Edge computing-based power transmission line steel structure welding quality evaluation method

Abstract

The application discloses a power transmission and transformation line steel structure welding quality evaluation method based on edge calculation, which comprises the following steps: deploying a federal edge multi-modal consensus network on an edge node, directly receiving process parameters such as welding current and voltage, and thermal radiation and molten pool shape detection data, and completing feature alignment between nodes relying on a consensus mechanism, thereby greatly reducing data transmission delay, deeply mining the internal correlation of multi-source data, and building a quality feature representation covering the whole welding process. The edge thermospectral lightweight perception model and the dynamic eddy current visual cooperative detection algorithm form a linkage processing, the thermospectral features and the molten pool boundary and flow state features are mutually verified, the linkage changes of the thermal field and the molten pool in the welding process are accurately captured, and the early identification efficiency of subtle defects is significantly enhanced. The edge side integrated architecture forms an efficient closed loop of data acquisition, fusion processing and quality evaluation, the method is suitable for the complex environment of the welding site, and promotes the evaluation mode to upgrade from traditional local sampling inspection to whole-process accurate monitoring.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of steel structure welding quality evaluation, and particularly relates to a power transmission line steel structure welding quality evaluation method based on edge computing. BACKGROUND

[0002] The power transmission line steel structure is the core support structure of the power transmission network, and the welding quality thereof is directly related to the stability and safety of the line operation. The welding process of such a steel structure involves multi-dimensional dynamic parameters such as arc characteristics, molten pool shape, welding current, voltage, and walking speed, and the welding environment is often accompanied by strong electromagnetic interference and complex thermal field changes, which puts strict requirements on the real-time and comprehensiveness of quality evaluation. The traditional evaluation method relies on manual sampling inspection or single sensor detection, which is difficult to synchronously capture the changes of multi-dimensional parameters, and cannot realize real-time monitoring and comprehensive evaluation of the welding process. With the development of edge computing technology, it is possible to deploy multi-modal data processing and intelligent algorithms on the edge nodes of the welding site, and to improve the timeliness of evaluation through the cooperation of federated networks and perception models, but the existing technology has not formed a mature edge-side multi-source data fusion evaluation system, which cannot fully adapt to the high-precision requirements of power transmission steel structure welding.

[0003] The existing technology has two significant shortcomings: first, the data processing and feature fusion lack efficient cooperation mechanisms, and rely mainly on centralized platforms for the aggregation and processing of welding parameters and detection data, without using federated edge multi-modal consensus networks to realize feature alignment between edge nodes, resulting in high data transmission delay and insufficient correlation mining of process parameters such as welding current and voltage, and detection data such as thermal radiation and molten pool shape, making it difficult to form a comprehensive quality feature representation. Second, the cooperation of perception detection models and algorithms is insufficient, and edge thermal spectrum perception and dynamic eddy current visual detection are independently operated, without establishing a deep correlation processing flow of thermal spectrum features and molten pool features, and unable to accurately capture the linkage features of thermal field changes and molten pool flow in the welding process through collaborative algorithms, resulting in weak early identification ability of welding defects and difficulty in meeting the fine evaluation requirements of power transmission steel structure welding quality. SUMMARY

[0004] In order to overcome the shortcomings and deficiencies of the existing technology, the present application provides a power transmission line steel structure welding quality evaluation method based on edge computing.

[0005] The technical scheme adopted by the present application is a power transmission and transformation line steel structure welding quality evaluation method based on edge computing, comprising the following steps: S1: deploying a federal edge multi-modal consensus network on an ESTUN Digital Robot Welding workstation edge node, collecting arc characteristics, molten pool shape, welding current, welding voltage, and walking speed data during the welding process of the power transmission and transformation line steel structure through a built-in sensing module of the workstation, and transmitting the data to the federal edge multi-modal consensus network node; S2: starting the edge thermal spectrum lightweight perception model to perceive and capture the welding area thermal radiation signal received by the edge node, generating a thermal spectrum feature vector in combination with the arc characteristic data, and transmitting the thermal spectrum feature vector to the federal edge multi-modal consensus network; S3: calling a dynamic eddy current visual collaborative detection algorithm to cooperatively process the molten pool shape data, extracting molten pool boundary contour, size change, and flow state characteristic parameters, and synchronizing the parameters to the federal edge multi-modal consensus network; S4: the federal edge multi-modal consensus network performs multi-modal data fusion on the received thermal spectrum feature vector, molten pool characteristic parameters, and welding current, welding voltage, and walking speed data, and completes data feature alignment between nodes through a consensus mechanism; S5: based on the aligned fusion feature data, combining the output results of the edge thermal spectrum lightweight perception model and the dynamic eddy current visual collaborative detection algorithm, a welding quality feature matrix is constructed; S6: the ESTUN Digital Robot Welding workstation edge processing unit analyzes the welding quality feature matrix and outputs the power transmission and transformation line steel structure welding quality evaluation result.

