Grounding down conductor corrosion cross-modal identification method based on shared latent space construction

Abstract

The application discloses a grounding down lead corrosion cross-modal identification method based on a shared hidden space construction, comprising the following steps: step S1: multi-source data synchronous acquisition; step S2: single-modal feature extraction; step S3: cross-modal feature alignment and fusion; step S4: corrosion grade identification based on a neural dynamics system; and step S5: visualization and alarm. The application synchronously acquires two types of heterogeneous data, respectively extracts acoustic features and visual features, and innovatively adopts a cross-modal hidden space alignment technology, uses a reconstruction error as a corrosion criterion, and realizes full-automatic, high-precision and interpretable identification of internal and external corrosion states of the grounding down lead. The application overcomes the limitation of single-modal detection, solves the bottleneck of the prior art in aspects of concealed corrosion identification, early warning, anti-interference capability and automation level, and is suitable for long-term health monitoring and intelligent operation and maintenance of a power transmission tower grounding down lead.

Description

Technical Field

[0001] This invention belongs to the field of power equipment condition detection and fault diagnosis technology, and particularly relates to a cross-modal identification method for grounding down conductor corrosion based on shared hidden space construction. Background Technology

[0002] Grounding down conductors are a critical component of lightning protection systems for transmission towers. Long-term exposure to harsh environments such as humidity and salinity makes them susceptible to metal corrosion. Severe corrosion significantly reduces the conductivity of the grounding system, threatening the safe operation of the power grid. Traditional detection methods, such as visual inspection and multimeter measurements, suffer from high subjectivity, low sensitivity, and difficulty in detecting internal corrosion.

[0003] In recent years, electromagnetic ultrasonic guided wave technology has been widely used in the non-destructive testing of long and narrow structures such as pipes and steel cables due to its advantages such as no need for coupling agents, long-distance propagation, and sensitivity to internal defects. When the guided wave propagates in the grounding down conductor, the corroded area will cause reflection, scattering, and mode conversion of the wave. By analyzing the amplitude, phase, spectrum, and other characteristics of the echo signal, defect information can be inverted.

[0004] Meanwhile, visible light image recognition technology, with the help of deep learning models, can effectively identify the color, texture, and morphological characteristics of surface rust, and is suitable for macroscopic rust assessment.

[0005] However, existing technologies have the following problems:

[0006] Single-mode limitations: Although electromagnetic ultrasonic guided waves can detect internal corrosion, they are difficult to locate and quantify the surface corrosion morphology; visible light images can only reflect the surface condition and cannot detect the degree of internal corrosion.

[0007] Multimodal fusion methods are outdated: existing fusion methods mostly use simple splicing or weighted averaging, which do not fully explore the potential semantic associations between acoustic signals and images.

[0008] Lack of cross-modal alignment mechanism: The two types of data have large differences in dimension, scale and semantics, making it difficult for conventional methods to achieve effective feature fusion.

[0009] In the prior art, patent CN202311218771.3 proposed an image recognition method based on improved StarNet-YOLOv8 and AKConv-YOLOv8. Although it improves the detection accuracy and robustness of conductive lead corrosion in complex scenarios, it relies entirely on visible light images for recognition. It can only detect surface rust and cannot detect hidden damage such as internal cross-sectional thinning and cavitation. It is also easily affected by environmental interference such as changes in lighting, shadows, stains, and reflections, which can lead to misjudgment or missed detection. At the same time, it lacks a physical correlation with the degree of material damage, making it difficult to achieve quantitative assessment and early warning of the degree of corrosion. Patent CN202111257871.8 employs electromagnetic ultrasonic guided wave technology, analyzing the number and location of echo pulse waves to determine corrosion defects. While it can detect internal structural changes and achieve non-contact detection, this method relies on manual observation of waveforms for interpretation, resulting in high subjectivity, low automation, and insufficient extraction of deeper features such as amplitude attenuation, phase shift, and spectral changes, leading to low information utilization. It also lacks a visual representation of rust morphology and fails to cross-verify with surface conditions, making it prone to misjudging structural abrupt changes as corrosion defects. Both methods share the limitation of using a single-modal detection approach, failing to achieve multi-source information fusion, resulting in significant deficiencies in detection completeness, accuracy, intelligence level, and engineering practicality. Summary of the Invention

[0010] To address the shortcomings of existing technologies, this invention aims to provide a high-precision, robust, and quantifiable method for identifying corrosion of grounding down conductors. Specifically, it addresses the problems of the pure image recognition method described in patent CN202311218771.3, which relies solely on visible light images for corrosion detection, failing to identify internal corrosion, being susceptible to light and background interference, and lacking physical quantification; and the pure electromagnetic ultrasonic guided wave method described in patent CN202111257871.8, which relies on manual pulse wave interpretation, only roughly assesses defects based on wavenumbers, and lacks automation and intelligent analysis. To achieve the above objectives, this invention provides a cross-modal identification method for grounding down conductor corrosion based on a shared latent space, comprising the following steps:

[0011] Step S1: Multi-source data synchronous acquisition, synchronously acquire electromagnetic ultrasonic guided wave echo signals and visible light image data at the same spatial location, establish a spatiotemporal alignment benchmark, and ensure the correspondence between the two types of data at the detection location;

[0012] Step S2: Single-mode feature extraction, performing independent feature modeling for electromagnetic ultrasonic guided wave signals and visible light images respectively;

[0013] Step S3: Cross-modal feature alignment and fusion;

[0014] Step S4: Corrosion level identification based on neurodynamic system. Based on fusion features, a neurodynamic classifier based on spiking neural network and liquid machine is proposed to replace the traditional SVM or MLP model, so as to realize automatic, dynamic, low-power classification and quantitative evaluation of the corrosion degree of grounding down conductor.

[0015] Step S5: Visualization and Alarms. Visualize the classification results and issue tiered alarms through multiple channels.

