelligent detection and positioning of zero-value insulators in transmission lines. This method combines infrared thermal imaging images acquired by UAV inspections, introduces main axis positioning (SAL), line interest pooling (LOI-Pooling), and pointer-type sequence decoding (PSH) into the YOLOv8 framework, and supplements it with pairwise thermal relation coding (PRT) and rhythm consistency coding (RCH). Through energy fusion and hysteresis (EFP), the zero-value piece number and confidence level are output, achieving high-precision and robust detection of zero-value insulators without relying on electrical parameters. To achieve the above objectives, this invention discloses relevant technical solutions, which are described in detail below with reference to flowchart illustrations:

[0079] like Figure 1 As shown, the insulator zero-value detection method integrating spindle positioning and line interest pooling proposed in this invention mainly includes the following steps:

[0080] Step S1: Insulator strings are inspected by using an infrared thermal imager mounted on a drone or ground platform. Infrared thermal images of the insulator strings are collected and input into the YOLOv8 network of the parallel main string axis positioning head SAL to predict the axial thermal map of the insulator strings, the positions of the two end points, and the axis sequence direction from the high voltage side to the grounding side, providing a basis for positioning and sorting for subsequent axis-based modeling.

[0081] In step S2, the backbone and neck network of YOLOv8 extract multi-scale features from infrared thermal imaging images, fuse and reduce their dimensions to obtain a fused feature map for subsequent modeling; at the same time, the infrared amplitude is relativized and standardized by combining the captured metadata and image statistics to reduce the impact of differences in ambient temperature, irradiance and viewing angle, and improve cross-scene consistency.

[0082] Step S3: Apply line interest pooling to the fused feature map using the principal axis coordinate system to align and reduce the dimensionality of the two-dimensional observations into a one-dimensional strip sequence. Based on the principal axis parameters output by SAL, construct differentiable sampling bands along the principal axis and radial direction on the fused feature map, and perform line interest pooling (LOI-Pooling) to achieve feature alignment, generating a one-dimensional strip feature sequence along the serial axis. This strip reduces the dimensionality of the two-dimensional observations into a one-dimensional sequence, effectively suppressing the effects of oblique shots and scale changes, and providing stable input for serialization decoding.

[0083] Step S4: Using Paired Thermal Relation Code (PRT), radial inner and outer bands and left and right neighbor sampling are performed on candidate strips on the strip to construct four relative thermal features: difference, ratio, derivative, and axial gradient, generating a relative thermal anomaly score. Rhythm Consistency (RCH) is used to perform autocorrelation frequency domain analysis on the one-dimensional strip sequence to measure whether the near-equidistant rhythm of the umbrella skirt is disrupted, generating a rhythm consistency index. The relative thermal anomaly score, rhythm consistency index, and one-dimensional strip features are concatenated to obtain a priori-enhanced strip sequence. The above priors are incorporated into the model in the form of trainable branches and constraint terms to improve the stability and interpretability of the results.

[0084] Step S5: The prior-blessed strip sequence is input into the pointer-type sequence decoder PSH for encoding. The sequence context is encoded through one-dimensional convolution and recurrent units. Under monotonic constraints, the slice position result is output using pointer decoding and back-projected into the image plane candidate box. The slice position result includes confidence.

[0085] Step S6, Anomaly Detection and Decision Fusion (EFP): The relative thermal anomaly score, rhythm consistency index, and confidence level are weighted and fused into a comprehensive energy index. Based on the comprehensive energy index, a window voting mechanism with confidence level smoothing and high and low threshold hysteresis is adopted to output the determination result of whether there is a zero-value insulator, and to locate the specific piece number of the faulty insulator and its confidence level.

[0086] The specific implementation of step S1 includes:

[0087] Step S11: An infrared thermal imager mounted on a drone or ground platform inspects the insulator strings of the transmission line to acquire single-frame or short-sequence infrared thermal imaging images, and performs image processing using a gain-bias model.

[0088] Because infrared radiation intensity varies with the fourth power of temperature, even a small temperature difference can create observable radiation differences. Infrared images naturally contain relative thermal clues reflecting the operating state of insulators; infrared radiation approximately follows the Stefan-Boltzmann law.

[0089] E=òσT 4 (1)

[0090] In formula (1): ω is the radiant power per unit area; ω is the surface emissivity (dimensionless, 0-1); σ is the Stefan-Boltzmann constant; Let be the absolute surface temperature. Considering the systematic deviations caused by different devices and imaging geometry, the imaging response can be written as a gain-bias model: (2)

[0091] In formula (2): Pixel response; For system gain; For local emissivity; Local temperature; It is an additive bias; These are pixel coordinates;

[0092] Step S12: Input the image into the YOLOv8 backbone and neck network. Connect the main string axis positioning head (SAL) in parallel to the YOLOv8 backbone and neck features {P3,P4,P5} to regress the main axis parameters and endpoints, automatically predict the axial thermal map and endpoint positions of the insulator string, and determine the sequential direction of "high voltage side → grounding side", providing a positioning and sorting basis for subsequent axial modeling. To reduce geometric mismatches in two-dimensional piece-by-piece search, connect the main string axis positioning head (SAL) in parallel to the YOLOv8 backbone and neck features {P3,P4,P5} to regress the main axis parameters and endpoints. (3)