[0006] Further, the feature fusion expression of the federal edge multi-modal consensus network is: wherein represents a multi-modal fusion feature, represents the number of federal edge nodes, represents the weight factor of the th node, represents a parameter matrix composed of welding current, voltage, and walking speed collected by the th node, represents a thermal spectrum feature vector output by the edge thermal spectrum lightweight perception model of the th node, represents molten pool characteristic parameters output by the dynamic eddy current visual collaborative detection algorithm of the th node, represents a feature splicing operation, represents an alignment matrix generated by a consensus mechanism, represents a matrix point multiplication operation.

[0007] Further, the thermal spectrum feature extraction expression of the edge thermal spectrum lightweight perception model is: wherein represents a thermal spectrum feature vector, Represents the activation function. Represents a two-dimensional convolution operation. Represents the thermal sensing convolution kernel. Represents convolution operation. Represents the thermal radiation signal of the welding area. This represents the thermal spectrum sensing bias term. Represents element-wise multiplication, and Pool represents pooling operations. Represents the characteristic mapping function of electric arc. This represents the characteristic data of the electric arc.

[0008] Furthermore, the expression for the molten pool feature extraction of the dynamic eddy current visual collaborative detection algorithm is as follows: ,in Represents the characteristic parameters of the molten pool. Represents gradient operation, Edge represents the edge detection operator, and E represents the vortex field feature extraction function. Data representing the morphology of the molten pool. The symbol represents the feature intersection operation, and Flow represents the flow state analysis function. Represents the visual feature transformation function. Represents the visual collaborative weights of eddy currents.

[0009] Furthermore, the expression for constructing the welding quality feature matrix is ​​as follows: ,in Represents the welding quality characteristic matrix. Represents the thermal spectral eigenvector. Represents the thermal spectrum feature weight matrix. Represents the characteristic parameters of the molten pool. Represents the feature weight matrix of the molten pool. A matrix representing welding current, voltage, and travel speed parameters. Represents the weight matrix of welding process parameters. Represents the federal consensus alignment transformation matrix. This represents the matrix transpose operation.

[0010] Furthermore, the evaluation result output expression of the ESTUN Digital Robot Welding workstation is as follows: ,in The sigmoid function represents the sigmoid activation function and represents the welding quality assessment result. Represents the welding quality characteristic matrix. Represents the evaluation weight matrix. Represents the evaluation bias term. Represents the set of standard characteristics for welding quality. This represents feature matching operations.

[0011] Further, step S2 includes the following sub-steps: S21: The edge thermal spectrum lightweight perception model receives the real-time thermal radiation signal of the welding area through the infrared sensing interface of the ESTUN Digital Robot Welding workstation, fragments and intercepts the signal at a preset time interval, and obtains a continuous thermal radiation signal sequence; S22: The built-in thermal signal enhancement module of the model is called to filter the intercepted signal sequence, retain the thermal radiation characteristic components related to the welding arc and the molten pool area, form a preliminary processed thermal signal, and S23: The feature correlation module is combined with the arc feature data transmitted in step S1 to establish a mapping relationship between the thermal radiation signal and the arc intensity and frequency features, and generate a correlation feature group; S24: The correlation feature group is dimensionally compressed by the feature dimension reduction module of the model to extract the core thermal spectrum features and construct a thermal spectrum feature vector, which is sent to the federated edge multi-modal consensus network.

[0012] Further, step S3 includes the following sub-steps: S31: The dynamic eddy current visual collaborative detection algorithm receives the molten pool morphology image data transmitted by the vision acquisition module of the ESTUN Digital Robot Welding workstation, divides the image into three regions of the molten pool core area, the transition area and the background area through the image segmentation module; S32: The eddy current detection sub-module is started to perform eddy current field simulation operation on the molten pool core area image, and extract the eddy current feature points and distribution law generated by the flow of substances in the molten pool; S33: The boundary of the molten pool transition area is identified by the visual detection sub-module, and the contour tracking algorithm is used to obtain the molten pool boundary contour coordinates and contour smoothness data; S34: The eddy current feature point distribution law is fused with the boundary contour coordinates and contour smoothness data, and the molten pool size change rate and flow uniformity parameters are calculated to form the molten pool feature parameters and synchronously to the federated edge multi-modal consensus network.