[0016] As a further improvement of the present invention, the multi-source data synchronous acquisition in step S1 includes establishing a precise mapping between the "probe center point" and the "image coordinate system origin" so that the subsequently acquired guided wave signals and image data strictly correspond in spatial position; configuring and starting the electromagnetic ultrasonic guided wave detection system to acquire high-quality echo signals; starting the visible light imaging module and completing the setting of image acquisition parameters; the electromagnetic ultrasonic guided wave detector outputs a TTL level trigger pulse at the same time as each excitation signal is emitted, and this pulse is connected to the external trigger input port of the visible light camera as the main synchronization signal, forcing the camera to synchronously capture a frame of image at the moment of guided wave emission, so that each group of guided wave signals and its corresponding visible light image are strictly aligned in time.

[0017] As a further improvement of the present invention, in step S2, the single-mode feature extraction first uses Hilbert transform to extract the envelope curve of the analytic signal of the guided wave echo signal, and then calculates key physical characteristic parameters, including the attenuation rate of the echo amplitude relative to the reference signal. Phase offset between the reflected wave and the reference wave The offset of the main frequency component And the energy entropy of the signal under multi-scale wavelet decomposition. Finally, the above four features are combined to form a four-dimensional acoustic feature vector. =[ , , , [This is used as a feature representation of the guided wave mode; simultaneously, visual feature extraction is performed on the synchronously acquired visible light images. First, the images are normalized and input into a pre-trained convolutional neural network based on the ResNet-18 architecture. The first few layers of the convolutional modules are frozen to retain general edge and texture features. The high-order semantic feature vector output by the penultimate fully connected layer is extracted.] This achieves the mapping from pixel space to semantic feature space; finally, and Both types of multi-source feature vectors are standardized to eliminate dimensional differences.

[0018] As a further improvement of the present invention, the cross-modal feature alignment and fusion in step S3 includes:

[0019] S3.1 Network Structure Design – Acoustic Coding Branch;

[0020] S3.2 Network Structure Design – Visual Encoding Branch;

[0021] S3.3 Shared Semantic Mapping Layer Design;

[0022] S3.4 Dual Decoding Branch: Feature Reconstruction Network Based on U-Net Architecture;

[0023] S3.5 Loss Function Design: Joint Optimization Objective;

[0024] S3.6 Corrosion Criterion Generation: Reconstruction Error as an Anomaly Indicator;

[0025] S3.7 Fusion Feature Output.

[0026] As a further improvement to the present invention, the S3.1 network structure design—acoustic coding branch—is an acoustic coding branch based on diffusion mechanism and Transformer architecture, which maps low-dimensional guided wave features to a highly discriminative latent space representation. The specific implementation process is as follows:

[0027] (1) Forward noise addition process: Constructing a multi-scale noise perturbation sequence;

[0028] Let the original guided wave eigenvector be... It includes four physical quantities: amplitude attenuation rate, phase shift, dominant frequency shift, and wavelet energy entropy. During the training phase, a forward noise addition process is defined to simulate potential electromagnetic interference, contact noise, and temperature drift in the field.

[0029] N(0,1);

[0030] Where t=1,2,...,T is the number of diffusion steps, and T is set to a total of 50 diffusion steps. This value can achieve a balance between computational efficiency and noise coverage.

[0031] ;

[0032] ;

[0033] in A cosine noise scheduling strategy is adopted. =0.008 is the offset coefficient; This represents the characteristic state after being disturbed by noise at step t;

[0034] (2) Inverse denoising encoder: Feature evolution network based on temporal conditional Transformer;

[0035] Define an inverse denoising encoder, starting from the noise-contaminated... The encoder gradually recovers clean feature representations related to the corrosion state; it uses a Transformer architecture as its core network. The specific structure is as follows:

[0036] Input concatenation: Combine current noise features diffusion steps t and original features Concatenate into joint input:

[0037] ;

[0038] in It is the sinusoidal position code for time step t, used to provide time information;

[0039] Through a fully connected layer Mapped to The dimensional embedding space is used as the input to the Transformer;

[0040] Transformer encoder layer: Stacks L=3 layers of standard Transformer blocks, each layer containing: four multi-head self-attention modules, query / key / value dimension. Two fully connected layers, with an intermediate dimension of 512, and GELU activation function; residual connections and layer normalization; the output of the last Transformer layer is then subjected to global average pooling, followed by dimensionality reduction through two MLP layers to output the final acoustic latent space representation:

[0041] ;

[0042] It is defined as: a low-dimensional implicit representation that is highly correlated with the internal structural damage of the grounding lead, which is evolved from the guided wave characteristics of noise disturbance by a time-conditional Transformer network during the diffusion denoising process.

[0043] As a further improvement to the present invention, the S3.2 network structure design—visual encoding branch—refers to the input being a 512-dimensional image feature vector. Based on the ResNet-18 model, the encoding branch employs multi-layer nonlinear transformations to compress features into a 256-dimensional visual latent space. .

[0044] As a further improvement to this invention, the S3.3 shared semantic mapping layer design introduces a shared semantic mapping layer and configures two independent decoding branches to form a complete cross-modal feature reconstruction path. This design maps heterogeneous features from different modalities to the same semantic space, ensuring the fidelity of each feature while establishing a potential consistency relationship between "internal structural state" and "surface visual representation." The acoustic latent space representation extracted in steps S3.1 and S3.2 is then used to... Representation of visual latent space Each input to the shared semantic mapping layer is mapped to a unified 128-dimensional shared latent space Z using a learnable linear transformation matrix. To achieve cross-modal semantic alignment;

[0045] ;

[0046] ;

[0047] The mapping weight matrix representing the acoustic modes; The mapping weight matrix representing the visual modality; , Represents the bias vector; , These represent the projection vectors of the waveguide features and the image features in the shared latent space Z, respectively.

[0048] Introduce cross-modal alignment loss during training:

[0049] ;

[0050] The two types of heterogeneous features are forced to be semantically consistent and should be highly aligned under normal conditions.