[0093] In equation (3), p(t) represents the coordinates of a point on the principal axis; c = (c x ,c y ) is the main axis center; v = (v x ,v y ) is the main axis unit direction; t is the axis parameter; S is the axis length; ||·||2 is the L2 norm; the comprehensive supervision is: (4)

[0094] In equation (4), L axis is the axis localization loss; Focal(·) is the focus loss; H is the predicted heatmap output from the model axis localization branch and normalized by the Sigmoid function; This is a Gaussian-annotated heatmap generated by projecting the center position of the insulator discs along the principal axis from the annotated sample. These are the predicted and labeled values ​​for both endpoints; The predicted and labeled values ​​of the direction vectors in axis order; "·" represents the dot product; λ e , λ o , λ L is the loss weight; ||·||1 is the L1 norm.

[0095] Step S1 achieves effective localization from infrared images to single-string principal axis geometry. The lightweight detection head ensures real-time performance in scenarios where UAV computing power is limited, while the axial heatmap and endpoint / direction regression provide high-quality input and directed sequence benchmarks for subsequent feature extraction and graph structure modeling, thus laying a solid foundation for the overall detection framework.

[0096] Step S2 performs cross-layer fusion and dimensionality reduction on the multi-scale features of YOLOv8 to obtain a fused feature map with uniform resolution; simultaneously, it extracts temperature statistical features, difference features, and texture statistical features within the slice / axial window, and uses Z-score normalization to eliminate the influence of different dimensions and environmental conditions. Specific implementation includes:

[0097] Step S21: Extract multi-scale features of the infrared thermal imaging image within the axial window along the axial sequence and perform normalization processing, including temperature statistical features, temperature difference features, and texture statistical features:

[0098] Temperature statistical characteristics include average temperature, peak temperature, minimum temperature, and temperature standard deviation. (5)

[0099] In the formula: T k,t The temperature value of the t-th pixel in the region of the k-th insulator; n i Let be the total number of pixels in the region of the k-th insulator. Let be the average temperature of the region of the k-th insulator; These represent the peak and minimum temperature values ​​in the region of the k-th insulator. denoted as the dispersion of the temperature distribution in the region of the k-th insulator; where k∈[1,N] and N is the total number of insulators in the string.

[0100] Temperature difference characteristics include temperature difference between adjacent sections and peak temperature difference: (6)

[0101] In the formula: △T k,k+1 The average temperature difference between two adjacent sections; The difference between the peak temperatures of two adjacent plates; The average temperature of adjacent plates; , The peak temperature of adjacent plates;

[0102] Texture statistical features include contrast, energy, entropy, and correlation. (7)

[0103] In the formula: P(gi,gj) is the co-occurrence probability of pixel gray values ​​gi and gj; μ gi μ gjσ is the mean of the grayscale distribution; gi σ gj is the standard deviation of the grayscale distribution; Con is the contrast; Energy is the energy value reflecting the uniformity of the texture distribution; Entropy is the entropy representing the texture complexity; Cor is the correlation, reflecting the linear correlation between grayscale values;

[0104] All extracted features are normalized using the Z-score normalization method: (8);

[0105] In equation (8): x i x represents the original feature values; μ represents the mean of the feature in the training set; σ represents the standard deviation of the feature; ’ i These are the normalized eigenvalues;

[0106] Step S22: To balance multi-scale semantics and spatial resolution, YOLOv8 performs upsampling alignment and channel dimensionality reduction fusion on the multi-scale features to obtain the fused feature map F. fuse : (9)

[0107] In equation (9): F (l) W represents the feature of the l-th layer, where l ∈ [1, L], and L is the number of feature layers involved in the fusion. (l) To align convolutions; This is the upsampling factor; For channel concatenation; Ψ represents first-order dimensionality reduction; * represents convolution. The temperature statistical features, temperature difference features, and texture statistical features are used as auxiliary feature channel inputs, along with YOLOv8's multi-scale features F. (l) Channel splicing or weighted fusion is performed to obtain a joint feature map.

[0108] After completing multi-scale fusion and standardized features, robust intra-slice statistics, adjacency difference and texture representation are formed, providing high-quality, dimensionless input for subsequent striped sampling and sequence modeling.

[0109] Step S3 uses the principal axis geometry obtained in step S1 to perform the following in axis-normal coordinates: It allows for fine-sampling and in-band attention aggregation to form a one-dimensional strip sequence. Specific implementation includes:

[0110] Step S31, based on the parameters of the principal axis, in axis-normal coordinates, fuse the feature map F fuse Construct differentiable sampling bands along the principal axis and radial direction, specifying the sampling location along the axis and with lateral offset: (10)

[0111] In equation (10): c(s) is the point at the arc length position s on the axis; s is the arc length position along the principal axis; P0 is the axis starting point; q(s,r) is the sampling point at s with a normal offset of r; r is the normal displacement; n is the normal unit vector; S is the axis length; R is the set of discrete normal radii;

[0112] Step S32, after determining the sampling point, from F fuse The corresponding features of the sampling points are read out and weighted summation is performed within the band to obtain a one-dimensional representation z(s) for each s. The two-dimensional features are compressed into a "strip" using bilinear sampling and in-band aggregation. (11) (12)

[0113] In the formula: Let be the vector read from the fused features at position q(s,r); BL-sample is the bilinear interpolation sampling function; z(s) is the one-dimensional strip feature; W r The learnable weights are the values ​​corresponding to the radius r; exp(·) is the exponential function; α r (s) represents the weights after Softmax normalization.

