A method and system for adaptive connection of multi-specification leads of surge arresters

Abstract

This invention relates to the field of line installation technology and discloses an adaptive docking method and system for multi-specification surge arrester leads. The method includes: determining the specification type and end oxidation degree of the leads to be docked using a multi-modal sensing model based on acquired multi-modal data; the multi-modal sensing model is configured to fuse lead image data and eddy current signal data based on an attention mechanism, and to use radial stiffness data as a guiding signal to weight and enhance the fusion result of the attention mechanism; matching the corresponding docking process procedure from a process database based on the specification type, end oxidation degree, and radial stiffness data from the multi-modal data; and performing cleaning processing based on a target preprocessing strategy and crimping operation based on target crimping process parameters according to the docking process procedure to obtain the docking result. The method provided in this application improves the adaptability and crimping quality of surge arrester lead docking.

Description

Technical Field

[0001] This invention relates to the field of line installation technology, and in particular to a method and system for adaptive connection of multi-specification leads of surge arresters. Background Technology

[0002] Surge arresters are core equipment for overvoltage protection in power systems. Through the volt-ampere characteristics of nonlinear resistive elements, they rapidly discharge surge currents to ground during lightning strikes or switching overvoltages, thus ensuring the insulation safety of transmission and transformation equipment. The connection leads between the surge arrester and the transmission line serve as the sole channel for overvoltage energy release, and their connection quality directly affects the reliability of the protection. Therefore, the lead connections must meet stringent requirements for low resistance, high mechanical strength, and resistance to environmental corrosion.

[0003] In existing engineering projects, surge arrester lead splicing still relies heavily on manual operation and experience-based judgment. On one hand, leads exhibit diverse specifications; their diameter, material, and cross-sectional shape vary depending on the voltage level and application scenario, lacking a unified standard for process parameter adaptation. On the other hand, leads operating outdoors for extended periods typically have an oxide layer at their ends, the thickness and density of which directly affect contact resistance, requiring targeted cleaning before splicing. Furthermore, differences in radial stiffness inherent in the leads themselves lead to varying metal deformation behavior during crimping, thus affecting connection tightness and long-term reliability. Traditional methods struggle to comprehensively quantify and evaluate these multi-dimensional influencing factors, often relying on fixed process parameters or adjustments based on experience, resulting in poor consistency in crimping quality and frequent occurrences of issues such as incomplete connections and overheating.

[0004] Therefore, improving the adaptability and crimping quality of surge arrester lead connections has become a technical problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] This invention provides a method and system for adaptive connection of multi-specification leads of surge arresters, in order to solve the technical problem of how to improve the adaptability and crimping quality of surge arrester lead connections.

[0006] In a first aspect, the present invention provides an adaptive docking method for multi-specification leads of a surge arrester, the method comprising: acquiring multi-modal data of the leads to be docked, the multi-modal data comprising: lead image data, eddy current signal data and radial stiffness data;

[0007] Based on the multimodal data, a pre-constructed multimodal sensing model is used to determine the specification type and end oxidation degree of the lead to be docked. The multimodal sensing model is set to fuse the lead image data and the eddy current signal data based on an attention mechanism, and to use the radial stiffness data as a guiding signal to weight and enhance the fusion result of the attention mechanism.

[0008] Based on the specification type, the degree of end oxidation, and the radial stiffness data, a corresponding butt welding process procedure is matched from a preset process database. The butt welding process procedure includes at least: target crimping process parameters, target dynamic impedance curve, and target pretreatment strategy for different degrees of oxidation.

[0009] According to the docking process specification, a cleaning process based on the target pretreatment strategy and a pressing operation based on the target pressing process parameters are performed to obtain the docking result. The pressing operation adopts a hybrid control strategy based on force feedback and impedance dual feedback, and controls the pressing process based on the collected real-time dynamic impedance curve and the target dynamic impedance curve.

[0010] Preferably, acquiring the multimodal data of the lead to be connected includes:

[0011] An industrial camera is used to acquire image data of the lead wire to be connected, and the lead wire image data includes the end of the lead wire to be connected and a preset length range;

[0012] Eddy current sensor is used to collect eddy current signal data of the lead to be connected. The eddy current signal data is the voltage characteristic sequence across the probe coil of the eddy current sensor. The voltage characteristic sequence includes material conductivity characteristics and surface condition characteristics. The material conductivity characteristics include the base value and amplitude of the voltage characteristic sequence. The surface condition characteristics include the mean, standard deviation, peak-to-peak value, and the number of signal fluctuations per unit length corresponding to the main frequency of the voltage characteristic sequence.

[0013] The eddy current sensor is used to collect the deformation data of the lead to be connected under a known radial force, and the radial stiffness data of the lead to be connected is obtained based on the radial force and the deformation data.

[0014] Preferably, the multimodal perception model includes: an image processing module, an eddy current processing module, a stiffness-guided fusion module, and an output module; the image processing module is constructed using a deep convolutional neural network, the eddy current processing module is constructed using a one-dimensional convolutional neural network, a Transformer encoder, and a feature pooling layer, and the stiffness-guided fusion module is constructed using a cross-modal attention fusion mechanism and a stiffness-guided weighted fusion network.

[0015] Preferably, determining the specification type and end oxidation degree of the lead to be connected using a pre-constructed multimodal sensing model based on the multimodal data includes:

[0016] The image processing module is used to sequentially perform feature extraction, cross-layer connection and multi-scale fusion processing on the lead image data to obtain a visual feature vector, which includes geometric contour features and color texture features.

[0017] The one-dimensional convolutional neural network is used to extract local features from the eddy current signal data to obtain the local fluctuation pattern and frequency response features of the voltage feature sequence; the Transformer encoder is used to extract global features from the position-encoded local fluctuation pattern and the position-encoded frequency response features to obtain global dependent features; the feature pooling layer is used to process the global dependent features to obtain the eddy current feature vector.

[0018] Based on the cross-modal attention fusion mechanism, the visual feature vector is linearly transformed into a query matrix, and the eddy current feature vector is linearly transformed into a key matrix and a value matrix. A visual enhancement feature vector is obtained based on the query matrix, the key matrix, and the value matrix. Furthermore, the weighted fusion network is used to perform element-wise multiplication of the eddy current feature vector and the radial stiffness data to obtain a stiffness-guided enhancement feature vector. The visual feature vector, the visual enhancement feature vector, and the stiffness-guided enhancement feature vector are then weighted and fused to obtain a weighted enhancement feature vector.

[0019] Based on the weighted enhanced feature vector, the specifications and end oxidation degree of the lead to be connected are obtained through the parallel output branch of the output module.

[0020] Preferably, the method for constructing the process database includes:

[0021] Based on the experimental data of crimping process of surge arrester lead samples with different specifications, different degrees of end oxidation, and different radial stiffness, an experimental dataset was obtained.

[0022] A multiphysics coupling model is constructed using simulation software, and the multiphysics coupling model is calibrated using the experimental dataset. Based on the calibrated multiphysics coupling model, a surrogate model is constructed through simulation design, and an extended dataset is generated based on the surrogate model.

[0023] Based on the experimental dataset and the expanded dataset, a process database is constructed. The process database adopts a multi-level index structure, wherein the first-level index is the specification type layer, the second-level index is the end oxidation degree layer, and the third-level index is the radial stiffness layer.

[0024] Preferably, the step of matching the corresponding docking process specification from a preset process database based on the specification type, the degree of end oxidation, and the radial stiffness data includes:

[0025] Based on the first-level index, using the specification type as the primary key, all docking process procedures applicable to the specification type are retrieved from the process database to obtain a first process procedure candidate set.

[0026] Based on the degree of end oxidation, a target pretreatment strategy is determined from the pretreatment strategy set in the process database, and the degree of oxidation matching between the lead to be docked and each docking process in the first process procedure candidate set is calculated.

[0027] Based on the oxidation degree matching degree, a docking process procedure with an oxidation degree matching degree greater than a preset oxidation degree matching threshold is selected from the first process procedure candidate set to obtain a second process procedure candidate set.

[0028] Based on the radial stiffness data, the deviation between the radial stiffness data and the radial stiffness applicable range of each docking process in the second process specification candidate set is calculated to obtain the radial stiffness deviation. Based on the radial stiffness deviation, the pressing process parameters of each docking process in the second process specification candidate set are linearly interpolated and corrected to obtain the corrected pressing process parameters of each docking process in the second process specification candidate set.

[0029] Based on the oxidation degree matching degree, the corrected pressing process parameters are weighted and fused to obtain the target pressing process parameters.

[0030] Based on the oxidation degree matching degree and the radial stiffness deviation, a comprehensive similarity index is obtained, and based on the comprehensive similarity index, a target quality monitoring index is determined from the second process specification candidate set. The target quality monitoring index includes: target dynamic impedance curve, mechanical pull-out force threshold, and contact resistance upper limit.