[0013] Further, step S4 includes the following sub-steps: S41: The nodes of the federated edge multi-modal consensus network receive the thermal spectrum feature vector output by the local edge thermal spectrum lightweight perception model, the molten pool feature parameters output by the dynamic eddy current visual collaborative detection algorithm, and the welding process parameters collected by the workstation, and perform local feature standardization processing; S42: Each node sends the standardized local feature data to the consensus coordination node through an encrypted communication link, and the consensus coordination node performs consistency check on the received multi-node feature data; S43: The federated average consensus strategy is adopted to fuse the multi-node feature data that passes the check, adjust the feature space of each node to a unified dimension, and complete the feature space alignment; S44: The aligned fusion feature data is fed back to each edge node to perform global consistency of the feature data in the federated edge multi-modal consensus network.

[0014] Further, step S5 includes the following sub-steps: S51: calling a feature matrix initialization module, constructing an empty welding quality feature matrix framework according to the power transmission line steel structure welding quality evaluation dimensions, and the framework dimensions are determined by the number of thermal spectrum features, molten pool features and process parameter features; S52: extracting the aligned thermal spectrum feature vector from the federated edge multi-modal consensus network, converting the thermal spectrum feature vector into a thermal spectrum feature column vector corresponding to the matrix framework through a feature mapping function, and filling the matrix at the specified position; S53: extracting the aligned molten pool feature parameters and welding process parameters, respectively converting them into corresponding column vectors in the same feature mapping manner, and sequentially filling them into the corresponding columns of the matrix framework; S54: scaling the elements of the filled matrix through a matrix normalization module to ensure that each feature column vector is in the same numerical range, and forming a final welding quality feature matrix.

[0015] Beneficial effects: The power transmission line steel structure welding quality evaluation method based on edge computing is proposed, the federated edge multi-modal consensus network is deployed on the edge node to replace the traditional centralized data processing mode, the welding current, voltage and other process parameters and the thermal radiation, molten pool morphology and other detection data are directly collected and aligned on the edge side, the data transmission delay is greatly reduced, and the multi-source data correlation is deeply mined through the consensus mechanism to form a comprehensive quality feature representation, effectively solving the problem of lack of efficient cooperation in data processing and feature fusion. At the same time, relying on the linkage operation of the edge thermal spectrum lightweight perception model and the dynamic eddy current visual cooperative detection algorithm, a deep correlation processing flow of thermal spectrum features and molten pool features is established, the linkage features of thermal field changes and molten pool flow in the welding process are accurately captured, the early identification ability of welding defects is significantly improved, the fine evaluation demand is met, and the real-time evaluation of the edge side algorithm deployment and data processing further guarantees the real-time evaluation of the severe requirements of the power transmission steel structure welding. BRIEF DESCRIPTION OF DRAWINGS

[0016] Figure 1 The method flowchart of the present application;

[0017] Figure 2 The method implementation unit composition diagram of the present application. DETAILED DESCRIPTION

[0018] It should be noted that the embodiments in the present application and the features in the embodiments can be combined with each other without conflict, and the present application will be further described in detail below in combination with the drawings and specific embodiments.

[0019] As shown in the drawings, the power transmission line steel structure welding quality evaluation method based on edge computing includes: Figure 1

[0020] ​Step S1: Deploy the federated edge multi-modal consensus network on the ESTUN Digital Robot Welding workstation edge node, collect the arc characteristics, molten pool shape, welding current, welding voltage, and walking speed data during the welding process of the power transmission line steel structure through the built-in sensing module of the workstation, and transmit them to the federated edge multi-modal consensus network node;

[0021] Specifically, step S1 completes the deployment of the federated edge multi-modal consensus network on the edge node. The edge node selects an embedded processing module with adaptive computing power, which has an operation capacity of processing no less than 100 MB of welding data per second and has multi-interface adaptation capability to connect sensing modules. During deployment, the identity authentication and communication configuration of the network node are completed, the encryption protocol for data transmission between nodes and the consensus trigger threshold are set, and the consensus trigger threshold is set to start the consensus process when the node receives 50 MB of data. Data collection is carried out through the built-in sensing module of the welding equipment. Arc characteristics are collected by an arc sensor, with a collection frequency of 200 Hz to capture the intensity fluctuation range and duration of the arc. The molten pool shape is collected by a high-speed industrial camera with a frame rate of 120 fps and a resolution of 1920x1080 to ensure clear capture of real-time changes in the molten pool shape. The welding current is collected by a current sensor with a range of 0-500 A and an accuracy of ±1 A. The welding voltage is collected by a voltage sensor with a range of 0-50 V and an accuracy of ±0.1 V. The walking speed is collected by a displacement sensor with a collection frequency of 50 Hz, a measurement range of 0-10 mm / s, and an accuracy of ±0.01 mm / s. The collected data is synchronized and encapsulated according to the timestamp, and the size of the encapsulated data block is controlled at 10 MB / block. The data is transmitted in real time to each node of the federated edge multi-modal consensus network through the Ethernet interface at a transmission rate of 100 Mbps.