[0114] Step S3 uses the principal axis coordinate system to perform line interest pooling, aligning the two-dimensional observations and reducing them to a one-dimensional strip sequence, providing a unified input for prior and decoding.

[0115] Step S4 enhances robustness using mechanistic priors, specifically including:

[0116] Step S41: To measure the "relative thermal anomaly of this wafer with the inner / outer zone and neighboring wafers", the Paired Thermal Relation Code (PRT) is used to analyze the relative thermal anomaly of candidate wafer k with the inner / outer zone and neighboring wafers, obtaining a thermal anomaly score A. PRT (s k ): (13)

[0117] In equation (13): s k R represents the axial position of the center of the k-th candidate slice; in R out Let R be the set of inner and outer zone radii; in |、|R out | represents the number of elements in the set; I(q) represents the normalized temperature at position q. It is the k-th candidate slice along the principal axis, within the inner and outer zone radius set R. in R out All sampling points; μ in μ outThe average temperature is the temperature of the inner and outer zones; ε is a very small positive number. For in s k The derivative along the axis at δ; s Adjacent slice spacing; φ(s) k ) represents the relative thermal feature vector; w represents the linear scoring parameter; σ(·) represents the Sigmoid function; the four relative thermal features—difference, ratio, derivative, and axial gradient—are μ, respectively. in -μ out , , μ in (s k )-μ in (s k ±δ s When a fault occurs in one insulator, its voltage approaches zero, while the voltages of the other insulators will change abruptly. This deviation provides a direct physical basis for fault detection.

[0118] Step S42: Perform autocorrelation / frequency domain analysis on the one-dimensional strip sequence using rhythm consistency RCH to obtain the rhythm consistency index sequence C. RCH : (14)

[0119] In equation (14): x(s) is a one-dimensional strip sequence response; R(τ) represents the mean; T represents the sequence length; R(τ) represents the autocorrelation. The desired pitch; max τ To maximize τ; ε is a minimum positive number; T-τ is the effective number of samples calculated by autocorrelation; This corresponds to the expected pitch autocorrelation value.

[0120] Step S43, assign the relative thermal anomaly score A PRT (s k ), rhythm consistency index sequence C RCH By splicing the data to one-dimensional strip features, a priori-blessed strip sequence is obtained.

[0121] Step S5 performs monotonic pointer decoding on the prior-blessed strip sequence: outputting the strip center, axial width, and confidence level sequentially from the high-voltage side to the ground side, and back-projecting them onto the image plane; during training, it is jointly optimized with the YOLO head. Specific implementation includes:

[0122] Step S51: Input the prior-enhanced stripe sequence into the pointer-type sequence decoder PSH. Encode the sequence context through one-dimensional convolution and recurrent units, unifying the encoding of the "stripe sequence + two priors" and capturing the context to obtain the encoded sequence H. enc :

[0123] H enc=BiGRU(ConvlD(Z||A PRT (s)||C RCH (15)

[0124] In equation (15): Z is the prior-bearing strip sequence; A PRT (s) is a paired thermal relation score sequence, A PRT (s)={A PRT (s k )};C RCH The sequence represents the rhythm consistency index; BiGRU represents a bidirectional gated cycle.

[0125] Step S52, in order to point to the next slice center in the "undecoded later segment", scoring is performed using pointer attention and monotonic masking to obtain a score e. k,s : (16)

[0126] In equation (16): exp(·) is the exponential function; s is the position of the candidate point on the principal axis; h k-1 The previous state; W q W k These are the query and key mapping matrices, respectively; v is a vector; tanh is the hyperbolic tangent; M k(s) For monotonic masks; π k(s) is the Softmax probability; T is the total length of the encoded sequence; s' is the Softmax normalized summation index; and k is the candidate slice.

[0127] Step S53: Process the image using position, width, confidence level, and reprojection information, and backproject the image into a candidate bounding box based on the image position result. (17)

[0128] In equation (17): Centered on the axis; Width of axis; For confidence; u, g, b u b g The header parameter is ReLU; ReLU is the activation function. is the center of the axial coordinate system; n is the normal unit vector; Proj is the projection function; Candidate bounding boxes for the image plane;

[0129] Step S54: During model training, sequence supervision is used in conjunction with the native YOLOv8 detection head for joint optimization. (18) (19)

[0130] In the formula, Let λ be the true position, true width, and confidence label of the k-th slice, respectively; ctr , λ w , λ conf , λ mono For loss weights; BCE is the binary cross-entropy; δ is the monotonic margin; max(0,·) is the ReLU form of monotonic constraint; L yolo The native detection loss for YOLOv8 includes bounding box regression loss, object confidence loss, and classification loss; L PRT For paired thermal relationship losses, used to constrain inter-plate thermal symmetry characteristics; L RCH Rhythm consistency loss is used to measure the periodicity and regularity of strip sequences; L joint The total loss is λ. prt , λ rch For the two prior loss weights;

[0131] Monotonic pointer decoding is used to achieve segment number-level structured output, and joint training with two-dimensional detection ensures sequence geometric consistency, thereby obtaining segment number-level detection results with continuous positions and ordered sequences.