[0051] As a further improvement to the present invention, S3.4 dual decoding branch refers to setting up two independent decoders based on the U-Net architecture, respectively from... and Reconstruct the original input features to form an end-to-end reconstruction path;

[0052] (1) Acoustic decoder based on U-Net guided wave feature reconstruction network: Treating the input as a "feature image" into a one-dimensional U-Net structure, the original 4-dimensional guided wave feature vector is reconstructed.

[0053] = ;

[0054] in, The reconstructed waveguide characteristics are used to calculate the reconstruction error with the true values;

[0055] (2) Visual decoder based on U-Net image semantic feature reconstruction network: A standard two-dimensional U-Net is used to reconstruct the image semantic features from... Reconstruct the semantic feature vector of the original 512-dimensional image.

[0056] = ;

[0057] The output is compared with the original visual features extracted by ResNet-18 to calculate the reconstruction error of the image modality.

[0058] As a further improvement of the present invention, (1) the S3.5 loss function design defines a multi-task loss function to drive the network to simultaneously complete the two major objectives of feature reconstruction and cross-modal alignment:

[0059] ;

[0060] To measure whether the waveguide characteristics are accurately recovered during the encoding-decoding process;

[0061] To measure whether image features are fully preserved;

[0062] Forced acoustic implicit representation Visual latent representation They tend to agree before even entering the shared space;

[0063] Weighting coefficient The importance of balancing the three losses;

[0064] (2) The corrosion criterion S3.6 is generated. In actual testing, a new "acoustic-image" sample is input, and its total reconstruction error is calculated:

[0065] ;

[0066] Reconstruction error It is used as a confidence index for corrosion.

[0067] As a further improvement to the present invention, in step S4, the corrosion level identification based on a neurodynamic system includes the following structural design of the neurodynamic classifier:

[0068] (1) Input coding layer: time-to-pulse conversion

[0069] Shared latent space feature vectors Encoded as a pulse sequence, using a time-to-pulse interval encoding strategy:

[0070] ;

[0071] in, The time when the i-th input neuron first fires a pulse; The maximum latency is 50ms, which is set based on the balance between edge device clock accuracy and response speed. , used to Map to the interval (0,1);

[0072] (2) Liquid Pool: High-dimensional dynamic state evolution;

[0073] Construct a recurrent connection network consisting of N=500 spiking neurons, called the "liquid pool"; each neuron follows the Leaky Integrate-and-Fire model:

[0074] ;

[0075] in, =10ms: Membrane time constant, reflecting the neuron's ability to "memorize" input; =1.0MΩ: Input resistance, normalized; Input connection weights, from the 128-dimensional input layer to the liquid pool, are initialized using a normal distribution; : Recursive connection weights, sparse random connections (connection probability p=30%), simulating the sparsity of biological neural networks; Input pulse signal; Neuronal impulses inside the liquid pool;

[0076] (3) Readout Layer: State decoding and classification;

[0077] After the pulse input ends, neural activity in the liquid pool is collected within a time window of t∈[50,100] ms, and the following statistical features are extracted to form the readout vector f. :

[0078] First 500 dimensions: Average firing frequency per neuron

[0079] ;

[0080] Dimension 501: Population Synchronization Index

[0081] ,in This represents the phase of neuronal activity, reflecting group synergy;

[0082] 502nd dimension: membrane potential variance Characterizes the activity level of the system;

[0083] These features are concatenated into a vector, which is then input into a lightweight linear classifier:

[0084] ;

[0085] in This represents the predicted probability distribution of the corrosion level; the final output is the corrosion level.

[0086] Compared with the prior art, the present invention has the following beneficial effects:

[0087] This invention proposes a cross-modal identification method for grounding down conductor corrosion based on shared latent space construction. By simultaneously acquiring two types of heterogeneous data—electromagnetic ultrasonic guided wave echo signals and visible light image data from the same spatial location—a spatiotemporal alignment benchmark is established to ensure the correspondence between the two types of data at the detection location. This solves the problem that traditional single-modal detection cannot correlate "internal structural changes" and "surface morphology." Acoustic and visual features are extracted separately, and a cross-modal latent space alignment technique is innovatively employed, using reconstruction error as a corrosion criterion to achieve fully automatic, high-precision, and interpretable identification of the internal and external corrosion states of grounding down conductors. This invention overcomes the limitations of single-modal detection and solves the bottlenecks of existing technologies in areas such as hidden corrosion identification, early warning, anti-interference capability, and automation level. It is suitable for long-term health monitoring and intelligent operation and maintenance of grounding down conductors on transmission towers. Attached Figure Description

[0088] Figure 1 This is a flowchart of the cross-modal identification method for grounding down conductor corrosion based on shared latent space construction according to the present invention. Detailed Implementation

[0089] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. These embodiments are based on the technical solution of the present invention and provide detailed implementation methods and specific operating procedures. However, the scope of protection of the present invention is not limited to the following embodiments.

[0090] The flowchart of the cross-modal identification method for grounding lead corrosion based on shared latent space construction of the present invention is as follows: Figure 1 As shown.

[0091] The steps of a specific embodiment are as follows:

[0092] Step S1: Synchronous Acquisition of Multi-Source Data

[0093] Before conducting the test, the testing equipment is first deployed and precisely positioned. The electromagnetic ultrasonic probe is securely mounted on the surface of the grounding lead wire using a magnetic clamp, ensuring that its two poles are aligned with the axis of the lead wire to effectively excite and receive ultrasonic guided waves propagating along the metal rod. Simultaneously, a high-definition visible light camera is fixedly installed near the probe, and its shooting angle and focal length are adjusted to ensure that its field of view completely covers the detection area corresponding to the probe and the surface of the lead wire within a range of at least 20 cm upstream and downstream. The probe and camera are connected by a rigid bracket or by establishing a precise mapping between the "probe center point" and the "origin of the image coordinate system" through a pre-calibrated spatial coordinate transformation relationship, thereby ensuring that the subsequently acquired guided wave signals and image data strictly correspond in spatial position.