[0022] Step S2: Start the edge thermal spectrum lightweight perception model to capture the welding area thermal radiation signal received by the edge node, generate a thermal spectrum feature vector combined with the arc characteristic data, and transmit it to the federated edge multi-modal consensus network;

[0023] Specifically, step S2 starts the edge thermal spectrum lightweight perception model immediately after the edge node receives the welding data. The model uses a lightweight convolutional neural network architecture, with a total of less than 5 million model parameters to adapt to the computing power limit of the edge node. The model receives the thermal radiation signals of the welding area through the infrared thermal image acquisition interface. The wavelength range is set to 8-14 μm, corresponding to the standard detection band of industrial infrared thermal imagers. The acquisition frequency is synchronized with the molten pool shape acquisition frequency, i.e. 120 fps, ensuring the time consistency of the thermal radiation signals and the molten pool shape data. The received thermal radiation signals are first divided into frames according to a time interval of 10 ms, with each frame corresponding to the thermal field distribution data at a time point. The signal intensity of each frame of thermal radiation signal is extracted, and the extraction range covers a circular area with a diameter of 200 mm in the welding area, which includes the arc core area, the molten pool area and the heat affected zone. The extracted signal intensity data is associated and matched with the arc characteristic data collected in step S1. The matching is based on the time stamp, and the arc intensity value, frequency value and thermal radiation signal intensity data at the same time stamp are bound to generate a thermal spectrum feature vector including 200 feature dimensions, which correspond to the intensity distribution of the thermal radiation signal at different spatial positions and the arc characteristic parameters. The generated thermal spectrum feature vector is data encoded in a predetermined format, and the data transmission rate after encoding is controlled at 50 Mbps. The encoded data is transmitted to the feature receiving module of the federated edge multi-modal consensus network through the internal data bus, and a data check bit is set during the transmission process. The check bit is generated using the CRC32 algorithm to ensure the integrity of data transmission.

[0024] Step S3: calling a dynamic eddy current visual collaborative detection algorithm to collaboratively process the molten pool shape data, extracting the molten pool boundary contour, size change and flow state characteristic parameters, and synchronizing to the federated edge multi-modal consensus network;

[0025] Specifically, when the dynamic eddy current visual collaborative detection algorithm is called in step S3, the initialization and configuration of the algorithm parameters are completed. In the eddy current detection parameters, the excitation current is set to 5A, the excitation frequency is adjusted to 10 kHz, and the distance between the eddy current probe and the welding area is fixed at 50 mm to ensure that the eddy current field can effectively cover the molten pool and the surrounding area. In the visual detection parameters, the gray threshold value of image preprocessing is set to 128, and the gradient threshold value of edge detection is set to 30 to accurately distinguish the molten pool from the background area. After the algorithm receives the molten pool shape image data transmitted in step S1, the image is first subjected to grayscale processing, converting the color image into an 8-bit grayscale image, and then subjected to image noise reduction processing. The median filter algorithm is used, and the filter window size is set to 3x3 to remove random noise in the image. The eddy current detection submodule is started, and an excitation signal is sent to the eddy current probe while receiving the eddy current response signal returned by the probe. The collection frequency of the response signal is 200 Hz, the response signal is converted into eddy current field distribution data, and the eddy current intensity variation, phase shift and other characteristics are extracted. These characteristics can reflect the change in the electrical conductivity of the material inside the molten pool. The visual detection submodule is started synchronously, and the edge of the denoised molten pool image is detected. The Canny operator is used to extract the molten pool boundary contour, and the coordinates of not less than 100 feature points on the boundary are obtained through the contour tracking algorithm. The length, width and other size parameters of the molten pool are calculated, and the calculation accuracy of the size parameters is controlled within ±0.1 mm. At the same time, through the matching of feature points in adjacent image frames, the change rate of the molten pool size is calculated, and the calculation interval of the change rate is 100 ms, and the change amount of the length and width of the molten pool within every 100 ms is obtained. In addition, through the image optical flow analysis algorithm, the flow direction and speed of the material inside the molten pool are obtained, the measurement range of the flow speed is 0-5 mm / s, and the accuracy is ±0.05 mm / s. Finally, the eddy current characteristics and the molten pool boundary contour, size change, flow state and other parameters extracted by vision are fused, the weights are assigned according to the importance of the characteristics, the weight proportion of the eddy current characteristics is 40%, and the weight proportion of the visual characteristics is 60%. After fusion, 150-dimensional molten pool feature parameters are formed, which are synchronized to the feature fusion module of the federal edge multi-modal consensus network through a data interface.

[0026] Step S4: The federal edge multi-modal consensus network performs multi-modal data fusion on the received thermal spectrum feature vector, molten pool feature parameter, and welding current, welding voltage, and walking speed data, and completes the alignment of data features between nodes through a consensus mechanism.