[0132] Step S6 includes:

[0133] Step S61 involves weighted fusion of the relative thermal anomaly score, rhythm consistency index, and detection confidence (summarizing the three types of evidence into slice energy and suppressing random fluctuations) to construct a comprehensive energy index E. k : (20) (twenty one)

[0134] In the formula: E k λ represents the original fusion energy of the k-th fragment. a , λ c , λ p Weights for the three-path fusion; To test the confidence level; γ is the energy after EMA smoothing; γ is the EMA coefficient;

[0135] Step S62: Using hysteresis, window voting, and probability mapping, output the determination result of whether a zero-value insulator exists, as well as the final determination and piece number-level confidence level of the zero-value insulator. To avoid fluctuations in results due to certain transient anomalies, the comprehensive score is smoothed over time or in the neighborhood. (twenty two) (twenty three)

[0136] In the formula: T high Tlow y represents the hysteresis threshold; 1[·] represents the indicator function; W represents the window length; k To determine whether the k-th slice is an anomaly; Score k α represents the confidence level of the final film number; α and β are the linear mapping parameters between energy and probability.

[0137] The method of the present invention will be described below with reference to a specific embodiment:

[0138] During the data acquisition phase, drones or ground platforms equipped with infrared thermal imagers are used to inspect the insulator strings of transmission lines. The infrared thermal imager's temperature measurement range is set to -20~120℃, with a thermal sensitivity of less than 0.05℃ to ensure accurate capture of minute temperature differences. The drone flies at a constant altitude and speed, acquiring continuous infrared image sequences, and adding insulator numbers and GPS timestamps to each frame to provide spatiotemporal reference for subsequent processing.

[0139] During the image processing stage, the system scales the original infrared frames to the YOLOv8 inference resolution and uses a lightweight detection network to locate candidate regions of the insulator string in real time at the edges. The detected string-level candidate boxes are then appropriately expanded to form ROIs, such as... Figure 2 As shown. In parallel with bounding box detection, the main axis positioning head (SAL) regresses the axial heatmap, endpoints, and unit direction vector on the multi-scale neck features of YOLOv8 to obtain the directed principal axis and axis length from the high-voltage side to the grounding side. The two-dimensional search is reduced to a sequential modeling problem along a line, and ROI clipping and principal axis parameters together constitute the coordinate reference for all subsequent calculations.

[0140] In the feature extraction stage, to ensure a comprehensive and stable characterization of the thermal state of a single insulator, the ROI is divided into equidistant axial segments or physically segmented based on pitch estimation under principal axis constraints, and multidimensional features are extracted from each segment. Intra-segment statistical features include average temperature, peak temperature, minimum temperature, and standard deviation; inter-segment relative features include the average temperature difference and peak temperature difference between adjacent segments; texture features use the contrast, energy, entropy, and correlation of the gray-level co-occurrence matrix to measure the organization of thermal textures. All features are uniformly standardized using Z-scores, with the mean and standard deviation obtained statistically from the training set, thereby eliminating the influence of dimensional differences and environmental drift on modeling, while preserving the original temperature channels so that subsequent physical priors can directly act on the temperature sequence.

[0141] Table 1 Summary of Thermal Characteristics and Energy Integration Judgment for Some Single-Piece Insulators

[0142] The strip modeling and line interest pooling (LOI-Pooling) stage: Based on the starting point, axis, and normal given by the SAL (System Algorithm), the system constructs an axis-normal coordinate grid on the fused feature map. For each grid point, the fused features are read using bilinear interpolation and weighted and aggregated within the normal band using learnable attention weights to obtain a one-dimensional strip feature sequence along the principal axis. The temperature map is sampled in the same way on the same grid to obtain the mean temperature curve within the band. Through this alignment and dimensionality reduction, oblique sampling, scale variations, and slight deformations are significantly suppressed, and all subsequent priors and decoding are performed on the same sequentially stable axial sequence.

[0143] In the physical prior introduction stage, the system constructs thermal relation coding (PRT) and rhythm consistency coding (RCH) in parallel on the strips. For PRT, at the center of each candidate strip, the temperature of the inner and outer bands is averaged, and inter-band difference, inner-outer ratio, axial temperature gradient, and difference with neighboring strips are constructed. After linear mapping and Sigmoid, a thermal anomaly score of 0 to 1 is obtained. RCH uses the average temperature within the band or the strip feature norm as the signal, calculating autocorrelation or the proportion of the dominant frequency peak to obtain a rhythm consistency index of 0 to 1. A higher value indicates a more complete structural rhythm with nearly equidistant skirts. These two prior curves are directly concatenated with the main strip features as channel features, preserving physical interpretability and providing prior constraints for subsequent sequence decoding.