[0031] Preferably, the step of performing cleaning treatment based on the target pretreatment strategy and pressing operation based on the target pressing process parameters according to the docking process specification to obtain the docking result includes:

[0032] According to the target pretreatment strategy of the docking process procedure, determine the cleaning tools and perform the cleaning process;

[0033] According to the target crimping process parameters in the aforementioned docking process specification, a visually guided positioning strategy is adopted to perform positioning processing.

[0034] Based on the target crimping process parameters and the target dynamic impedance curve, a hybrid control strategy based on force feedback and electrical impedance dual feedback is adopted to perform the crimping operation.

[0035] Based on the real-time dynamic impedance curve collected after crimping and the target dynamic impedance curve, the joint is automatically detected to obtain the mating result.

[0036] Preferably, the step of performing the crimping operation based on a hybrid control strategy using force feedback and impedance dual feedback according to the target crimping process parameters and the target dynamic impedance curve includes:

[0037] Based on the target pressing force and holding time in the aforementioned docking process specification, the desired force curve is obtained;

[0038] Based on the deviation between the real-time pressing force and the expected pressing force from the expected force curve, the desired position of the robotic arm performing the pressing operation is corrected.

[0039] Based on the impedance deviation between the target dynamic impedance curve and the acquired real-time impedance, the crimping force of the crimping operation performing robotic arm is corrected.

[0040] Preferably, the method further includes:

[0041] After the crimping is completed, a corresponding digital record of the crimping is obtained based on the specification type, the degree of end oxidation, the radial stiffness data, and the collected real-time docking parameters and quality inspection results.

[0042] The digital record of the crimping process is used as feedback input to iteratively optimize the process database, so that the process parameters of the process database are dynamically updated.

[0043] Secondly, the present invention also provides an adaptive docking system for multi-specification leads of surge arresters, used to realize the adaptive docking method for multi-specification leads of surge arresters described above. The system includes: a data acquisition unit, a multi-modal recognition unit, a process specification matching unit, and a docking execution unit.

[0044] The data acquisition unit is used to acquire multimodal data of the lead wire to be connected, including lead wire image data, eddy current signal data, and radial stiffness data.

[0045] The multimodal recognition unit is used to determine the specification type and end oxidation degree of the lead to be docked based on the multimodal data using a pre-constructed multimodal perception model. The multimodal perception model is set to fuse the lead image data and the eddy current signal data based on an attention mechanism, and to use the radial stiffness data as a guiding signal to weight and enhance the fusion result of the attention mechanism.

[0046] The process matching unit is used to match the corresponding butt welding process from a preset process database according to the specification type, the degree of end oxidation and the radial stiffness data. The butt welding process includes at least: target crimping process parameters, target dynamic impedance curve and target pretreatment strategy for different degrees of oxidation.

[0047] The docking execution unit is used to perform cleaning processing based on the target pretreatment strategy and crimping operation based on the target crimping process parameters according to the docking process specification, so as to obtain the docking result. The crimping operation adopts a hybrid control strategy based on force sensing and electrical impedance dual feedback, and controls the crimping process based on the collected real-time dynamic impedance curve and the target dynamic impedance curve.

[0048] This invention provides a method and system for adaptive connection of multi-specification leads in surge arresters. Compared with the prior art, the embodiments of this invention have the following advantages:

[0049] This application provides an adaptive docking method for multi-specification surge arrester leads. It constructs a multimodal perception model based on an attention mechanism, achieving simultaneous and accurate identification of surge arrester lead specifications and end oxidation levels through the collaborative work of an image processing module, an eddy current processing module, and a stiffness-guided fusion module. In conventional attention fusion, this application introduces radial stiffness data as a guiding signal, providing a physical constraint anchor point for the multimodal perception model. Since radial stiffness directly reflects the mechanical properties of the lead and is strongly correlated with physical states such as lead material density and porosity caused by oxidation, using it as a guiding signal forces the attention mechanism to focus on the areas in image and eddy current features most relevant to mechanical performance degradation. This avoids misjudgments by the multimodal perception model based solely on visual color and eddy current baseline values, achieving deeper semantic alignment between modes. This application's multimodal perception model not only solves the semantic alignment problem during multimodal data fusion but also enhances its robustness and interpretability by injecting prior knowledge of radial stiffness. In specific scenarios such as dust, oil stains, and changes in lighting, data from different modes may point to contradictory results. Stiffness variation is a comprehensive reflection of the material's internal structure and is not easily affected by surface coatings. The multimodal perception model dynamically adjusts the fusion weights of image features and eddy current features based on radial stiffness data, prioritizing information highly correlated with stiffness features. This significantly improves the model's recognition accuracy and anti-interference capability in unstructured environments. It also concretizes abstract quality requirements into monitoring dynamic impedance curves, providing a quantifiable benchmark for real-time process monitoring and improving crimping quality. Attached Figure Description

[0050] Figure 1 This is a schematic diagram of the steps of a preferred embodiment of the present invention for an adaptive connection method of multi-specification leads for a surge arrester;

[0051] Figure 2 This is a schematic diagram of the overall architecture of a multimodal sensing model provided in a preferred embodiment of the present invention;

[0052] Figure 3 This is a schematic diagram of a multi-specification lead adaptive docking system for surge arresters provided in a preferred embodiment of the present invention. Detailed Implementation

[0053] The embodiments of the present invention are described in detail below with reference to the accompanying drawings. The embodiments are provided for illustrative purposes only and should not be construed as limiting the scope of the invention. The accompanying drawings are for reference and illustration only and do not constitute a limitation on the scope of protection of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of this invention.

[0054] In the description of this invention, it should be noted that, unless otherwise defined, all technical and scientific terms used in this invention have the same meaning as commonly understood by one of ordinary skill in the art. The terminology used in this specification is for the purpose of describing specific embodiments only and is not intended to limit the invention. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0055] Please see Figure 1 In an embodiment of the present invention, an adaptive connection method for multi-specification leads of a surge arrester is provided, the method comprising:

[0056] S1. Obtain multimodal data of the leads to be connected. The multimodal data includes: lead image data, eddy current signal data, and radial stiffness data. Compared with ordinary conductors, surge arrester leads need to meet specific requirements in terms of electrical performance and mechanical strength: In terms of material specifications, copper core insulated wire is generally used. The cross-sectional area of ​​the upper lead should be greater than or equal to 16 square millimeters, and the cross-sectional area of ​​the lower lead should be greater than or equal to 25 square millimeters. In terms of laying path, it must be short and straight, avoiding coiling and bending. In terms of connection requirements, terminal crimping or welding is used, and the contact surface is clean and coated with conductive grease. In terms of mechanical flexibility, flexible connections or expansion joints should be used to avoid direct connection of rigid busbars. In terms of stiffness characteristics, the lead must have sufficient tensile strength and fatigue resistance to prevent breakage under wind, vibration, or earthquake. The surge arrester lead should be able to withstand the electrodynamic force when a short-circuit current passes through without melting or being torn off. In a preferred embodiment of this application, to comprehensively perceive the external physical morphology, surface condition, and internal mechanical properties of the leads to be connected, and to provide a data foundation for subsequent accurate identification and adaptive process matching, it is necessary to simultaneously collect lead image data, eddy current signal data, and radial stiffness data. The adaptive connection method for multi-specification surge arrester leads of this application is applied to a dedicated surge arrester lead self-connection device. A high-resolution industrial camera and an LED ring light source are mounted on the side of the end effector of the six-degree-of-freedom robotic arm of the surge arrester lead self-connection device. After the robotic arm moves to the predetermined observation point at the lead connection station, image acquisition is triggered. The ring light source is then controlled to uniformly illuminate the lead end at a preset brightness. The industrial camera takes pictures perpendicular to the lead axis, acquiring clear images including the lead end and a preset length range (e.g., 20 mm), thus obtaining lead image data. The acquired lead image data is an RGB color image or a grayscale image, mainly used for subsequent analysis of the lead's geometric contour, end color characteristics, and physical damage to the insulation layer or conductor surface.