[0027] Specifically, the step S4 federal edge multi-modal consensus network receives the thermal spectrum feature vector, the molten pool feature parameter, and the welding current, welding voltage, and walking speed data, and performs format standardization processing on each type of data. The 200 dimensions of the thermal spectrum feature vector, the 150 dimensions of the molten pool feature parameter, and the 3 dimensions (current, voltage, and walking speed) of the welding process parameter are uniformly converted to 32-bit floating point number format to ensure data type consistency. Each edge node starts local feature preprocessing, scales the range of the standardized data, and maps all feature values to the 0-1 interval. The scaling is based on the historical maximum and minimum values of each feature, where the historical maximum value of the welding current is set to 500 A, the minimum value is set to 100 A, the historical maximum value of the welding voltage is set to 50 V, the minimum value is set to 10 V, the historical maximum value of the walking speed is set to 10 mm / s, and the minimum value is set to 1 mm / s. The thermal spectrum and molten pool features are mapped according to the extreme values calibrated earlier. After preprocessing, each node sends local feature data to the consensus coordination node through an encrypted communication link. The communication link uses the SSL / TLS encryption protocol, and the data transmission delay is controlled within 50 ms. After receiving the data from each node, the consensus coordination node performs consistency checking. The checking standard is that the timestamp deviation of each node data does not exceed 10 ms, and the proportion of abnormal values of the feature values does not exceed 5%. The abnormal value is determined according to the 3σ criterion, i.e., data exceeding the mean value ± 3 times the standard deviation is determined as an abnormal value. After passing the verification, the consensus mechanism is started, and the federal average algorithm is used to weight and fuse the multi-node feature data. The weight of each node is determined according to its data quality score, which considers data integrity, collection accuracy, and other indicators. Nodes with integrity of 98% or more and accuracy meeting the preset standard are set to a weight of 0.8, and the remaining nodes meeting the basic requirements are set to a weight of 0.2. In the fusion process, the feature spaces of each node are aligned, and the feature vectors of different nodes are converted to a unified feature space through a feature mapping matrix. The dimension of the mapping matrix is 353x353 (353 is the total feature dimension), ensuring that the fused feature data remains consistent in spatial dimension. After fusion, the aligned fused feature data is fed back to each edge node. The update frequency of the feedback data is consistent with the data collection frequency, realizing the global consistency of the feature data in the network.

[0028] Step S5: Based on the aligned fused feature data, the output results of the edge thermal spectrum lightweight perception model and the dynamic eddy current visual collaborative detection algorithm are combined to construct a welding quality feature matrix.

[0029] Specifically, step S5 calls the welding quality feature matrix construction module, which preloads the evaluation dimension configuration of the welding quality of the power transmission and transformation steel structure. The evaluation dimensions include three categories: weld forming quality, internal defects, and mechanical property correlation features, a total of 30 evaluation sub-dimensions. Therefore, the dimension of the welding quality feature matrix is set to 353x30, with the rows corresponding to the total feature dimensions and the columns corresponding to the evaluation sub-dimensions. The aligned fusion feature data is extracted from the feature storage unit of the federal edge multi-modal consensus network. The fusion feature data includes 200-dimensional thermal spectrum feature vectors, 150-dimensional molten pool feature parameters, and 3-dimensional welding process parameters. When extracting, verify whether the timestamp of the validation data matches the evaluation period, and control the matching deviation within 20ms. Next, the thermal spectrum feature vectors are processed. The feature selection algorithm is used to select thermal spectrum features with a correlation higher than 0.7 with each evaluation sub-dimension. The Pearson correlation coefficient method is used for correlation calculation. The selected thermal spectrum features are sequentially filled into the corresponding columns of the matrix according to the evaluation sub-dimensions. For example, thermal spectrum features related to weld forming quality are filled into columns 1-10, internal defect-related features are filled into columns 11-20, and mechanical property correlation features are filled into columns 21-30. The molten pool feature parameters are processed synchronously. Features with a correlation higher than 0.7 are selected and filled into the corresponding columns. The welding process parameters are directly filled into the matrix according to their association with the evaluation sub-dimensions. The welding current corresponds to columns 1, 11, and 21, the welding voltage corresponds to columns 2, 12, and 22, and the walking speed corresponds to columns 3, 13, and 23. After the feature filling is completed, the matrix is normalized. The min-max normalization method is used to map the numerical value of each element in the matrix to the 0-1 interval. The normalization parameters (maximum value, minimum value) of each column are recorded during the mapping process for subsequent reverse analysis of the evaluation results. Finally, a welding quality feature matrix with a dimension of 353x30 is formed. The matrix data is stored in the local storage module of the edge node in binary format, with a storage capacity of not less than 10GB to meet the continuous evaluation requirements.