[0144] In the sequence decoding and slice number localization (PSH) stage, the system feeds "strip features + PRT thermal anomaly + RCH rhythm" into an encoder composed of one-dimensional convolution and bidirectional GRU to model local and long-range context. The decoding end employs pointer-type attention with a monotonic mask: each time, the position of the next slice is searched only within the axial range after the previous center, thus naturally satisfying the "high-voltage side to grounding side" sequence. For each selected position, the axial width and confidence are regressed, and the slice center and width are projected back onto the image plane using the principal axis start point, axial direction, and normal to obtain rectangular candidate boxes. During training, a joint loss is used: YOLO's detection loss ensures 2D localization consistency, SAL loss constrains principal axis geometry, and the PSH loss includes L1 terms for center and width, cross-entropy of confidence, and monotonic constraint terms. PRT and RCH can optionally be added through consistency regularization or contrast targets. The overall convergence is achieved using AdamW optimization and cosine annealing scheduling.

[0145] In the Decision Fusion (EFP) phase, the system linearly weights three independent evidence sources (PRT thermal anomaly, RCH rhythm disruption, and PSH detection uncertainty) into slice energy and smooths it temporally and spatially using an exponential moving average. To avoid jitter, a high-low threshold hysteresis strategy combined with sliding window voting is adopted: when the smoothed energy is consistently above the upper threshold and the proportion of "abnormally open" segments within the window reaches a set threshold, the slice is determined to be a zero-value insulator; when the energy falls below the lower threshold, the status is revoked. The final output includes the insulator string number, slice number, image plane positioning frame, zero-value determination label, and comprehensive confidence score, while retaining the visualization of strip curves, PRT, and RCH to support inspection interpretation. Based on the above process, YOLOv8-LOI can achieve a closed loop at the UAV edge, from rapid spindle positioning, strip alignment, physical prior enhancement to slice number-level pointer decoding and robust decision-making, demonstrating high accuracy and strong robustness for zero-value insulators in actual ultra-high voltage line scenarios.

[0146] Table 5. Partial Composite Energy (E_EMA)

[0147] This invention provides an insulator zero-value detection system that integrates spindle positioning and line interest pooling, for implementation Figure 1 The method shown includes:

[0148] The acquisition module is used to acquire infrared thermal imaging images of the insulator string;

[0149] The prediction module uses the YOLOv8 network of the parallel main string shaft positioning head SAL to predict the axial thermal map of the insulator string, the positions of the two endpoints, and the axial sequence direction from the high voltage side to the grounding side.

[0150] The extraction module is used to extract multi-scale features from infrared thermal imaging images from the backbone and neck network of YOLOv8, perform fusion and dimensionality reduction, and obtain a fused feature map.

[0151] The line interest pooling module is used to perform line interest pooling on the fused feature map using the principal axis coordinate system, aligning the two-dimensional observations and reducing them to a one-dimensional strip sequence.

[0152] The prior module is used to sample candidate slices radially within the inner and outer bands and left and right neighborhoods on the strip using Paired Thermal Relation Code (PRT), constructing four relative thermal features: difference, ratio, derivative, and axial gradient, and generating a relative thermal anomaly score. Rhythm Consistency (RCH) is used to perform autocorrelation frequency domain analysis on the one-dimensional strip sequence to measure whether the near-equidistant rhythm of the umbrella skirt is disrupted, generating a rhythm consistency index. The relative thermal anomaly score, rhythm consistency index, and one-dimensional strip features are concatenated to obtain the prior-enhanced strip sequence.

[0153] The slice output module is used to encode the prior-enhanced strip sequence into the pointer-type sequence decoder PSH, encode the sequence context through one-dimensional convolution and recurrent units, and output the slice result using pointer decoding under monotonic constraints. The back projection is used to generate candidate boxes for the image plane. The slice result includes confidence.

[0154] The judgment module is used for anomaly detection and decision fusion (EFP): it weights and fuses relative thermal anomaly scores, rhythm consistency indicators and confidence levels into a comprehensive energy index. Based on the comprehensive energy index, it adopts a window voting mechanism with confidence level smoothing and high and low threshold hysteresis to output the judgment result of whether there is a zero-value insulator, and locates the specific segment number of the faulty insulator and its confidence level.

[0155] The above-mentioned unit modules can be embedded in the processor of the electronic device in hardware form or independent of it, or they can be stored in the memory of the electronic device in software form, so that the processor can call and execute the corresponding operations of the above modules.

[0156] This embodiment also provides an electronic device, which can be a terminal. The electronic device includes a processor, a memory, a communication interface, a display screen, and an input device connected via a system bus. The processor provides computing and control capabilities. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. When the computer program is executed by the processor, it implements an insulator zero-value detection method that integrates spindle positioning and line interest pooling.