[0057] In a preferred embodiment of this application, the end effector also integrates a high-precision eddy current sensor, such as the German Mi-Iddy NCDT 3010 (3mm probe diameter, 1mm linear range) or other eddy current sensor modules with similar or superior performance. The eddy current displacement sensor is also mounted on the end effector of the six-degree-of-freedom robotic arm, achieving time-division multiplexing through the movement of the robotic arm. After image acquisition, the six-degree-of-freedom robotic arm is moved to vertically align the eddy current sensor probe with the conductor end face or side near the end of the lead to be connected, maintaining the initial gap between the probe and the lead surface at the midpoint of the eddy current sensor's linear range. A high-frequency excitation signal is applied to the eddy current sensor probe coil. The alternating magnetic field generated by the probe induces eddy current signals on the metal surface of the lead. These eddy current signals react on the probe coil, changing its equivalent impedance, thereby causing a change in the voltage across the probe coil. The base value and fluctuation of the voltage signal across the probe coil imply the electromagnetic characteristics and displacement information of the lead surface material. By amplifying and filtering the voltage signal across the probe coil, a voltage characteristic sequence reflecting the material's inherent properties is obtained. This voltage characteristic sequence is closely related to the oxidation degree of the lead surface to be connected. The voltage characteristic sequence includes material conductivity characteristics and surface state characteristics. Material conductivity characteristics include the base value and amplitude of the voltage characteristic sequence, which are closely related to the oxidation degree of the lead. Surface state characteristics include the mean, standard deviation, peak-to-peak value, and the number of signal fluctuations per unit length corresponding to the dominant frequency of the voltage characteristic sequence. The mean of the voltage characteristic sequence reflects the average oxide layer thickness on the lead surface; the thicker the oxide layer, the greater the equivalent distance between the probe and the substrate, and the smaller the mean value usually is. The standard deviation of the voltage characteristic sequence reflects the overall uniformity of the oxide layer on the lead surface; the higher the standard deviation, the worse the overall uniformity of the oxide layer. The peak-to-peak value of the voltage characteristic sequence reflects the difference in oxide layer thickness on the lead surface; severe localized corrosion bulges or peeling pits on the lead surface will lead to an increase in the peak-to-peak value. The number of signal fluctuations per unit length corresponding to the main frequency reflects the density of the oxide layer on the surface of the lead wire to be connected. During the measurement process, the probe scans laterally along the lead wire surface. Therefore, the spatial frequency of the signal fluctuation is related to the roughness of the surface morphology. Dense, fine particles cause high-frequency fluctuations, while loose, large particles cause low-frequency fluctuations.

[0058] The radial stiffness data in this application are the deformation data of the lead wire to be connected under a known force, measured by an eddy current sensor. Specifically, the end effector integrates a miniature linear motor and connects it in series with a force sensor to apply a precise radial force to the lead wire to be connected via servo control. This radial force range is much smaller than the yield strength of the lead wire material, such as 0.5N~5N, ensuring that the radial stiffness data measurement is performed within the elastic range and can cover the stiffness differences of leads with different cross-sections. The eddy current sensor is calibrated before measurement to establish a relationship curve between output voltage and displacement. Based on the real-time acquired voltage signal, it is converted into a displacement value d in real time through this calibration curve. The force value F and the displacement value d output by the eddy current sensor are recorded during the force application process. The Fd curve is fitted using the least squares method, and its slope is the radial contact stiffness of the part of the lead wire to be connected, which is used as the radial stiffness data.

[0059] In a preferred embodiment of this application, high-resolution industrial cameras are used to acquire lead wire image data, clearly revealing the geometric contours and end color characteristics of the leads, providing an intuitive basis for determining lead wire specifications and macroscopic physical damage. Simultaneously, by combining eddy current signal data and analyzing the original characteristic voltage, the surface oxide layer or stains can be penetrated to perceive changes in the conductivity and permeability of the lead wire surface material, thereby accurately assessing the degree of end oxidation. The combination of macroscopic and microscopic data overcomes the limitation of a single sensor in simultaneously identifying surface conditions and internal material properties. For surge arrester leads, mechanical strength and electrical performance are equally important. Traditional methods struggle to obtain lead wire stiffness information online and non-destructively. This application uses eddy current sensors to acquire deformation data of the leads to be connected under a known radial force, calculating radial stiffness data and providing crucial mechanical parameters for subsequent matching and connection processes.

[0060] S2. Based on the multimodal data, a pre-constructed multimodal sensing model is used to determine the specification type and end oxidation degree of the lead to be connected. The multimodal sensing model is configured to fuse the lead image data and the eddy current signal data based on an attention mechanism, and to use the radial stiffness data as a guiding signal to weight and enhance the fusion result of the attention mechanism. In a preferred embodiment of this application, the multimodal sensing model adopts a multimodal fusion neural network based on an attention mechanism, including: an image processing module, an eddy current processing module, a stiffness guiding fusion module, and an output module. The overall architecture of the multimodal sensing model is as follows: Figure 2 As shown.

[0061] In a preferred embodiment of this application, the image processing module is used to process the lead wire image data. The image processing module is constructed using a deep convolutional neural network, specifically an improved CSPDarkNet53 network (Cross-Stage Local Connections DarkNet53) as the backbone feature extraction network. The CSPDarkNet53 network includes an input layer, an initial convolutional layer, a DarkNet53 backbone network, and feature pyramid enhancement. The initial convolutional layer uses 32 3×3 convolutional kernels for preliminary feature extraction with a stride of 2 to achieve downsampling. The DarkNet53 backbone network contains five stages of CSP (Cross-Stage Local Connections) modules, each stage consisting of multiple residual units. Gradient fusion is achieved through cross-stage connections, reducing computation while maintaining feature richness. Feature pyramid enhancement, located after the CSPDarkNet53 backbone network, performs multi-scale fusion of feature maps at different scales to accommodate the size differences of lead wires of different specifications. The image processing module outputs a visual feature vector, which encodes geometric contour features and color texture features. Geometric profile features mainly include geometric information such as the diameter, cross-sectional shape, and insulation layer boundary of the lead wire, used to identify the lead wire specification type. Color and texture features mainly include the color distribution at the lead wire end, oxide mottle texture, and surface roughness, used to preliminarily determine the degree of oxidation.

[0062] In a preferred embodiment of this application, the eddy current processing module is used to process eddy current signal data, specifically voltage feature sequences. The eddy current processing module employs a hybrid network based on a one-dimensional convolutional neural network and a Transformer encoder, including: a one-dimensional convolutional layer, a position encoding layer, a Transformer encoder, and a feature pooling layer. The one-dimensional convolutional layer comprises three layers of one-dimensional convolutional neural networks with kernel sizes of 7, 5, and 3, respectively. This one-dimensional convolutional layer is used to extract local features from the voltage feature sequence, capturing local fluctuation patterns and frequency response features. The feature sequence output from the one-dimensional convolutional layer is input to the position encoding layer, where sinusoidal position encoding is used to add absolute position information, and then input to two Transformer encoder layers. Each Transformer encoder layer includes a multi-head attention mechanism and a feedforward neural network to capture long-range dependencies and global patterns in the voltage feature sequence, obtaining global dependency features. The feature pooling layer transforms the global dependency features into a fixed-dimensional feature vector through global average pooling, obtaining the eddy current feature vector.

[0063] The stiffness-guided fusion module uses radial stiffness data as a guiding signal to fuse visual feature vectors and eddy current feature vectors. Specifically, based on a cross-modal attention fusion mechanism, the visual feature vectors are linearly transformed into a query matrix, and the eddy current feature vectors are linearly transformed into a key matrix and a value matrix. The attention weights of the visual feature vectors to the eddy current feature vectors are calculated using a cross-attention mechanism, and the value matrices of the eddy current feature vectors are weighted and summed. This allows the multimodal perception model to focus on the eddy current response region most relevant to the visual content. The calculation formula is as follows:

[0064]

[0065] in, , , , Represents the query vector. Represents the key vector. Represents a value vector. Represents visual feature vectors. Represents the eigenvector of eddy currents. , and Represents a learnable linear transformation matrix. This represents the dimension of the key vector. This represents the vector transpose symbol.

[0066] The output of the cross-modal attention fusion mechanism is a visual enhancement feature vector. The visual enhancement feature vector takes visual information as a guide and dynamically aggregates the electric eddy current signal response corresponding to the image region, thereby improving the accuracy of material oxidation degree assessment.

[0067] There is a significant interaction between the degree of oxidation at the ends of the surge arrester's mating leads and the radial stiffness of the leads; the degree of oxidation at the ends alters the radial stiffness of the leads. Specifically, severe oxidation consumes the base metal, leading to a reduction in the effective load-bearing cross-sectional area of ​​the mating leads and a decrease in the radial stiffness of the leads. Some oxides are inherently brittle, or the oxidation process can cause intergranular corrosion, resulting in damage to the microstructure of the surface material of the mating leads. This makes them more susceptible to microcracks when subjected to radial forces, thereby reducing the overall elastic modulus or bending resistance. Uneven oxide pits can form microscopic gaps on the lead surface. When the lead is bent, stress concentration occurs at these gaps, leading to local yielding or cracking, which macroscopically manifests as a nonlinear change in stiffness or a decrease in overall stiffness. Therefore, to improve the accuracy of end oxidation degree identification, radial stiffness data is used as a guiding signal to weight and enhance the attention fusion result. Specifically, a weighted fusion network is used to multiply the eddy current feature vector and radial stiffness data element-wise to obtain a stiffness-guided enhancement feature vector. The visual feature vector, visual enhancement feature vector, and stiffness-guided enhancement feature vector are then weighted and fused to obtain a weighted enhancement feature vector. The calculation formula for the stiffness-guided weighted fusion network is as follows:

[0068]

[0069] in, This represents a weighted enhanced feature vector. Representation layer normalization, Represents radial stiffness data. and This represents the learnable fusion weight parameters. This indicates element-wise multiplication. Represents a learnable linear transformation matrix. This represents the stiffness-guided enhancement eigenvector.