[0094] Subsequently, the electromagnetic ultrasonic guided wave testing system was configured and started to acquire high-quality echo signals. The main unit of the electromagnetic ultrasonic guided wave testing instrument was turned on, and the testing interface was accessed. The excitation signal was set to a Hanning window modulated sine pulse, the center frequency was selected as 1.0MHz, the excitation voltage was set to 300V, the sampling rate was set to no less than 10MHz, and the signal acquisition time window was set to 80 microseconds, sufficient to capture the reflected echo after the guided wave has propagated several meters in the downleader. Simultaneously, the signal accumulation function was enabled, and the averaging frequency was set to 16 times to suppress random noise and improve the signal-to-noise ratio to above 20dB. The preamplifier gain was set to 40dB, and the dynamic range was set to no less than 80dB to ensure that even weak defect echoes could be effectively captured. At the same time, the visible light imaging module was started, and the image acquisition parameters were set. Before the inspection begins, the camera enters working mode, is set to automatic exposure mode and locks the white balance parameters to prevent image color distortion caused by ambient light fluctuations. In scenes with insufficient light or backlight interference, a ring LED fill light is activated, with the light source color temperature set to 5500K and the illuminance not less than 500 lux, to ensure clear surface texture and moderate contrast of the lead wire. A fixed-focus lens is selected and the focal length is adjusted so that the smallest resolvable detail in the image is less than 0.5 mm, ensuring the recognition accuracy of rust texture. The output image format is RGB three-channel digital image with a resolution of not less than 1280×720 pixels.

[0095] To achieve precise temporal synchronization between acoustic signals and visual images, an electromagnetic ultrasonic guided wave detector outputs a TTL-level trigger pulse simultaneously with each excitation signal transmission. This pulse serves as the main synchronization signal, connected to the external trigger input port of a visible light camera. This forces the camera to capture a frame of image synchronously at the instant the guided wave is emitted, ensuring strict temporal alignment between each set of guided wave signals and its corresponding visible light image. This avoids timing deviations caused by software delays or system clock asynchrony. The response delay of the entire synchronization system is controlled within ±1 millisecond, meeting the time accuracy requirements for cross-modal data correlation analysis and providing a reliable timing reference for subsequent feature fusion.

[0096] Step S2: Single-modal feature extraction

[0097] After completing the synchronous acquisition of acoustic signals and visual image data, the single-mode feature extraction stage begins, where independent feature modeling is performed on the electromagnetic ultrasonic guided wave signal and the visible light image respectively. First, for the guided wave echo signal, the Hilbert transform is used to extract the envelope curve of its analytic signal, and then key physical characteristic parameters are calculated, including the attenuation rate of the echo amplitude relative to the reference signal. (defined as) = 1 - / (used to reflect energy loss caused by defects), the phase offset between the reflected wave and the reference wave. (Unit: radians; sensitive to minute structural changes), offset of the dominant frequency component. (Obtained through short-time Fourier transform or wavelet transform, reflecting changes in dispersion characteristics), and the energy entropy of the signal under multi-scale wavelet decomposition. (where p) i (The energy proportion of wavelet coefficients at each scale is used to characterize signal complexity and scattering degree). Finally, the above four features are combined to form a four-dimensional acoustic feature vector. =[ , , , [This is used as a feature representation of the guided wave mode.] Simultaneously, visual feature extraction is performed on the synchronously acquired visible light images. First, the images are normalized to 640×640 pixels and input into a pre-trained convolutional neural network based on the ResNet-18 architecture. The first few convolutional layers are frozen to preserve general edge and texture features. The 512-dimensional high-order semantic feature vector output from the penultimate fully connected layer is then extracted. This vector implicitly contains deep visual information such as color distribution, texture roughness, and patch morphology of the rusted area, thus achieving a mapping from pixel space to semantic feature space; finally, it will... and Both types of multi-source feature vectors were standardized to eliminate dimensional differences, laying a data foundation for subsequent cross-modal fusion.

[0098] Step S3: Cross-modal feature alignment and fusion

[0099] S3.1 Network Structure Design — Acoustic Coding Branch (Feature Implicit Enhancement Encoder Based on Diffusion Mechanism)

[0100] To enhance the robustness and semantic representation of guided wave features in complex electromagnetic environments, this invention proposes an acoustic coding branch based on a diffusion mechanism and Transformer architecture to map low-dimensional guided wave features to a highly discriminative latent space representation. Inspired by the "denoising evolution" concept of the diffusion model, this encoder is not used for signal generation but focuses on extracting essential damage features related to corrosion from noisy observations. It is particularly suitable for detecting long-term exposed power equipment such as grounding down conductors, where the signal-to-noise ratio fluctuates significantly. The network is implemented as follows:

[0101] (1) Forward noise addition process: constructing a multi-scale noise perturbation sequence

[0102] Let the original guided wave eigenvector be... It includes four physical quantities: amplitude attenuation rate, phase shift, dominant frequency shift, and wavelet energy entropy. During the training phase, a forward noise addition process is defined to simulate potential electromagnetic interference, contact noise, and temperature drift in the field.

[0103] N(0,1)

[0104] Where t=1,2,...,T is the number of diffusion steps, and T is set to a total of 50 diffusion steps. Experiments have verified that this value can achieve a balance between computational efficiency and noise coverage. ,in A cosine noise scheduling strategy is adopted. =0.008 is the offset coefficient; This represents the characteristic state after being disturbed by noise at step t.

[0105] (2) Inverse denoising encoder: Feature evolution network based on temporal conditional Transformer

[0106] Define an inverse denoising encoder whose goal is to denoise from noise-contaminated text. The encoder gradually recovers clean feature representations related to the corrosion state. It employs a Transformer architecture as its core network. The specific structure is as follows:

[0107] Input concatenation: Combine current noise features diffusion steps t and original features Concatenate into joint input:

[0108]

[0109] in It is the sinusoidal position code of time step t, used to provide time information.

[0110] Through a fully connected layer Mapped to The dimensional embedding space is used as the input to the Transformer.