[0157] This embodiment also provides a computer-readable storage medium on which a computer program is stored, and when the computer program is executed by a processor, it performs the following steps:

[0158] Step S1: Acquire infrared thermal imaging images of the insulator string, input them into the YOLOv8 network of the parallel main string shaft positioning head SAL, and predict the axial thermal map, the position of the two end points, and the axial sequence direction from the high voltage side to the grounding side of the insulator string.

[0159] In step S2, the backbone and neck network of YOLOv8 extract multi-scale features from the infrared thermal imaging image, fuse and reduce the dimensions to obtain a fused feature map.

[0160] Step S3: Use the principal axis coordinate system to perform line interest pooling on the fused feature map, align the two-dimensional observations and reduce the dimension to a one-dimensional strip sequence;

[0161] Step S4: Using Paired Thermal Relation Code (PRT), radial inner and outer bands and left and right neighbor sampling are performed on candidate slices on the strip to construct four relative thermal features: difference, ratio, derivative, and axial gradient, generating a relative thermal anomaly score; using Rhythm Consistency Code (RCH), autocorrelation frequency domain analysis is performed on the one-dimensional strip sequence to measure whether the near-equidistant rhythm of the umbrella skirt is disrupted, generating a rhythm consistency index; the relative thermal anomaly score, rhythm consistency index, and one-dimensional strip features are concatenated to obtain the prior-bearing strip sequence.

[0162] Step S5: The prior-blessed strip sequence is input into the pointer-type sequence decoder PSH for encoding. The sequence context is encoded through one-dimensional convolution and recurrent units. Under monotonic constraints, the slice position result is output using pointer decoding and back-projected into the image plane candidate box. The slice position result includes confidence.

[0163] Step S6, Anomaly Detection and Decision Fusion (EFP): The relative thermal anomaly score, rhythm consistency index, and confidence level are weighted and fused into a comprehensive energy index. Based on the comprehensive energy index, a window voting mechanism with confidence level smoothing and high and low threshold hysteresis is adopted to output the determination result of whether there is a zero-value insulator, and to locate the specific piece number of the faulty insulator and its confidence level.

[0164] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

[0165] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0166] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. An insulator null detection method that fuses principal axis positioning and line interest pooling, characterized by, The method comprises the following steps: Step S1, collect the infrared thermal imaging image of the insulator string, input the YOLOv8 network of the parallel main string shaft positioning head SAL, and predict the axial thermal map of the insulator string, the two end point positions, and the shaft sequence direction from the high voltage side to the grounding side; Step S2, the backbone and neck network of YOLOv8 extract multi-scale features of the infrared thermal imaging image, perform fusion and dimension reduction, and obtain a fused feature map; Step S3, use the main shaft coordinate system to implement line interest pooling on the fused feature map, align the two-dimensional observation and reduce it to a one-dimensional strip sequence; Step S4, use the pair-wise thermal relationship coding PRT to sample the radial inner and outer bands and the left and right neighborhoods on the strip to generate relative thermal anomaly scores; use the rhythm consistency RCH to perform autocorrelation frequency domain analysis on the one-dimensional strip sequence to measure whether the umbrella skirt near the equidistant rhythm is damaged, and generate a rhythm consistency index; the relative thermal anomaly score, the rhythm consistency index and the one-dimensional strip feature are spliced to obtain a priori enhanced strip sequence; Step S5, input the priori enhanced strip sequence into the pointer sequence decoding head PSH for encoding, encode the sequence context through one-dimensional convolution and cyclic unit, and output the slice position result under the monotonic constraint by using the pointer decoding, and project it to the image plane candidate box, the slice position result including the confidence; Step S6, anomaly detection and decision fusion EFP: the relative thermal anomaly score, the rhythm consistency index and the confidence are weighted and fused into a comprehensive energy index, according to the comprehensive energy index, the confidence is smoothed, and a high and low threshold hysteresis window voting mechanism is adopted to output the judgment result of whether there is a zero value insulator, and locate the specific slice number and the confidence of the fault insulator.

2. The insulator null detection method of claim 1, wherein The step S1 comprises: Step S11, a UAV or a ground platform carries an infrared thermal imager to inspect the insulator string of the power transmission line, acquires a single frame or short time sequence infrared thermal imaging image, and processes the image by using a gain-bias model; Step S12, the backbone and neck features {P3, P4, P5} of YOLOv8 are connected in parallel with the main string shaft positioning head SAL to regress the main shaft parameters and the end points: ； where p(t) is the coordinate of the point on the principal axis; c = (c x ,c y ) is the center of the principal axis; v = (v x ,v y ) is the unit direction of the principal axis; t is the axis parameter; S is the axis length; ||·||2 is the two-norm; and the comprehensive supervision is: ； In the formula, L axis is the axis positioning loss; Focal(·) is the focal loss; is the predicted heat map output by the model axis positioning branch and normalized by the Sigmoid function; is the Gaussian labeled heat map generated according to the projection of the center position of the insulator sheet in the labeled sample along the main axis; are the predicted value and the labeled value of the two end points; are the predicted value and the labeled value of the direction vector of the axis sequence; is the dot product; λ e , λ o , λ L is the loss weight; ||·||1 is the L1 norm.