[0070] The weighted enhancement method employed in this application directly involves radial stiffness data in the interaction, multiplying it element-wise with the eddy current feature vector to generate stiffness-guided enhancement features, achieving fine-grained interaction across modalities. The final output is a weighted sum of the visual feature vector, the visual enhancement feature vector, and the stiffness-guided enhancement feature vector, allowing the guiding signal to influence the extraction of the eddy current feature vector earlier and preserving original information through residual connections, preventing the loss of details in deep networks. The radial stiffness of the lead depends not only on the geometry but also on the Young's modulus of the material, which is closely related to the material's microstructure (such as lattice defects, dislocation density, and oxide brittleness). Changes in these microstructures also affect the material's conductivity, i.e., the baseline and fluctuations of the voltage feature sequence. The weighted enhancement method using radial stiffness data as the guiding signal provided in this application is designed specifically for the scenario described in this application, enabling the multimodal perception model to output more accurate specifications and end oxide levels of the lead to be docked.

[0071] In a preferred embodiment of this application, based on a weighted enhanced feature vector, the output module is configured with two parallel output branches to respectively complete specification type identification and end oxidation degree identification. The specification type identification branch is a multi-classification task, employing a two-layer fully connected network. Based on the structural standards and actual application scenarios of the surge arrester leads, the specification types include:

[0072] Top lead: 16mm², 25mm², 35mm², 50mm² copper core insulated wire;

[0073] Bottom lead: 25mm², 35mm², 50mm², 70mm² copper core insulated wire.

[0074] The last fully connected layer of the specification type identification branch uses the Softmax function (normalized exponential function) to output the probability distribution of each category. The category with the highest probability is taken as the identified specification type, and a confidence score is also output for the reliability assessment of subsequent process matching.

[0075] The task type of the end oxidation degree recognition branch is ordered multi-class classification. It adopts a two-layer fully connected network, with the last layer being a linear layer, to quantify the oxidation degree as a continuous value between 0 and 1.

[0076] Furthermore, the constructed multimodal sensing model was trained, and multimodal data of a large number of surge arrester lead samples with different specifications and oxidation levels were collected using the multimodal data sampling method of this application. The surge arrester lead samples were labeled, with the specification type labeled using a combination of manual labeling based on the lead nameplate and the measured cross-sectional area; the degree of oxidation at the ends was labeled by obtaining objective quantitative values ​​through metallographic microscopy, energy dispersive spectroscopy analysis, or electrochemical impedance spectroscopy measurement.

[0077] In a preferred embodiment of this application, the total loss function of the multimodal perception model for multi-task learning is a weighted sum of the loss function for the specification type recognition branch and the loss function for the end oxidation degree recognition branch, specifically expressed as follows:

[0078]

[0079] in, This represents the total loss function of the multimodal sensing model. The cross-entropy loss represents the branch used for specification type identification. The mean square error loss representing the degree of end oxidation is used to identify branches. The weight hyperparameters representing the specification type identification branch are automatically determined through grid search. The weight hyperparameters representing the degree of oxidation at the ends of the branches are automatically determined through grid search.

[0080] Furthermore, the image data processing branch is pre-trained on a large-scale visual dataset to accelerate convergence. Then, based on the multimodal data and corresponding labels of the above-mentioned lightning arrester lead samples, a training dataset is constructed. The training dataset is used to perform end-to-end joint training of the multimodal perception model as a whole, realizing feature sharing and task collaborative optimization, and obtaining the trained multimodal perception model.

[0081] In practical applications of adaptive docking of surge arrester self-connected leads, real-time acquired multimodal data is input into a trained multimodal sensing model, which outputs the specifications and end oxidation degree of the lead to be docked, for example: specifications, top lead -25mm. 2 The degree of end oxidation is 0.35 (moderate oxidation).

[0082] In a preferred embodiment of this application, the constructed attention-based multimodal perception model achieves simultaneous and accurate identification of the surge arrester lead specifications and end oxidation levels through the collaborative work of the image processing module, eddy current processing module, and stiffness-guided fusion module. In conventional attention fusion, this application introduces radial stiffness data as a guiding signal, providing a physical constraint anchor point for the multimodal perception model. Since radial stiffness directly reflects the mechanical properties of the lead and is strongly correlated with physical states such as material density and porosity caused by oxidation, using it as a guiding signal forces the attention mechanism to focus on the areas in image and eddy current features most relevant to mechanical performance degradation. This avoids misjudgments by the multimodal perception model based solely on visual color and eddy current baseline values, achieving deeper semantic alignment between modes. This multimodal perception model not only solves the semantic alignment problem during multimodal data fusion but also enhances its robustness and interpretability by injecting prior knowledge of radial stiffness. In specific scenarios such as dust, oil stains, and changes in lighting, data from different modalities may point to contradictory results. Stiffness variation is a comprehensive reflection of the material's internal structure and is not easily affected by surface coatings. The multimodal perception model dynamically adjusts the fusion weights of image features and eddy current features based on radial stiffness data, prioritizing information highly correlated with stiffness features. This significantly improves the model's recognition accuracy and anti-interference capability in unstructured environments. Radial stiffness data itself has clear physical meaning. When used as a guiding signal to enhance the attention map, it is equivalent to injecting prior physical knowledge into the neural network. The generated attention heatmap not only reflects data features but also the degree of potential physical damage. This makes the identification of lead type and end oxidation degree in this application no longer isolated data-driven but based on the deep integration of physical properties and appearance features, laying a solid perceptual foundation for the subsequent high-reliability adaptive docking of surge arrester leads.

[0083] S3. Based on the specification type, the degree of end oxidation, and the radial stiffness data, a corresponding butt welding process procedure is matched from a preset process database. The butt welding process procedure includes at least: target crimping process parameters, a target dynamic impedance curve, and a target pretreatment strategy for different degrees of oxidation. In a preferred embodiment of this application, based on the specification type, end oxidation degree, and radial stiffness data of the lead to be butted, a butt welding process procedure most suitable for the state of the lead to be butted is matched from a preset process database to guide subsequent adaptive butt welding operations. The process database of this application is not a simple set of empirical parameters, but a knowledge base built based on a large amount of process experimental data, theoretical simulation analysis, and real-time field data. In a preferred embodiment of this application, the construction of the process database includes the following steps:

[0084] Extensive crimping process experiments were conducted on surge arrester lead samples of different specifications, degrees of end oxidation, and radial stiffness. The crimping process parameters and corresponding quality monitoring indicators were recorded under each condition. A hexagonal crimping die was used for the crimping operation. The crimping process parameters included at least: target crimping force, crimping depth, holding time, and crimping speed. The quality monitoring indicators included at least: target dynamic impedance curve, mechanical pull-out force threshold, and upper limit of contact resistance. During the experiments, the surge arrester lead samples were installed on a dedicated crimping fixture and crimped according to preset parameters. Each set of parameters was repeated five times, and the mechanical pull-out force, contact resistance, and dynamic impedance curve after crimping were recorded. Through orthogonal experimental design, the optimal process parameters that maximized the pull-out force and minimized the contact resistance under each condition were selected, forming an experimental dataset. Each experimental data point in the dataset includes the applicable specification type, degree of end oxidation, radial stiffness characteristics, crimping process parameters, mechanical pull-out force after crimping, contact resistance during crimping, and dynamic impedance curve. Furthermore, a multiphysics coupling model for adaptive connection of surge arrester leads, constructed using simulation software, is employed. Experimental datasets are input into this model for calibration, resulting in a surrogate model. Based on this surrogate model, an expanded dataset is generated. Specifically, firstly, a three-dimensional model of lead terminal crimping is established in ANSYS (a leading global engineering simulation software widely used in structural mechanics, fluid dynamics, electromagnetic fields, and thermal analysis), considering the helical structure of the wire strands, the insulation peeling zone, and the geometric characteristics of the crimping mold. Next, a bilinear elastoplastic model is used for modeling. The yield strength and tangent modulus of the bilinear elastoplastic model vary with the degree of end oxidation and are calibrated through nanoindentation testing. The oxide layer is considered an isotropic brittle layer, assigned a corresponding thickness and elastic modulus. Finally, frictional contacts between wire strands and between wire strands and terminals are established. The friction coefficient is determined by the degree of oxidation; for example, 0.3 for light oxidation and 0.5 for heavy oxidation. Finally, based on electrical contact theory, the shrinkage resistance of the contact spots after crimping deformation is calculated. Simultaneously, the Joule heating effect caused by the temperature rise during the large current flow at the moment of crimping is considered, and the material parameters are corrected. Based on the experimental dataset, the mapping relationship between crimping process parameters and quality monitoring indicators is fitted to obtain a surrogate model. This surrogate model is then used to simulate and complete the experimentally uncovered conditions, generating an expanded dataset. Box plots are used to identify outliers in the mechanical pull-out force and contact resistance in the expanded dataset. Combined with post-crimping cross-sectional metallographic analysis, it is determined whether these outliers are caused by crimping defects; if so, the corresponding expanded data entry needs to be deleted.