[0030] Step S6: The ESTUN Digital Robot Welding workstation edge processing unit analyzes the welding quality feature matrix and outputs the welding quality evaluation results of the power transmission and transformation steel structure.

[0031] Specifically, step S6 reads the welding quality feature matrix from the local storage module by the edge processing unit through the internal bus, and the reading rate is set to 200 MB / s, ensuring that the complete reading of the matrix data is completed within 1 s. After reading is completed, the edge processing unit starts the matrix analysis program, which adopts a feature weighted summation algorithm, pre-configures a weight coefficient for each row of features of the feature matrix according to the corresponding relationship between historical welding quality data and evaluation results, the weight coefficient has a value range of 0-1, and the sum of all row weight coefficients is 1, wherein the weight coefficients of the molten pool boundary contour feature, the thermal spectrum core region feature and the welding current feature are relatively high, and are respectively set to 0.08, 0.07 and 0.06. In the analysis process, the score of each of the 30 evaluation sub-dimensions is obtained by multiplying each row of feature values of the matrix by the corresponding weight coefficient and then performing column-wise summation, and the score range is 0-1, and the higher the score, the better the quality performance corresponding to the sub-dimension. The scores of the 30 evaluation sub-dimensions are classified and summarized, the scores of the weld forming quality class (columns 1-10) are averaged to obtain the total score of the class, the scores of the internal defect class (columns 11-20) are averaged to obtain the total score of the class, and the scores of the mechanical performance related feature class (columns 21-30) are averaged to obtain the total score of the class, and the value ranges of the total scores of the three classes are all 0-1. Finally, the evaluation result is output according to the pre-set quality grade division standard, wherein all three total scores are greater than 0.8 to determine that the quality is excellent, two total scores are greater than 0.8 and one total score is greater than 0.6 to determine that the quality is qualified, and the rest is determined to be unqualified, and the evaluation result is output in a structured data format, including the three total scores and the final quality grade, the output data is transmitted to the welding quality monitoring terminal through the communication interface, the transmission delay is controlled within 100 ms, and the entire welding quality evaluation process is completed.

[0032] Preferably, the feature fusion expression of the federated edge multi-modal consensus network is: wherein represents a multi-modal fusion feature, represents the number of federated edge nodes, represents the weight factor of the th node, represents the parameter matrix composed of the welding current, voltage and walking speed collected by the th node, represents the thermal spectrum feature vector output by the thermal spectrum lightweight perception model of the th node edge, represents the molten pool feature parameter output by the dynamic eddy current visual collaborative detection algorithm of the th node, represents a feature splicing operation, represents an alignment matrix generated by a consensus mechanism, represents a matrix point multiplication operation.

[0033] Specifically, in the operation of the federal edge multi-modal consensus network, the number of edge nodes participating in data fusion is first determined, combined with the sensing coverage range of the steel structure welding site of the power transmission line, the number of nodes is usually set to 3-5, each node corresponds to a different welding area monitoring point, ensuring the comprehensiveness of data collection. The weight factor of each node is pre-configured according to its data collection quality, the node weight is set to 0.8 when the data integrity reaches more than 98% and the collection accuracy meets the preset standard, and the weight of the remaining nodes that meet the basic requirements is set to 0.2, and the sum of the weight factors is 1, to ensure the rationality of the fusion result. The welding parameter matrix is composed of welding current, voltage and walking speed data collected by each node, each parameter is arranged in time sequence to form a 100-row 3-column matrix, the row corresponds to the collection time point, and the column corresponds to different parameters. The thermal spectrum feature vector is the 200-dimensional feature data output by the edge thermal spectrum lightweight perception model, and the molten pool feature parameter is the 150-dimensional feature data output by the dynamic eddy current visual collaborative detection algorithm. The three are combined into a 353-dimensional joint feature vector through feature splicing operation. The alignment matrix generated by the consensus mechanism has a dimension of 353x353, and its element value is calculated according to the similarity of the features between nodes. The position element with a similarity higher than 0.9 is set to 1, and the rest is set to 0.1, which is used to correct the deviation of the feature space of different nodes. Finally, the fusion feature is obtained through weighted summation and matrix point multiplication operation, which is completed locally in the edge node, and the operation time is controlled within 50ms to ensure the real-time requirement, and the fusion feature can be directly used for the construction of the subsequent quality feature matrix.

[0034] Preferably, the thermal spectrum feature extraction expression of the edge thermal spectrum lightweight perception model is: wherein represents a thermal spectrum feature vector, represents an activation function, represents a two-dimensional convolution operation, represents a thermal spectrum perception convolution kernel, represents a convolution operation, represents a welding area thermal radiation signal, represents a thermal spectrum perception bias term, represents an element-wise product, and Pool represents a pooling operation. represents an arc feature mapping function, represents arc feature data.