[0111] Transformer encoder layer: Stacks L=3 layers of standard Transformer blocks, each layer containing: four multi-head self-attention modules, query / key / value dimension. Two fully connected layers, with an intermediate dimension of 512, and GELU activation function; residual connections and layer normalization. The output of the final Transformer layer is then subjected to global average pooling, followed by dimensionality reduction through two MLP layers to output the final acoustic latent space representation:

[0112]

[0113] It is defined as: a low-dimensional implicit representation that is highly correlated with the internal structural damage of the grounding lead, which is evolved from the guided wave characteristics of noise disturbance by a time-conditional Transformer network during the diffusion denoising process.

[0114] S3.2 Network Structure Design—Visual Encoding Branch

[0115] The input is a 512-dimensional image feature vector. Derived from the ResNet-18 model, it contains visual semantics such as color distribution, texture roughness, and patch continuity of the rusted area. Due to its high dimensionality and rich information, the encoding branch employs multi-layer nonlinear transformation (ReLU activation function + BatchNorm) to compress the features into a 256-dimensional visual latent space. .

[0116] S3.3 Shared Semantic Mapping Layer Design

[0117] To achieve effective semantic alignment and fusion of electromagnetic ultrasonic guided wave signals and visible light images, this invention introduces a shared semantic mapping layer and configures two independent decoding branches to form a complete cross-modal feature reconstruction path. This design aims to map heterogeneous features from different modalities to the same semantic space, establishing a potential consistency relationship between "internal structural state" and "surface visual representation" while ensuring the fidelity of each feature. This is particularly suitable for critical power equipment such as grounding down conductors that are exposed for long periods and prone to hidden corrosion.

[0118] The acoustic latent space extracted in steps S3.1 and S3.2 is represented as... Representation of visual latent space Each input to the shared semantic mapping layer is mapped to a unified 128-dimensional shared latent space Z using a learnable linear transformation matrix. To achieve cross-modal semantic alignment.

[0119]

[0120]

[0121] The mapping weight matrix representing the acoustic modes; The mapping weight matrix representing the visual modality; , Represents the bias vector; , Let Z represent the projection vectors of the guided wave features and the image features in the shared latent space Z, respectively.

[0122] Furthermore, a cross-modal alignment loss is introduced during training:

[0123]

[0124] This is used to force two types of heterogeneous features to be semantically consistent, especially to ensure high alignment under normal conditions.

[0125] S3.4 Dual Decoding Branch: Feature Reconstruction Network Based on U-Net Architecture

[0126] To verify whether the shared latent space Z effectively preserves the key information of the original modality, this invention sets up two independent decoders based on the U-Net architecture, respectively from... and The original input features are reconstructed to form an end-to-end reconstruction path. U-Net's symmetric encoder-decoder structure and skip connection mechanism enable it to exhibit excellent fidelity in low-dimensional to high-dimensional and non-image to image / vector inverse mapping tasks.

[0127] (1) Acoustic decoder: U-Net-based guided wave feature reconstruction network

[0128] Although guided wave characteristics are low-dimensional vectors, they contain multi-scale physical information (such as amplitude, phase, and spectrum). Therefore, this invention will... The input, considered as a "feature image," is a one-dimensional U-Net structure (1D U-Net) used to progressively reconstruct the original 4-dimensional guided wave feature vector. The specific implementation is as follows:

[0129] P1 will Expanded to a 128×1 vector, it serves as the input to the 1D U-Net;

[0130] The P2 encoding path contains 3 downsampling blocks, each consisting of 1D convolution (kernel=3, stride=1) + BatchNorm + LeakyReLU + MaxPool1d (kernel=2), progressively extracting multi-scale features;

[0131] The P3 decoding path contains 3 upsampling blocks, each consisting of a 1D transposed convolution + a 1D convolution;

[0132] P4 skip connections concatenate the outputs of each layer in the encoding path with the corresponding decoding layer, preserving local details;

[0133] The final output layer of P5 is a fully connected layer, which maps the features of the last layer to a 4-dimensional space:

[0134] =

[0135] in, The reconstructed waveguide characteristics are used to calculate the reconstruction error with the true values.

[0136] (2) Visual decoder: Image semantic feature reconstruction network based on U-Net

[0137] The visual decoder uses a standard two-dimensional U-Net (2D U-Net) for processing data from... Reconstructing the semantic feature vector of the original 512-dimensional image The specific implementation is as follows:

[0138] Q1 will Reshape it into an 8×4×4 tensor and use it as input to a 2D U-Net;

[0139] The Q2 encoding path contains four downsampling blocks (convolution + pooling) to progressively extract high-level semantics;

[0140] The Q3 decoding path contains four upsampling blocks, which restore spatial resolution through transposed convolution;

[0141] Q4 skip connections fuse the corresponding layer features of the encoding and decoding paths;

[0142] Q5 finally compresses the output feature map into a 512-dimensional vector using global average pooling:

[0143] =

[0144] The output is compared with the original visual features extracted by ResNet-18 to calculate the reconstruction error of the image modality.

[0145] S3.5 Loss Function Design: Joint Optimization Objective

[0146] To train the network, a multi-task loss function is defined to drive the network to simultaneously achieve the two major objectives of feature reconstruction and cross-modal alignment:

[0147]

[0148] This measure assesses whether the guided wave characteristics are accurately recovered during the encoding-decoding process. If the grounding lead has severe corrosion, the guided wave signal distortion will be large, and this error will increase, but under normal circumstances it should be as small as possible.

[0149] This measure assesses whether image features are fully preserved. This metric may increase when the image is affected by lighting or dirt, reflecting uncertainty in visual information.

[0150] It forces acoustic implicit representation Visual latent representation They tend to conform before even entering the shared space.

[0151] For a properly functioning grounding lead, its internal structure is stable (stable guided wave signal) and its surface is rust-free (uniform image texture), and both are highly consistent. It's very small. However, when corrosion occurs, two inconsistent situations may arise: internal corrosion but an intact exterior; or abnormal waveguide operation. Significant changes, but the image appears normal. Unchanged, significant differences; surface contamination but intact structure: image misidentified as corrosion. Abnormal; waveguide normal. The difference remains the same, but the variation is equally significant. Therefore, it essentially reflects the "consistency between internal and external conditions" and is an important indicator for judging whether corrosion exists.