3. The insulator null detection method of claim 2, wherein, The step S2 comprises: Step S21, along the axial sequence direction, multi-scale features of the infrared thermal imaging image are extracted in the axial window and normalized, including temperature statistical features, temperature difference features and texture statistical features: The temperature statistical features include average temperature, peak temperature, minimum temperature and temperature standard deviation: ； In the formula, T k,t The temperature value of the tth pixel point in the kth insulating sub-area; n i The total number of pixels in the kth insulating sub-area; the average temperature of the kth insulating sub-area; the temperature peak value and the minimum temperature value in the kth insulating sub-area; The dispersion degree of the temperature distribution of the kth insulating sub-area; The temperature difference features include adjacent slice temperature difference and peak temperature difference: ； wherein: ΔT k,k+1 is the average temperature difference between two adjacent pieces; is the peak temperature difference between two adjacent pieces; is the average temperature of an adjacent piece; , is the peak temperature of an adjacent piece; The texture statistical features include contrast, energy, entropy and correlation: ； In the formula: P(gi,gj) is the co-occurrence probability of pixel gray values ​​gi and gj; μ gi μ gj σ is the mean of the grayscale distribution; gi σ gj is the standard deviation of grayscale distribution; Con is the contrast; Energy is the energy value reflecting the uniformity of texture distribution; Entropy is the entropy representing texture complexity; Cor is the correlation. All the extracted features are normalized by using the Z-score standardization method; Step S22, the multi-scale features are up-sampled, aligned and fused with channel dimension reduction to obtain a fused feature map F fuse : ； In the formula: F (l) is the l-th layer feature, l∈[1, L], L is the number of layers of features participating in fusion; W (l) is the alignment convolution; is the up-sampling rate; is the channel splicing; Ψ is the first dimension reduction; * is the convolution; the temperature statistical feature, the temperature difference feature and the texture statistical feature are auxiliary feature channels input, and the multi-scale feature F (l) is spliced or weighted and fused to obtain a joint feature map.

4. The insulator null detection method of claim 3, wherein, The step S3 comprises: Step S31, based on the parameters of the principal axes, in the axis-normal coordinates, in the fused feature map F fuse Constructing a differentiable sampling band along the principal axes and radial, explicit to sampling locations along the axis and lateral offset: ； In the formula, c(s) is a point at the arc length position s on the shaft; s is the arc length position along the main shaft; P0 is the starting point of the shaft; q(s, r) is a sampling point at s with a normal offset r; r is the normal displacement; n is the normal unit vector; S is the shaft length; and R is the normal discrete radius set; Step S32, after determining the sampling points, the corresponding features of the sampling points are read out and weighted summed in the band to obtain one-dimensional representation z(s) of each s: fuse The corresponding features of the sampling points are read out and weighted summed in the band to obtain one-dimensional representation z(s) of each s: ； ； where: is the vector read from the fused feature at position q(s, r); BL-sample is the bilinear interpolation sampling function; z(s) is the one-dimensional strip feature; W r is the learnable weight corresponding to radius r; exp(·) is the exponential function; a r (s) is the weight after Softmax normalization.

5. The insulator null detection method of claim 4, wherein, The step S4 comprises: Step S41, using the paired thermal relation coding PRT to analyze the relative thermal anomaly of the candidate slice k and the inner / outer belt and adjacent slices, and obtaining a thermal anomaly score A PRT (s k ): ； where s k is the axial position of the kth candidate slice center; R in , R out is the set of inner and outer band radii; |R in |, |R out | is the number of set elements; I(q) is the normalized temperature at position q, is the set of all sampling points along the principal axis at the kth candidate slice, within the set of inner and outer band radii R in , R out ; μ in , μ out is the average temperature of the inner and outer bands; ε is a small positive number; is the derivative along the axis at s k ; δ s is the adjacent slice distance; φ(s k ) is the relative heat feature vector; w is the linear scoring parameter; σ(·) is the Sigmoid; the four relative heat features of difference, ratio, derivative, and axial gradient are μ in - μ out , , , μ in (s k ) - μ in (s k ± δ s ). Step S42, using the rhythm consistency RCH to carry out autocorrelation / frequency domain analysis on the one-dimensional strip sequence, and obtaining a rhythm consistency index sequence C RCH : ； where: x(s) is the one-dimensional strip sequence response; is the mean; T is the sequence length; R(τ) is the autocorrelation; τ * is the expected pitch; max τ is the maximum over τ; ε is a small positive number; T-τ is the effective number of samples for the autocorrelation calculation; R(τ * ) is the autocorrelation value corresponding to the expected pitch. Step S43, relative thermal anomaly score A PRT (s k ), rhythm consistency index sequence C RCH Spliced to one-dimensional strip features, get prior possession strip sequence.