[0035] Specifically, after the edge thermal spectrum lightweight perception model is started, preset thermal spectrum perception convolution kernels are loaded first, the convolution kernel size is set to 3x3, and a total of 64 different convolution kernels are included, which correspond to different thermal radiation signal feature modes, such as point heat source features and surface heat source features. After the welding area thermal radiation signal is collected by the infrared thermal imager, it is converted into a 256x256 grayscale image data, and the grayscale value of each pixel point represents the thermal radiation intensity of the corresponding position. In the convolution operation process, each convolution kernel performs sliding window calculation with the thermal radiation signal image, the window sliding step is set to 1, and 64 feature maps are generated, each with a size of 254x254. A thermal spectrum perception bias term is added, the bias term is 64 constants, the value range is 0.01-0.05, and is used to adjust the overall numerical level of the feature map. The activation function adopts the ReLU function, which performs nonlinear transformation on the convolution operation result, sets the negative value to 0, retains the positive value feature, and enhances the extraction ability of the model to key thermal spectrum features. The arc feature mapping function converts the collected arc intensity and frequency data into a 256x256 feature map, and performs element-by-element multiplication operation with the convolution result to realize the correlation enhancement of the thermal spectrum features and the arc features. Finally, the feature map is reduced in dimension through the maximum pooling operation, the pooling window size is 2x2, the step is 2, the feature map size is compressed to 127x127, and the final thermal spectrum feature vector is formed. The total amount of model parameters in this process is controlled within 5 million, which is suitable for the power limit of the edge node, and the feature extraction accuracy is above 92%.

[0036] Preferably, the molten pool feature extraction expression of the dynamic eddy current visual collaborative detection algorithm is: wherein represents a molten pool feature parameter, represents a gradient operation, Edge represents an edge detection operator, and E represents an eddy current field feature extraction function, represents molten pool morphology data, represents a feature intersection operation, and Flow represents a flow state analysis function, represents a visual feature conversion function, represents an eddy current visual collaborative weight.

[0037] Specifically, the dynamic eddy current visual collaborative detection algorithm is initialized, the parameters of the eddy current field feature extraction function are configured, the excitation current is set to 5A, the excitation frequency is adjusted to 10kHz, and according to the material characteristics of the welding area, the eddy current penetration depth parameter is set to 2mm, which can ensure that the feature change in the molten pool can be detected. The molten pool shape data is a 1920x1080 resolution image collected by a high-speed industrial camera, which is input into the eddy current field feature extraction function after grayscale and noise reduction processing, and the corresponding eddy current field distribution image is generated. The image size is consistent with the original image, and the pixel value represents the eddy current intensity. The edge detection operator uses the Canny operator, and the gradient threshold is set to 30. The edge detection of the eddy current field distribution image is performed, the boundary features of the eddy current field are extracted, and the binary edge image is obtained. The visual feature conversion function converts the molten pool shape image into a flow state feature map. The pixel displacement of adjacent frame images is calculated by the optical flow analysis algorithm, and a 1920x1080 optical flow vector map is generated. The direction of the vector represents the flow direction of the molten pool, and the size represents the flow speed. The feature intersection operation is to take the common effective area of the edge detection result and the flow state feature map. This area includes both the eddy current boundary and the molten pool flow features. The effective area accounts for more than 60% of the original image, otherwise the detection parameters need to be adjusted. The gradient operation processes the intersection area to calculate the spatial change rate of the features and generates a gradient feature map. The eddy current visual collaborative weight is a 150-dimensional vector with a value range of 0.01-0.1. The eddy current intensity change feature weight is set to 0.1, and the molten pool flow speed feature weight is set to 0.08. The molten pool feature parameter is obtained by weighted operation, which can comprehensively reflect the physical state change of the molten pool and provide key basis for welding quality evaluation.

[0038] Preferably, the welding quality feature matrix construction expression is: wherein represents the welding quality feature matrix, represents the thermal spectrum feature vector, represents the thermal spectrum feature weight matrix, represents the molten pool feature parameter, represents the molten pool feature weight matrix, represents the welding current, voltage, and walking speed parameter matrix, represents the welding process parameter weight matrix, represents the federal consensus alignment conversion matrix, represents the matrix transposition operation.