[0152] Weighting coefficient settings:

[0153] To balance the importance of the three losses, let... =0.6, because the signal-to-noise ratio of guided wave signals is high, its fidelity is given priority; =0.3, the image is easily disturbed, so the weight should be reduced appropriately; =0.1. While the alignment term is crucial, it should not be too strong to avoid suppressing the modal characteristics of each component.

[0154] S3.6 Corrosion Criterion Generation: Reconstruction Error as an Anomaly Indicator

[0155] After training, the deep neural network learns the prior knowledge that "sound-image features should remain consistent under normal conditions." In actual detection, a new sound-image sample is input, and its total reconstruction error is calculated:

[0156]

[0157] This error It is used as a confidence index for corrosion.

[0158] S3.7 Fusion Feature Output

[0159] Finally, a fused feature vector is extracted from the shared latent space Z. This vector contains both the physical damage response of the guided wave and the visual semantics of the image, and is a highly abstract and discriminative joint representation, which serves as the input to the corrosion level classifier in the next stage (step S4).

[0160] Step S4: Corrosion level identification based on neurodynamic system

[0161] After completing cross-modal feature alignment and fusion, a 128-dimensional shared latent space feature vector is obtained. This vector integrates internal structural damage information reflected by electromagnetic ultrasonic guided waves (such as cross-sectional thinning and intergranular corrosion) with surface corrosion morphology features characterized by visible light images (such as rust distribution and texture fragmentation). The next step is to leverage this fused feature to achieve automatic, dynamic, and low-power classification and quantitative assessment of the corrosion degree of grounding down conductors. This invention innovatively proposes a neurodynamic classifier based on spiking neural networks and liquid dynamics machines, replacing traditional SVM or MLP models, for achieving high-efficiency and robust corrosion level identification. Specific classifier structure design:

[0162] (1) Input coding layer: time-to-pulse conversion;

[0163] Shared latent space feature vectors Encoded as a pulse sequence. A time-to-pulse interval encoding strategy is employed.

[0164]

[0165] in, The time when the i-th input neuron first fires a pulse; The maximum latency is 50ms, which is set based on the balance between edge device clock accuracy and response speed. , used to Map to the interval (0,1);

[0166] (2) Liquid Pool: High-dimensional dynamic state evolution;

[0167] Construct a recurrent connection network consisting of N=500 spiking neurons, called the "liquid pool". Each neuron follows the Leaky Integrate-and-Fire model:

[0168]

[0169] in, =10ms: Membrane time constant, reflecting the neuron's ability to "memorize" input; =1.0MΩ: Input resistance, normalized; Input connection weights, from the 128-dimensional input layer to the liquid pool, are initialized using a normal distribution; : Recursive connection weights, sparse random connections (connection probability p=30%), simulating the sparsity of biological neural networks; Input pulse signal; Neuronal pulses inside the liquid pool.

[0170] The core function of the liquid pool is to map the input pulse pattern to a high-dimensional dynamic state space, whose instantaneous state is determined by the membrane potential and pulse history of all neurons. Even if the input is static, it can generate rich dynamic responses, enhance sensitivity to subtle differences, and is particularly suitable for distinguishing between "mild" and "moderate" corrosion.

[0171] (3) Readout Layer: State Decoding and Classification

[0172] After the pulse input ends (i.e., t=100 ms), the neural activity of the liquid pool is collected within the time window t∈[50,100] ms, and the following statistical features are extracted to form the readout vector f. :

[0173] First 500 dimensions: Average firing frequency per neuron ;

[0174] Dimension 501: Population Synchronization Index ,in This represents the phase of neuronal activity, reflecting group synergy;

[0175] 502nd dimension: membrane potential variance , characterizing the activity level of the system.

[0176] These features are concatenated into a vector, which is then input into a lightweight linear classifier: ,in This represents the predicted probability distribution of corrosion levels.

[0177] Final output corrosion level:

[0178] .

[0179] The classification task is defined as a four-class classification problem, and the output corrosion level is as follows:

[0180] Level 0: No rust, smooth metal surface of grounding down conductor with no visible oxidation marks, stable guided wave signal, low reconstruction error, suitable for newly commissioned or well-maintained towers.

[0181] Level 1: Slight corrosion (section loss <20%), with scattered reddish-brown rust spots on the surface, without forming a continuous rust layer; the guided wave echo is slightly distorted, the phase shift is small, and the main frequency drifts slightly; this condition does not affect the conductivity and is recommended to be included in the routine inspection checklist.

[0182] Level 2: Moderate corrosion (section loss 20%~50%), with continuous rust patches and localized peeling; the guided wave signal shows obvious reflected wave packets, and the amplitude attenuation rate A... r >0.3, phase offset ∆ ∅ >0.5π indicates a significant increase in wavelet energy entropy; this condition has affected the grounding resistance, and it is recommended to arrange for replacement or anti-corrosion treatment within 3 months.

[0183] Level 3: Severe corrosion (section loss >50% or cracks present), the metal surface is severely corroded, with local honeycomb or pitted cavitation, and some areas are left with only a thin shell; the guided wave signal shows multiple reflections and scattering, the main frequency is severely deviated, and the reconstruction error is extremely high; this state poses a risk of breakage and requires immediate power outage and replacement of the down conductor.

[0184] Step S5: Visualization and Alerts

[0185] To facilitate operations and maintenance personnel's intuitive understanding of the test results and quick decision-making, this step visualizes the classification results and issues tiered alarms through multiple channels.

[0186] First, a corrosion heatmap overlay image is generated, using the original visible light image as the base image. Based on the location of the detection points, corresponding areas are marked on the image. Then, using color mapping technology, the corrosion level is converted into a visual heatmap.