6. The insulator null detection method of claim 5, wherein, The step S5 comprises: Step S51, input the a priori held strip sequence into the pointer sequence decoding head PSH, and get the encoded sequence H by one-dimensional convolution and the context of the cyclic unit encoded sequence enc : H enc = BiGRU(Conv1D(Z||A PRT (s)||C RCH )); where: Z is the prior held strip sequence; A PRT (s) is the pairwise thermal relationship score sequence, A PRT (s) = {A PRT (s k )}; C RCH is the rhythm consistency index sequence; BiGRU is bidirectional gated recurrent; Step S52, scoring using pointer attention and monotonic mask, resulting in score e k,s : ； where exp(·) is the exponential function; s is the candidate point position on the principal axis; h k-1 is the previous state; W q , W k are the query and key mapping matrices, respectively; v is a vector; tanh is the hyperbolic tangent; M k(s) is the monotonic mask; π k(s) is the Softmax probability; T is the total length of the encoding sequence; s' is the Softmax normalization sum index; Step S53, processing is performed using the position, width, confidence and re-projection information, and the slice result is back-projected to the image plane candidate box: ； In the formula: is the axis center; is the axis width; is the confidence; u, g, b u , b g is the head parameter; ReLU is the activation function; is the axis coordinate center; n is the normal unit vector; Proj is the projection function; is the image plane candidate frame; Step S54, during model training, sequence supervision is used and combined with YOLOv8 native detection head optimization: ； ； wherein, are the true position, true width, and confidence label of the k-th slice, respectively; λ ctr , λ w , λ conf , λ mono is the loss weight; BCE is the binary cross-entropy; δ is the monotonicity margin; max(0, ·) is the monotonicity constraint in the form of ReLU; L yolo is the YOLOv8 native detection loss, including the bounding box regression loss, target confidence loss, and classification loss; L PRT is the pairwise thermal relation loss, used to constrain the thermal symmetry feature between slices; L RCH is the rhythm consistency loss, used to measure the periodicity and regularity of the strip sequence; L joint is the total loss; λ prt , λ rch are the two prior loss weights; The slice number level structured output is realized by using the monotone pointer decoding, and the sequence geometry consistency is ensured by joint training with two-dimensional detection, so that the slice number level detection result with continuous position and ordered sequence is obtained.

7. The insulator null detection method of claim 6, wherein, The step S6 includes: Step S61, the relative thermal anomaly score, the rhythm consistency index and the detection confidence are weighted and fused to construct a comprehensive energy index E k : ； ； wherein: E k is the kth patch raw fusion energy; λ a , λ c , λ p is the three-pass fusion weight; is the detection confidence; is the EMA smoothed energy; γ is the EMA coefficient; Step S62, using hysteresis, window voting, probability mapping, outputing the determination result of whether there is a zero value insulator, and the final determination of the zero value insulator and the slice number level confidence: ； ； In the formula, T high , T low is a hysteresis high-low threshold; 1[·] is an indicator function; W is a window length; y k is whether the k-th slice is judged as abnormal; Score k is the final slice number confidence; and α and β are energy-probability linear mapping parameters.

8. An insulator null detection system that fuses principal axis positioning with line interest pooling, characterized by, For implementing the method of any one of claims 1-7, comprising: The acquisition module is used for acquiring the infrared thermal imaging image of the insulator string; The prediction module uses the YOLOv8 network of the parallel main string shaft positioning head SAL to predict the axial thermal map of the insulator string, the positions of the two ends, and the axial sequence direction from the high-voltage side to the ground side; The extraction module is used for extracting the multi-scale features of the infrared thermal imaging image by the backbone and neck network of YOLOv8, fusing and reducing the dimension to obtain a fused feature map; The line interest pooling module is used for implementing line interest pooling on the fused feature map by using the main shaft coordinate system, aligning and reducing the dimension of the two-dimensional observation to a one-dimensional strip sequence; The prior module is used for using the paired thermal relationship coding PRT to sample the candidate slices on the strip in the radial inner and outer bands and left and right neighborhoods, constructing four relative thermal features of difference, ratio, derivative and axial gradient, and generating a relative thermal anomaly score; using the rhythm consistency RCH to perform autocorrelation frequency domain analysis on the one-dimensional strip sequence, measuring whether the umbrella skirt near equidistance rhythm is destroyed, and generating a rhythm consistency index; splicing the relative thermal anomaly score, the rhythm consistency index and the one-dimensional strip features to obtain a prior-aided strip sequence; The slice output module is used for inputting the prior-aided strip sequence into the pointer sequence decoding head PSH for encoding, encoding the sequence context by one-dimensional convolution and cyclic unit, and outputting the slice result under the monotone constraint by using the pointer decoding, back-projecting to the image plane candidate box, the slice result including the confidence; The determination module is used for anomaly detection and decision fusion EFP: the relative thermal anomaly score, the rhythm consistency index and the confidence are weighted and fused into a comprehensive energy index, according to the comprehensive energy index, the confidence is smoothed, and the high and low threshold hysteresis window voting mechanism is used to output the determination result of whether there is a zero value insulator, and to locate the specific slice number and its confidence of the fault insulator. 9.An electronic device comprising a memory and a processor, the memory storing a computer program, wherein, The processor executes the computer program to realize the steps of the insulator zero value detection system of claim 1-7.

10. A computer-readable storage medium having stored thereon a computer program, characterized in that, The computer program is executed by the processor to realize the steps of the insulator zero value detection system of claim 1-7.

Images