[0085] Furthermore, based on the experimental and extended datasets, a process database is constructed, employing a multi-level index structure. The first-level index is the specification type layer, enumerating all specification types of the surge arrester's self-connecting leads. The second-level index is the end oxidation degree layer, using discretized intervals, such as 0-0.2 (none / light), 0.2-0.5 (medium), 0.5-0.8 (heavy), and 0.8-1.0 (severe). The third-level index is the radial stiffness layer. Each docking process specification in this application's process database is encapsulated using an object-oriented data structure. Examples of the core fields included in the docking process specification are as follows:

[0086] Docking process specification ID: PROC-25-03-085

[0087] Applicable conditions: {Specification type: top lead - 25mm², applicable range of end oxidation degree: [0.2, 0.5], applicable range of radial stiffness: [75, 95], unit N / mm}

[0088] Target preprocessing strategy: {Strategy type: Mechanical grinding}

[0089] Target crimping process parameters: {Target crimping force: 2.8kN; Crimping depth: 4.2mm; Holding time: 1.5s; Crimping speed: 2.0mm / s}

[0090] Target quality monitoring indicators: {Target dynamic impedance curve: Impedance curve ID: IMP-25-03, Mechanical pull-out force threshold: 1.2kN; Upper limit of contact resistance: 50μΩ}.

[0091] Pretreatment strategy is a key component of the docking process. Its design fully considers the special requirements of the surge arrester leads to thoroughly remove the oxide layer without damaging the conductor substrate or reducing its mechanical strength. Table 1 shows an example of the mapping between the degree of oxidation and the pretreatment strategy.

[0092] Table 1

[0093]

[0094] The conductive grease cleaning strategy is executed by a precision grease applicator at the end of the robotic arm, which includes a grease reservoir, a metering pump, and a scraper head. It automatically calculates the required amount of conductive grease based on the lead wire specifications. The formula for calculating the required amount of conductive grease is as follows:

[0095]

[0096] in, This indicates the amount of conductive grease applied per unit area. This indicates the perimeter of the lead cross-section. Indicates the coating length.

[0097] Furthermore, the metering pump is precisely controlled to extrude the conductive grease, and the scraper head contacts the lead surface at a 45° angle, moving uniformly along the axial direction at a speed of 5 mm / s to ensure a uniform coating without bubbles. After coating, the surface condition is re-measured using an eddy current sensor to ensure that the quantified value of the end oxidation level is reduced to below the first oxidation level target value.

[0098] The mechanical polishing strategy is executed by a miniature mechanical polishing head equipped at the end of the robotic arm, which can be fitted with different grit sandpaper, such as 400 grit, 600 grit, and 800 grit. Optimized polishing parameter combinations are pre-stored for different lead wire specifications, as shown in Table 2, which provides examples of polishing parameter combinations.

[0099] Table 2

[0100]

[0101] During the polishing process, a constant force control mode is employed. A six-dimensional force sensor at the end of the mechanical polishing head provides real-time feedback on the contact force, dynamically adjusting the robotic arm's posture to ensure constant polishing pressure and prevent damage to the leads from localized over-polishing. Force control utilizes admittance control to ensure stable polishing force and avoid lead damage. After polishing, the robotic arm switches to the eddy current sensor probe to rescan the polished area, verifying whether the quantified value of the end oxidation level has decreased below the second oxidation level target value. If not, supplementary polishing is automatically performed. The second oxidation level target value is greater than the first oxidation level target value to match the actual effect after mechanical polishing.

[0102] The laser cleaning strategy is executed by a pulsed fiber laser cleaning head integrated into the end effector of a robotic arm, with a wavelength of 1064nm and a maximum average power of 50W. The high-energy laser beam irradiates the lead surface; the oxide layer has a high laser absorption rate, causing it to expand, vaporize, or peel off instantly upon heating, while the metal substrate has a high laser reflectivity and absorbs less energy, thus achieving selective removal without damaging the substrate. The power density of the pulsed fiber laser cleaning head is automatically adjusted according to the degree of oxidation; the higher the degree of oxidation, the higher the power density.

[0103] Based on power density and lead diameter, the pulsed fiber laser cleaning head automatically plans a spiral scanning path to ensure coverage of the entire end area, with an overlap rate of ≥30% between adjacent spots. After laser cleaning, the robotic arm switches to the eddy current sensor probe to rescan the laser-cleaned area, verifying whether the quantified value of the end oxidation level has dropped below the third oxidation level target value. If not, supplementary cleaning is automatically performed. The third oxidation level target value is greater than the second oxidation level target value to match the actual effect after laser cleaning.

[0104] When the oxidation degree quantification value exceeds 0.8, it is determined that the lead wire has been severely oxidized and the substrate may have been damaged. At this time, an oxidation over-limit alarm is output to prompt the operator to check.

[0105] If further processing is still necessary, employ a combined cleaning strategy: first, laser cleaning to remove the loose surface oxide layer; then, mechanical polishing to expose the fresh metal; and finally, protective conductive grease. Record the lead information in the quality traceability system to enhance quality monitoring.

[0106] The target dynamic impedance curve of this application includes annotations of key feature points, which include: initial impedance, inflection point 1 (elastic limit), inflection point 2 (plastic flow initiation point), plateau region, final stable impedance, and springback, as shown in the following example:

[0107] Curve ID: IMP-25-03

[0108] Applicable conditions: {Specification type: top lead - 25mm², applicable range of end oxidation degree: [0.2, 0.5], applicable range of radial stiffness: [75, 95], unit N / mm}

[0109] Key feature points:

[0110] Initial impedance: Z0 = 120 mΩ (t=0s)

[0111] Inflection point 1 (elastic limit): t1 = 0.3s, Z1 = 45 mΩ

[0112] Inflection point 2 (start of plastic flow): t2 = 0.8 s, Z2 = 28 mΩ

[0113] Plateau region: t=0.8s-1.8s, Z value fluctuates smoothly between 25mΩ and 28mΩ.

[0114] Final steady-state impedance: Z∞ = 26 mΩ (t=2.0s)

[0115] Springback amount: ΔZ = 2 mΩ.

[0116] In a preferred embodiment of this application, a multi-level matching strategy is adopted to match the docking process specifications based on specification type, end oxidation degree, and radial stiffness data. The first level involves coarse screening of specification types, using the identified specification types as the primary key to quickly retrieve all docking process specifications applicable to that specification type from the process database, forming a first candidate set of process specifications. .

[0117] In a preferred embodiment of this application, the degree of oxidation matching is used as the basis for the pretreatment strategy, and the target pretreatment strategy is determined from the pretreatment strategy set in the process database.

[0118] The second level involves matching the degree of end oxidation. Based on the degree of end oxidation, the results are compared with the first set of candidate process procedures. The oxidation degree matching degree of each docking process specification is calculated using the following formula:

[0119]

[0120] in, This represents the docking process specification index of the first set of candidate process specifications. Indicates the first candidate process specification. The degree of oxidation matching between the docking process specifications This indicates the measured degree of oxidation at the ends of the leads to be connected. Indicates the first candidate process specification. The midpoint of the applicable range for the degree of end oxidation in each docking process specification. Indicates the first candidate process specification. The applicable half-width range of the end oxidation degree for each docking process specification. Based on the oxidation degree matching degree, from the first set of process specification candidates... The docking process procedures with an oxidation degree matching degree greater than the preset oxidation degree matching threshold are selected to obtain the second process procedure candidate set. .