[0187] Level 0: Green (RGB: 0, 255, 0)

[0188] Level 1: Yellow (RGB: 255, 255, 0)

[0189] Level 2: Orange (RGB: 255, 165, 0)

[0190] Level 3: Red (RGB: 255, 0, 0)

[0191] The heat map is overlaid on the corresponding area of ​​the original image in a semi-transparent manner, clearly showing the location and severity of corrosion; at the same time, metadata such as timestamp, tower number, detection height, and confidence level are marked to form a complete inspection report image.

[0192] Subsequently, a multi-level alarm mechanism was activated: Local alarm: The results were indicated by LED indicator lights or buzzers at the detection equipment end - a solid green light indicates normal operation, and a flashing red light indicates severe corrosion; Remote alarm: The detection results, heat map and raw data packets were uploaded to the intelligent operation and maintenance platform of the transmission line via 4G / 5G or power private network.

[0193] Level 0~1: Records and archives data; no alerts required.

[0194] Level 2: The platform generates a "medium-risk warning work order" and pushes it to the regional operation and maintenance APP, reminding the user to handle it within a specified period.

[0195] Level 3: Triggers a "high-risk emergency alarm", automatically notifies the operations and maintenance manager, and links the GIS system to locate the tower position. It is recommended to activate the emergency plan.

[0196] The alarm information includes fields such as: tower ID, GPS coordinates, corrosion level, recommended measures, inspection time, and responsible person, and supports seamless integration with the work order system.

[0197] In addition, all inspection data is automatically archived to a cloud database, supporting historical trend analysis. For example, periodic scanning of the same inspection point can generate a "reconstruction error-time" curve to monitor the corrosion development trend, realizing a shift from "passive maintenance" to "predictive maintenance".

[0198] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of this invention is defined by the appended claims and their equivalents.

Claims

1. A cross-modal identification method for grounding down conductor corrosion based on shared latent space construction, characterized in that, Includes the following steps: Step S1: Multi-source data synchronous acquisition, synchronously acquire electromagnetic ultrasonic guided wave echo signals and visible light image data at the same spatial location, establish a spatiotemporal alignment benchmark, and ensure the correspondence between the two types of data at the detection location; Step S2: Single-mode feature extraction, performing independent feature modeling for electromagnetic ultrasonic guided wave signals and visible light images respectively; Step S3: Cross-modal feature alignment and fusion; Step S4: Corrosion level identification based on neurodynamic system. Based on fusion features, a neurodynamic classifier based on spiking neural network and liquid machine is proposed to replace the traditional SVM or MLP model, so as to realize automatic, dynamic, low-power classification and quantitative evaluation of the corrosion degree of grounding down conductor. Step S5: Visualization and Alarms. Visualize the classification results and issue tiered alarms through multiple channels.

2. The cross-modal identification method for grounding down conductor corrosion based on shared latent space construction according to claim 1, characterized in that: The multi-source data synchronous acquisition in step S1 includes establishing a precise mapping between the "probe center point" and the "image coordinate system origin" so that the subsequently acquired guided wave signals and image data strictly correspond in spatial position; configuring and starting the electromagnetic ultrasonic guided wave detection system to acquire high-quality echo signals; starting the visible light imaging module and completing the setting of image acquisition parameters; the electromagnetic ultrasonic guided wave detector outputs a TTL level trigger pulse at the same time as each excitation signal is emitted, and this pulse is connected to the external trigger input port of the visible light camera as the main synchronization signal, forcing the camera to synchronously capture a frame of image at the moment of guided wave emission, so that each group of guided wave signals and its corresponding visible light image are strictly aligned in time.

3. The cross-modal identification method for grounding down conductor corrosion based on shared latent space construction according to claim 1, characterized in that: In step S2, single-mode feature extraction first involves using Hilbert transform to extract the envelope curve of the analytic signal of the guided wave echo signal, and then calculating key physical characteristic parameters, including the attenuation rate of the echo amplitude relative to the reference signal. Phase offset between the reflected wave and the reference wave The offset of the main frequency component And the energy entropy of the signal under multi-scale wavelet decomposition. Finally, the above four features are combined to form a four-dimensional acoustic feature vector. =[ , , , [This is used as a feature representation of the guided wave mode; simultaneously, visual feature extraction is performed on the synchronously acquired visible light images. First, the images are normalized and input into a pre-trained convolutional neural network based on the ResNet-18 architecture. The first few layers of the convolutional modules are frozen to retain general edge and texture features. The high-order semantic feature vector output by the penultimate fully connected layer is extracted.] This achieves the mapping from pixel space to semantic feature space; finally, and Both types of multi-source feature vectors are standardized to eliminate dimensional differences.

4. The cross-modal identification method for grounding down conductor corrosion based on shared latent space construction according to claim 1, characterized in that, The cross-modal feature alignment and fusion in step S3 includes: S3.1 Network Structure Design – Acoustic Coding Branch; S3.2 Network Structure Design – Visual Encoding Branch; S3.3 Shared Semantic Mapping Layer Design; S3.4 Dual Decoding Branch: Feature Reconstruction Network Based on U-Net Architecture; S3.5 Loss Function Design: Joint Optimization Objective; S3.6 Corrosion Criterion Generation: Reconstruction Error as an Anomaly Indicator; S3.7 Fusion Feature Output.