[0121] The third level involves adaptive stiffness adjustment. This involves calculating the deviation between the measured radial stiffness data and the applicable radial stiffness range for each butt joint process specification in the second set of candidate process specifications. The radial stiffness deviation is then used to linearly interpolate and correct the pressing process parameters of the butt joint process specification based on this deviation. The correction formula is as follows:

[0122]

[0123] in, This represents the docking process specification index of the second set of candidate process specifications. Indicates the first in the second process specification candidate set The revised crimping process parameters for each butt welding procedure include: target crimping force, crimping depth, holding time, and crimping speed. Indicates the first in the second process specification candidate set The basic crimping process parameters of the first butt welding process specification, namely the first... The crimping process parameters recommended in the specific butt welding process specifications. This represents the measured radial stiffness data of the current lead wire. Indicates the first in the second process specification candidate set The median value of the applicable range of radial stiffness for each docking process specification. Indicates the first in the second process specification candidate set The stiffness-crimping process parameter sensitivity coefficient of each butt joint process specification represents the amount of compensation required by the crimping process parameters for a unit change in stiffness. Each crimping process parameter is different and is calibrated through experimental data.

[0124] Based on the degree of oxidation matching, all modified pressing process parameters in the second process specification candidate set are weighted and fused to obtain the target pressing process parameters. The calculation formula for the target pressing process parameters is as follows:

[0125]

[0126] in, Indicates the target crimping process parameters. Indicates the first in the second process specification candidate set The degree of oxidation matching of each docking process specification.

[0127] For the selection of target quality monitoring indicators, a comprehensive similarity index combining oxidation matching degree and stiffness deviation is adopted. The calculation formula of the comprehensive similarity index is as follows:

[0128]

[0129] in, Indicates the first in the second process specification candidate set The comprehensive similarity index of the docking process specifications The weights representing the degree of matching in oxidation are... The weights representing radial stiffness deviations are fitted based on historical data. Indicates the first in the second process specification candidate set The applicable half-width of the radial stiffness of each docking process specification is determined. The quality monitoring index of the docking process specification with the highest comprehensive similarity index in the second set of candidate process specifications is selected as the target quality monitoring index.

[0130] If the final matching result is empty, that is, there is no fully matching docking process procedure, the system will automatically switch to the conservative process mode. The conservative process mode adopts a specific smaller pressure connection force and a longer pressure holding time, and will alarm in the system to prompt manual review.

[0131] In a preferred embodiment of this application, a multi-level indexed process database is constructed, and based on the specification type and oxidation degree output by the multimodal sensing model, as well as radial stiffness data, hierarchical and refined matching is performed. This upgrades the fixed butt welding process procedure to an adaptive butt welding process procedure, achieving precise adaptation of the crimping process procedure and significantly improving the consistency of crimping quality for leads in different states. For the continuous variable of radial stiffness data, linear interpolation correction is used instead of simple discrete matching, allowing the crimping process parameters to be dynamically adjusted according to the deviation between the actual radial stiffness value and the applicable range in the procedure. This solves the problem of parameter mismatch in non-standard working conditions caused by the traditional table lookup method. By constructing and calibrating a multi-physics coupling model, and combining it with simulation design to generate an expanded dataset, the limitations of the limited physical experimental samples are effectively compensated for, enabling the process database to cover a wider range of lead specifications, oxidation degrees, and stiffness characteristic combinations.

[0132] S4. According to the docking process specifications, perform cleaning treatment based on the target pretreatment strategy and crimping operation based on the target crimping process parameters to obtain the docking result. The crimping operation adopts a hybrid control strategy based on force feedback and impedance dual feedback, and controls the crimping process based on the collected real-time dynamic impedance curve and the target dynamic impedance curve. In a preferred embodiment of this application, the multi-degree-of-freedom robotic arm has high-precision position control and force control interfaces, and a quick-change device is installed at the end to switch tools. The end effector is an integrated multi-functional actuator head, which integrates at least: an adaptive clamp for holding the lead wire, an adjustable depth cutter for rotating and stripping the insulation layer, a cleaning tool for cleaning the conductor, a hydraulic or electric crimping module for crimping operations, and a micro-resistance measurement contact for detecting the connection quality.

[0133] In a preferred embodiment of this application, different cleaning actions are performed based on a pre-processing strategy. The robotic arm switches between corresponding cleaning tools via a quick-change device to perform the corresponding cleaning operation. After cleaning, a positioning operation is further performed. In this preferred embodiment, a vision-guided positioning strategy is adopted. An industrial camera identifies the approximate position of the lead wire, and the end effector of the robotic arm is moved to a first distance threshold near the lead wire. Through vision guidance, the motor is controlled to move precisely, accurately measuring the three-dimensional coordinates and orientation of the lead wire to be connected, and adjusting the robotic arm to align the crimping connector with the center line of the lead wire. When the crimping connector reaches a second threshold distance from the lead wire, the robotic arm is switched to force control mode, and it slowly approaches with a small force until contact force is detected, the contact position is recorded, and the positioning is completed.

[0134] After positioning, the crimping operation is further performed through the crimping joint of the end effector. The crimping operation adopts a hybrid control strategy based on force feedback and impedance dual feedback, using impedance as the basis for process monitoring and state switching. The real-time dynamic impedance curve is a direct feedback of the evolution of the electrical contact state throughout the crimping process. The shape of the real-time dynamic impedance curve directly reflects the formation quality and mechanical stability of the crimped joint. The real-time dynamic impedance curve records the dynamic change of contact impedance over time from the moment the probe contacts the lead wire to the completion of crimping and the end of the pressure holding process. During the crimping process, as the crimping force increases, the contact interface between the lead wire and the terminal undergoes the following changes: In the initial contact stage, only a few micro-protrusions are in contact, resulting in high contact resistance and high impedance; in the elastic deformation stage, the micro-protrusions are flattened, the contact area increases, and the impedance decreases rapidly; in the plastic deformation stage, the base metal undergoes plastic flow, the interface forms metallic bonds, and the impedance tends to stabilize; in the pressure holding stage, the interface stress relaxes, and elastic rebound causes a slight increase in impedance.

[0135] In a preferred embodiment of this application, the control phases of the crimping operation are divided into: rapid approach, contact positioning, crimping forming, pressure holding, and release. In the rapid approach phase, the wire moves quickly to the vicinity of the lead at a preset first speed without making contact; in this phase, only the position is controlled, and no feedback is provided. In the contact positioning phase, the wire slowly approaches the lead to be crimped at a preset second speed until the contact force reaches a threshold. In the crimping forming phase, pressure is applied according to the desired force curve. During the crimping process, as the crimping force increases, the contact interface between the lead and the terminal undergoes the following changes: In the initial contact phase, only a few micro-protrusions are in contact, resulting in high contact resistance and high impedance; in the elastic deformation phase, the micro-protrusions are flattened, the contact area increases, and the impedance decreases rapidly; in the plastic deformation phase, the base metal undergoes plastic flow, and metallic bonding forms at the interface, causing the impedance to stabilize; in the pressure holding phase, the interface stress relaxes, and elastic rebound causes a slight increase in impedance.

[0136] In a preferred embodiment of this application, the target pressing force is determined according to the process specifications. and holding time Generate the expectation force curve For example, in the pressing stage, the desired force is to increase linearly to the target pressing force. During the holding pressure phase, the expected force remains constant. ; During the release phase, the force is withdrawn and the expected force is restored. The following is an example of the desired force curve in a preferred embodiment of this application:

[0137] During the press forming stage, the desired force curve is expressed as:

[0138]

[0139] in, Indicates the time it takes for the expectation force to rise.

[0140] During the pressure holding phase, the expected force curve is represented as:

[0141]

[0142] The duration of the pressure holding phase is as follows: .

[0143] During the release phase, the expectation force curve is represented as:

[0144]

[0145] in, Indicates the duration of the release phase. This indicates the total duration of the crimping operation.

[0146] The control architecture for the crimping operation includes inner-loop control, outer-loop control, and impedance monitoring. The inner-loop control is the position servo control of the robotic arm, receiving the desired position command. The outer-loop control is an admittance controller that corrects the desired position of the robotic arm based on the deviation between the real-time crimping force and the desired crimping force from the desired force curve, achieving crimping force tracking. Impedance monitoring compares the collected real-time impedance with the target dynamic impedance curve in the docking process specification. The resulting impedance deviation is used to adjust the crimping force of the robotic arm performing the crimping operation or to trigger an alarm. Specifically, the outer-loop control calculates the desired position correction amount at the end of the robotic arm based on the deviation between the real-time crimping force and the desired crimping force from the desired force curve, using the following formula:

[0147]

[0148] in, This represents the desired position correction amount at the end of the robotic arm. The expected pressing force represents the expected force curve. This indicates the real-time crimping force during the crimping process. Represents the coefficient of inertia. Indicates the damping coefficient. These represent the stiffness coefficient, inertia coefficient, damping coefficient, and rigidity coefficient, which can be adjusted according to different stages. This represents the Laplace operator.

[0149] In this application, the desired position correction amount of the robotic arm end effector during the crimping process is represented in a discretized form, as shown below:

[0150]

[0151] in, express Position correction amount at any given time. express Position correction amount at any given time. and Indicated by admittance parameters , , and the discretization coefficients determined by the sampling period, express Real-time pressure at all times.