5. The cross-modal identification method for grounding down conductor corrosion based on shared latent space construction according to claim 4, characterized in that, The S3.1 network structure design—the acoustic coding branch—is an acoustic coding branch based on the diffusion mechanism and Transformer architecture, which maps low-dimensional guided wave features to a highly discriminative latent space representation. The specific implementation process is as follows: (1) Forward noise addition process: Constructing a multi-scale noise perturbation sequence; Let the original guided wave eigenvector be... It includes four physical quantities: amplitude attenuation rate, phase shift, dominant frequency shift, and wavelet energy entropy. During the training phase, a forward noise addition process is defined to simulate potential electromagnetic interference, contact noise, and temperature drift in the field. N(0,1)； Where t=1,2,...,T is the number of diffusion steps, and T is set to a total of 50 diffusion steps. This value can achieve a balance between computational efficiency and noise coverage. ； ； in A cosine noise scheduling strategy is adopted. =0.008 is the offset coefficient; This represents the characteristic state after being disturbed by noise at step t; (2) Inverse denoising encoder: Feature evolution network based on temporal conditional Transformer; Define an inverse denoising encoder, starting from the noise-contaminated... The clean features associated with the corrosion state are gradually restored in the middle; The encoder uses a Transformer architecture as its core network. The specific structure is as follows: Input concatenation: Combine current noise features diffusion steps t and original features Concatenate into joint input: ； in It is the sinusoidal position code for time step t, used to provide time information; Through a fully connected layer Mapped to The dimensional embedding space is used as the input to the Transformer; Transformer encoder layer: Stacks L=3 layers of standard Transformer blocks, each layer containing: four multi-head self-attention modules, query / key / value dimension. Two fully connected layers, with an intermediate dimension of 512, and GELU activation function; residual connections and layer normalization; the output of the last Transformer layer is then subjected to global average pooling, followed by dimensionality reduction through two MLP layers to output the final acoustic latent space representation: ； It is defined as: a low-dimensional implicit representation that is highly correlated with the internal structural damage of the grounding lead, which is evolved from the guided wave characteristics of noise disturbance by a time-conditional Transformer network during the diffusion denoising process.

6. The cross-modal identification method for grounding down conductor corrosion based on shared latent space construction according to claim 4, characterized in that, The S3.2 network structure design—visual encoding branch—refers to the input being a 512-dimensional image feature vector. Based on the ResNet-18 model, the encoding branch employs multi-layer nonlinear transformations to compress features into a 256-dimensional visual latent space. .

7. The cross-modal identification method for grounding down conductor corrosion based on shared latent space construction according to claim 4, characterized in that, The S3.3 shared semantic mapping layer design introduces a shared semantic mapping layer and configures two independent decoding branches to form a complete cross-modal feature reconstruction path. This design maps heterogeneous features from different modalities to the same semantic space, ensuring the fidelity of each feature while establishing a potential consistency relationship between "internal structural state" and "surface visual representation." The acoustic latent space representation extracted in steps S3.1 and S3.2 is then used to... Representation of visual latent space Each input to the shared semantic mapping layer is mapped to a unified 128-dimensional shared latent space Z using a learnable linear transformation matrix. To achieve cross-modal semantic alignment; ； ； The mapping weight matrix representing the acoustic modes; The mapping weight matrix representing the visual modality; , Represents the bias vector; , These represent the projection vectors of the waveguide features and the image features in the shared latent space Z, respectively. Introduce cross-modal alignment loss during training: ； The two types of heterogeneous features are forced to be semantically consistent and should be highly aligned under normal conditions.

8. The cross-modal identification method for grounding down conductor corrosion based on shared latent space construction according to claim 4, characterized in that, The S3.4 dual decoding branch refers to setting up two independent decoders based on the U-Net architecture, respectively from... and Reconstruct the original input features to form an end-to-end reconstruction path; (1) Acoustic decoder based on U-Net guided wave feature reconstruction network: The input is treated as a "feature image" and a one-dimensional U-Net structure is used to reconstruct the original 4-dimensional guided wave feature vector. = ； in, The reconstructed waveguide characteristics are used to calculate the reconstruction error with the true values; (2) Visual decoder based on U-Net image semantic feature reconstruction network: A standard two-dimensional U-Net is used to reconstruct the image semantic features from... Reconstruct the semantic feature vector of the original 512-dimensional image. = ； The output is compared with the original visual features extracted by ResNet-18 to calculate the reconstruction error of the image modality.

9. The cross-modal identification method for grounding down conductor corrosion based on shared latent space construction according to claim 4, characterized in that, (1) The S3.5 loss function design defines a multi-task loss function to drive the network to simultaneously complete the two major objectives of feature reconstruction and cross-modal alignment: ； To measure whether the waveguide characteristics are accurately recovered during the encoding-decoding process; To measure whether image features are fully preserved; Forced acoustic implicit representation Visual latent representation They tend to agree before even entering the shared space; Weighting coefficient The importance of balancing the three losses; (2) The corrosion criterion S3.6 is generated. In actual testing, a new "acoustic-image" sample is input, and its total reconstruction error is calculated: ； Reconstruction error It is used as a confidence index for corrosion.

10. The cross-modal identification method for grounding down conductor corrosion based on shared latent space construction according to claim 1, characterized in that: In step S4, the corrosion level identification is based on a neurodynamic system. The structural design of the neurodynamic classifier includes... (1) Input coding layer: time-to-pulse conversion; Shared latent space feature vectors Encoded as a pulse sequence, using a time-to-pulse interval encoding strategy: ； in, The time when the i-th input neuron first fires a pulse; The maximum latency is 50ms, which is set based on the balance between edge device clock accuracy and response speed. , used to Map to the interval (0,1); (2) Liquid Pool: High-dimensional dynamic state evolution; Construct a recurrent connection network consisting of N=500 spiking neurons, called the "liquid pool"; each neuron follows the Leaky Integrate-and-Fire model: ； in, =10ms: Membrane time constant, reflecting the neuron's ability to "memorize" input; =1.0MΩ: Input resistance, normalized; Input connection weights, from the 128-dimensional input layer to the liquid pool, are initialized using a normal distribution; : Recursive connection weights, sparse random connections (connection probability p=30%), simulating the sparsity of biological neural networks; Input pulse signal; Neuronal impulses inside the liquid pool; (3) Readout Layer: State decoding and classification; After the pulse input ends, neural activity in the liquid pool is collected within a time window of t∈[50,100] ms, and the following statistical features are extracted to form the readout vector f. : First 500 dimensions: Average firing frequency per neuron ； Dimension 501: Population Synchronization Index ,in This represents the phase of neuronal activity, reflecting group synergy; 502nd dimension: membrane potential variance Characterizes the activity level of the system; These features are concatenated into a vector, which is then input into a lightweight linear classifier: ； in The predicted probability distribution for corrosion level; The final output is the corrosion level.

Images