[0152] In impedance monitoring, impedance feedback and outer loop control are integrated. During the contact positioning stage, when the real-time impedance value collected jumps from infinity to a finite value, it is considered that contact has been made, the progress stops and the process transitions to the pressing and forming stage.

[0153] During the pressing and pressure holding stages, the impedance deviation value is calculated in real time. The calculation formula is as follows:

[0154]

[0155] in, express The impedance deviation value at time 10:00. express Real-time impedance value at time t. This represents the target impedance value of the target dynamic impedance curve at the corresponding time.

[0156] Furthermore, if , If the threshold value is set by the docking process specification, the crimping force of the crimping operation robot arm is adjusted to compensate for the actual impedance and bring it closer to the target value of the target dynamic impedance curve. If the target impedance value cannot be reached after compensation, an alarm is triggered, the crimping is paused, and the abnormality is recorded.

[0157] To enable the process database to be adaptive, this application designs a closed-loop feedback update mechanism: after each crimping, the identified specification type, end oxidation degree, matching docking process procedure, and collected real-time docking parameters and final quality inspection results are stored as a complete process case in a temporary database. Cluster analysis is periodically performed on the accumulated process cases to identify outliers with identical process parameters but abnormal quality results. When a sufficient number of new process cases are accumulated, incremental learning of the process database is triggered, updating the mapping relationship between crimping process parameters and quality monitoring indicators, and optimizing the docking process procedures in the process database. Each update retains a version number to ensure traceability; for critical process changes, manual review is required before they take effect.

[0158] In existing technologies, the quality of the crimping result is generally predicted based on the comparison between the real-time mechanical pull-out force and a preset mechanical pull-out force threshold. In the preferred embodiment of this application, the similarity between the target dynamic impedance curve and the dynamic impedance curve during the crimping process is strongly correlated with quality indicators such as the mechanical pull-out force and contact resistance after crimping. Therefore, after crimping is completed, this application calculates the similarity between the target dynamic impedance curve and the dynamic impedance curve, compares it with a preset similarity threshold to obtain the impedance curve similarity, and combines the pull-out force comparison result to comprehensively evaluate the quality of the crimped joint and obtain the mating result.

[0159] In a preferred embodiment of this application, a cleaning strategy is dynamically selected based on the degree of end oxidation, ensuring cleaning effectiveness while avoiding substrate damage. Abstract quality requirements are concretized into monitoring dynamic impedance curves, providing a quantifiable benchmark for real-time process monitoring. Through a closed-loop feedback mechanism, the process database possesses continuous learning capabilities, adapting to changes in materials and environment, and achieving dynamic optimization of the process knowledge base.

[0160] Accordingly, such as Figure 3 As shown, based on an adaptive docking method for multi-specification leads of surge arresters, this embodiment of the invention also provides an adaptive docking system for multi-specification leads of surge arresters, used to implement the adaptive docking method for multi-specification leads of surge arresters disclosed in this embodiment of the invention. The system includes: a data acquisition unit, a multi-modal recognition unit, a process specification matching unit, and a docking execution unit.

[0161] The data acquisition unit is used to acquire multimodal data of the lead wire to be connected, including lead wire image data, eddy current signal data, and radial stiffness data.

[0162] The multimodal recognition unit is used to determine the specification type and end oxidation degree of the lead to be docked based on the multimodal data using a pre-constructed multimodal perception model. The multimodal perception model is set to fuse the lead image data and the eddy current signal data based on an attention mechanism, and to use the radial stiffness data as a guiding signal to weight and enhance the fusion result of the attention mechanism.

[0163] The process matching unit is used to match the corresponding butt welding process from a preset process database according to the specification type, the degree of end oxidation and the radial stiffness data. The butt welding process includes at least: target crimping process parameters, target dynamic impedance curve and target pretreatment strategy for different degrees of oxidation.

[0164] The docking execution unit is used to perform cleaning processing based on the target pretreatment strategy and crimping operation based on the target crimping process parameters according to the docking process specification, so as to obtain the docking result. The crimping operation adopts a hybrid control strategy based on force sensing and electrical impedance dual feedback, and controls the crimping process based on the collected real-time dynamic impedance curve and the target dynamic impedance curve.

[0165] Specific limitations regarding the adaptive docking system for multi-specification leads of surge arresters can be found in the above-described limitations regarding the adaptive docking method for multi-specification leads of surge arresters, and will not be repeated here. Those skilled in the art will recognize that the various modules and steps described in conjunction with the embodiments disclosed in this invention can be implemented in hardware, software, or a combination of both. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this invention.

[0166] This embodiment provides a method and system for adaptive docking of multi-specification leads of surge arresters, which is used to solve the technical problem. The method includes: acquiring multimodal data of the leads to be docked, including lead image data, eddy current signal data, and radial stiffness data; determining the specification type and end oxidation degree of the leads to be docked using a pre-constructed multimodal sensing model based on the multimodal data, wherein the multimodal sensing model is configured to fuse lead image data and eddy current signal data using an attention mechanism, and to use radial stiffness data as a guiding signal to weight and enhance the fusion result of the attention mechanism; matching the corresponding docking process procedure from a pre-set process database based on the specification type, end oxidation degree, and radial stiffness data, wherein the docking process procedure includes at least the target crimping process parameters, the target dynamic impedance curve, and the target pretreatment strategy for different oxidation degrees; performing cleaning treatment based on the target pretreatment strategy and crimping operation based on the target crimping process parameters according to the docking process procedure to obtain the docking result, wherein the crimping operation adopts a hybrid control strategy based on force feedback and impedance dual feedback, and controls the crimping process based on the acquired real-time dynamic impedance curve and the target dynamic impedance curve. The adaptive docking method for multi-specification surge arrester leads provided in this application constructs a multimodal perception model based on an attention mechanism. Through the collaborative work of an image processing module, an eddy current processing module, and a stiffness-guided fusion module, it achieves simultaneous and accurate identification of surge arrester lead specification types and end oxidation degrees. In conventional attention fusion, this application introduces radial stiffness data as a guiding signal, providing a physical constraint anchor point for the multimodal perception model. Since radial stiffness directly reflects the mechanical performance of the lead, it is strongly correlated with the lead's material density and the porosity caused by oxidation. Using it as a guiding signal forces the attention mechanism to focus on the areas in image features and eddy current features most relevant to mechanical performance degradation, avoiding misjudgments by the multimodal perception model based solely on visual color and eddy current baseline values, and achieving deeper semantic alignment between modes. This application's multimodal perception model not only solves the semantic alignment problem during multimodal data fusion but also enhances the robustness and interpretability of the multimodal perception model by injecting prior knowledge of radial stiffness. In specific scenarios such as dust, oil stains, and changes in lighting, data from different modes may point to contradictory results. Stiffness variation is a comprehensive reflection of the material's internal structure and is not easily affected by surface coatings. The multimodal perception model dynamically adjusts the fusion weights of image features and eddy current features based on radial stiffness data, prioritizing information highly correlated with stiffness features. This significantly improves the model's recognition accuracy and anti-interference capability in unstructured environments. It also concretizes abstract quality requirements into monitoring dynamic impedance curves, providing a quantifiable benchmark for real-time process monitoring and improving crimping quality.

[0167] The various embodiments in this specification are described in a progressive manner. For directly identical or similar parts of the embodiments, refer to each other. Each embodiment focuses on describing the differences from other embodiments. In particular, the system embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions in the method embodiments. It should be noted that the technical features of the above embodiments can be combined arbitrarily. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as the combination of these technical features does not contradict each other, it should be considered within the scope of this specification.

[0168] The above-described embodiments are merely preferred embodiments of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the invention. It should be noted that those skilled in the art can make various improvements and substitutions without departing from the principles of the present invention, and these improvements and substitutions should also be considered within the scope of protection of the present invention. Therefore, the scope of protection of this invention should be determined by the scope of the claims.

Claims

1. A method for adaptive connection of multi-specification leads of surge arresters, characterized in that, The method includes: Acquire multimodal data of the lead wire to be connected, the multimodal data including: lead wire image data, eddy current signal data and radial stiffness data; Based on the multimodal data, a pre-constructed multimodal sensing model is used to determine the specification type and end oxidation degree of the lead to be docked. The multimodal sensing model is set to fuse the lead image data and the eddy current signal data based on an attention mechanism, and to use the radial stiffness data as a guiding signal to weight and enhance the fusion result of the attention mechanism. Based on the specification type, the degree of end oxidation, and the radial stiffness data, a corresponding butt welding process procedure is matched from a preset process database. The butt welding process procedure includes at least: target crimping process parameters, target dynamic impedance curve, and target pretreatment strategy for different degrees of oxidation. According to the docking process specification, a cleaning process based on the target pretreatment strategy and a pressing operation based on the target pressing process parameters are performed to obtain the docking result. The pressing operation adopts a hybrid control strategy based on force feedback and impedance dual feedback, and controls the pressing process based on the collected real-time dynamic impedance curve and the target dynamic impedance curve.

2. The adaptive connection method for multi-specification leads of surge arresters as described in claim 1, characterized in that, The acquisition of multimodal data of the leads to be connected includes: An industrial camera is used to acquire image data of the lead wire to be connected, and the lead wire image data includes the end of the lead wire to be connected and a preset length range; Eddy current sensor is used to collect eddy current signal data of the lead to be connected. The eddy current signal data is the voltage characteristic sequence across the probe coil of the eddy current sensor. The voltage characteristic sequence includes material conductivity characteristics and surface condition characteristics. The material conductivity characteristics include the base value and amplitude of the voltage characteristic sequence. The surface condition characteristics include the mean, standard deviation, peak-to-peak value, and the number of signal fluctuations per unit length corresponding to the main frequency of the voltage characteristic sequence. The eddy current sensor is used to collect the deformation data of the lead to be connected under a known radial force, and the radial stiffness data of the lead to be connected is obtained based on the radial force and the deformation data.

3. The adaptive connection method for multi-specification leads of surge arresters as described in claim 2, characterized in that, The multimodal perception model includes an image processing module, an eddy current processing module, a stiffness-guided fusion module, and an output module. The image processing module is constructed using a deep convolutional neural network. The eddy current processing module is constructed using a one-dimensional convolutional neural network, a Transformer encoder, and a feature pooling layer. The stiffness-guided fusion module is constructed using a cross-modal attention fusion mechanism and a stiffness-guided weighted fusion network.

4. The adaptive connection method for multi-specification leads of surge arresters as described in claim 3, characterized in that, The step of determining the specification type and end oxidation degree of the lead to be connected using a pre-constructed multimodal sensing model based on the multimodal data includes: The image processing module is used to sequentially perform feature extraction, cross-layer connection and multi-scale fusion processing on the lead image data to obtain a visual feature vector, which includes geometric contour features and color texture features. The one-dimensional convolutional neural network is used to extract local features from the eddy current signal data to obtain the local fluctuation pattern and frequency response features of the voltage feature sequence; the Transformer encoder is used to extract global features from the position-encoded local fluctuation pattern and the position-encoded frequency response features to obtain global dependent features; the feature pooling layer is used to process the global dependent features to obtain the eddy current feature vector. Based on the cross-modal attention fusion mechanism, the visual feature vector is linearly transformed into a query matrix, and the eddy current feature vector is linearly transformed into a key matrix and a value matrix. A visual enhancement feature vector is obtained based on the query matrix, the key matrix, and the value matrix. Furthermore, the weighted fusion network is used to perform element-wise multiplication of the eddy current feature vector and the radial stiffness data to obtain a stiffness-guided enhancement feature vector. The visual feature vector, the visual enhancement feature vector, and the stiffness-guided enhancement feature vector are then weighted and fused to obtain a weighted enhancement feature vector. Based on the weighted enhanced feature vector, the specifications and end oxidation degree of the lead to be connected are obtained through the parallel output branch of the output module.

5. The adaptive connection method for multi-specification leads of surge arresters as described in claim 1, characterized in that, The method for constructing the process database includes: Based on the experimental data of crimping process of surge arrester lead samples with different specifications, different degrees of end oxidation, and different radial stiffness, an experimental dataset was obtained. A multiphysics coupling model is constructed using simulation software, and the multiphysics coupling model is calibrated using the experimental dataset. Based on the calibrated multiphysics coupling model, a surrogate model is constructed through simulation design, and an extended dataset is generated based on the surrogate model. Based on the experimental dataset and the expanded dataset, a process database is constructed. The process database adopts a multi-level index structure, wherein the first-level index is the specification type layer, the second-level index is the end oxidation degree layer, and the third-level index is the radial stiffness layer.

6. The adaptive connection method for multi-specification leads of surge arresters as described in claim 5, characterized in that, The step of matching the corresponding docking process procedure from a preset process database based on the specification type, the degree of end oxidation, and the radial stiffness data includes: Based on the first-level index, using the specification type as the primary key, all docking process procedures applicable to the specification type are retrieved from the process database to obtain a first process procedure candidate set. Based on the degree of end oxidation, a target pretreatment strategy is determined from the pretreatment strategy set in the process database, and the degree of oxidation matching between the lead to be docked and each docking process in the first process procedure candidate set is calculated. Based on the oxidation degree matching degree, a docking process procedure with an oxidation degree matching degree greater than a preset oxidation degree matching threshold is selected from the first process procedure candidate set to obtain a second process procedure candidate set. Based on the radial stiffness data, the deviation between the radial stiffness data and the radial stiffness applicable range of each docking process in the second process specification candidate set is calculated to obtain the radial stiffness deviation. Based on the radial stiffness deviation, the pressing process parameters of each docking process in the second process specification candidate set are linearly interpolated and corrected to obtain the corrected pressing process parameters of each docking process in the second process specification candidate set. Based on the oxidation degree matching degree, the corrected pressing process parameters are weighted and fused to obtain the target pressing process parameters. Based on the oxidation degree matching degree and the radial stiffness deviation, a comprehensive similarity index is obtained, and based on the comprehensive similarity index, a target quality monitoring index is determined from the second process specification candidate set. The target quality monitoring index includes: target dynamic impedance curve, mechanical pull-out force threshold, and contact resistance upper limit.

7. The adaptive connection method for multi-specification leads of surge arresters as described in claim 6, characterized in that, The process of performing cleaning based on the target pretreatment strategy and pressing operation based on the target pressing process parameters according to the docking process specification to obtain the docking result includes: According to the target pretreatment strategy of the docking process procedure, determine the cleaning tools and perform the cleaning process; According to the target crimping process parameters in the aforementioned docking process specification, a visually guided positioning strategy is adopted to perform positioning processing. Based on the target crimping process parameters and the target dynamic impedance curve, a hybrid control strategy based on force feedback and electrical impedance dual feedback is adopted to perform the crimping operation. Based on the real-time dynamic impedance curve collected after crimping and the target dynamic impedance curve, the joint is automatically detected to obtain the mating result.

8. The adaptive connection method for multi-specification leads of surge arresters as described in claim 7, characterized in that, The step of performing the crimping operation based on the target crimping process parameters and the target dynamic impedance curve, using a hybrid control strategy based on force feedback and impedance dual feedback, includes: Based on the target pressing force and holding time in the aforementioned docking process specification, the desired force curve is obtained; Based on the deviation between the real-time pressing force and the expected pressing force from the expected force curve, the desired position of the robotic arm performing the pressing operation is corrected. Based on the impedance deviation between the target dynamic impedance curve and the acquired real-time impedance, the crimping force of the crimping operation performing robotic arm is corrected.

9. The adaptive connection method for multi-specification leads of surge arresters as described in claim 1, characterized in that, The method further includes: After the crimping is completed, a corresponding digital record of the crimping is obtained based on the specification type, the degree of end oxidation, the radial stiffness data, and the collected real-time docking parameters and quality inspection results. The digital record of the crimping process is used as feedback input to iteratively optimize the process database, so that the process parameters of the process database are dynamically updated.

10. A surge arrester multi-specification lead adaptive connection system, used to implement the surge arrester multi-specification lead adaptive connection method according to any one of claims 1-9, characterized in that, The system includes: a data acquisition unit, a multimodal recognition unit, a process specification matching unit, and a docking execution unit; The data acquisition unit is used to acquire multimodal data of the lead wire to be connected, including lead wire image data, eddy current signal data, and radial stiffness data. The multimodal recognition unit is used to determine the specification type and end oxidation degree of the lead to be docked based on the multimodal data using a pre-constructed multimodal perception model. The multimodal perception model is set to fuse the lead image data and the eddy current signal data based on an attention mechanism, and to use the radial stiffness data as a guiding signal to weight and enhance the fusion result of the attention mechanism. The process matching unit is used to match the corresponding butt welding process from a preset process database according to the specification type, the degree of end oxidation and the radial stiffness data. The butt welding process includes at least: target crimping process parameters, target dynamic impedance curve and target pretreatment strategy for different degrees of oxidation. The docking execution unit is used to perform cleaning processing based on the target pretreatment strategy and crimping operation based on the target crimping process parameters according to the docking process specification, so as to obtain the docking result. The crimping operation adopts a hybrid control strategy based on force sensing and electrical impedance dual feedback, and controls the crimping process based on the collected real-time dynamic impedance curve and the target dynamic impedance curve.

Images