[0039] Specifically, before the welding quality feature matrix is constructed, the weight matrix of each feature is configured first. The weight matrix of the thermal spectrum feature has a dimension of 200x30, and each element represents the associated weight of the corresponding thermal spectrum feature and the evaluation sub-dimension. The weight value related to the weld forming quality ranges from 0.05 to 0.1, the weight value related to the internal defect ranges from 0.06 to 0.12, and the weight value related to the mechanical property associated feature ranges from 0.04 to 0.09. The weight matrix of the molten pool feature has a dimension of 150x30. The corresponding weight of the molten pool boundary profile feature is set to 0.08-0.1, the corresponding weight of the size change feature is set to 0.07-0.09, and the corresponding weight of the flow state feature is set to 0.06-0.08. The weight matrix of the welding process parameter has a dimension of 3x30. The corresponding weight of the welding current is set to 0.09-0.11, the corresponding weight of the welding voltage is set to 0.08-0.1, and the corresponding weight of the walking speed is set to 0.07-0.09. The thermal spectrum feature vector, the molten pool feature parameter, and the welding process parameter are subjected to matrix multiplication operation with the corresponding weight matrix to obtain feature contribution matrices with dimensions of 200x30, 150x30, and 3x30, respectively. The three feature contribution matrices are transposed and longitudinally spliced to form an initial matrix with a dimension of 353x30. The federal consensus alignment conversion matrix has a dimension of 353x353. According to the feature alignment result between nodes, the position elements with an alignment error of less than 5% are set to 1, and the rest are set to 0.05, which are used to correct the feature deviation in the initial matrix. The matrix multiplication operation is completed in the edge processing unit, and the operation uses floating-point precision calculation with an error controlled within 0.001. The generated welding quality feature matrix is stored in the local storage module in binary format to ensure fast data reading and subsequent analysis.

[0040] Preferably, the evaluation result output expression of the ESTUN Digital Robot Welding workstation is: wherein represents the welding quality evaluation result, Sigmoid represents the Sigmoid activation function, represents the welding quality feature matrix, represents the evaluation weight matrix, represents the evaluation bias term, represents the welding quality standard feature set, represents the feature matching operation.

[0041] Specifically, after the edge processing unit of the ESTUN Digital Robot Welding workstation reads the welding quality feature matrix, a preset evaluation weight matrix is loaded, which has a dimension of 30x1, and each element represents the weight of the corresponding evaluation sub-dimension in the final evaluation. The weight of the weld forming quality class sub-dimension is in the range of 0.03-0.04, the weight of the internal defect class sub-dimension is in the range of 0.035-0.045, and the weight of the mechanical property related feature class sub-dimension is in the range of 0.025-0.035. The sum of all elements is 1. The evaluation bias term is 30 constants with a value range of 0.01-0.03, which is used to adjust the baseline score of each evaluation sub-dimension. Matrix multiplication operation multiplies the welding quality feature matrix and the evaluation weight matrix to obtain a 353x1 feature score vector, and then adds the evaluation bias term to adjust the numerical level of the score vector. The Sigmoid activation function operates on the adjusted score vector to map the values to the 0-1 interval to obtain the normalized score of each feature. Features with a score higher than 0.8 are determined to have a positive contribution to the welding quality, and features with a score lower than 0.3 are determined to have a negative contribution. The welding quality standard feature set includes feature score data of historical high-quality welding samples, a total of 1000 sample data, each data is a 353x1 vector. The feature matching operation calculates the similarity between the current normalized score and the sample data in the standard feature set. If the number of samples with a similarity higher than 0.9 accounts for more than 70%, the corresponding evaluation result is high quality. If the number of samples with a similarity between 0.7 and 0.9 accounts for more than 70%, the result is qualified, otherwise it is unqualified. After the evaluation result is generated, it is packaged in JSON format, including three types of total score and final quality level, and transmitted to the monitoring terminal through the Ethernet interface at a transmission rate of 100 Mbps with a delay of no more than 100 ms, ensuring that the welding quality information can be obtained by the staff in a timely manner.

[0042] Preferably, step S2 comprises the following sub-steps: S21: the edge thermal spectrum lightweight perception model receives the real-time thermal radiation signal of the welding area through the infrared sensing interface of the ESTUN Digital Robot Welding workstation, fragments the signal at a preset time interval to obtain a continuous thermal radiation signal sequence; S22: a built-in thermal signal enhancement module of the model is called to filter the noise of the intercepted signal sequence, retain the thermal radiation characteristic components related to the welding arc and molten pool area, and form a preliminary processed thermal signal; S23: combined with the arc feature data transmitted in step S1, the mapping relationship between the thermal radiation signal and the arc intensity and frequency features is established through a feature correlation module to generate a correlation feature group; S24: the correlation feature group is dimensionally compressed through a feature dimension reduction module of the model to extract core thermal spectrum features and construct a thermal spectrum feature vector, which is sent to the federated edge multi-modal consensus network.

[0043] Specifically, after step S2 is started, first, S21 is executed, the edge thermal spectrum lightweight perception model is connected with the detection equipment through a special infrared sensing interface, the interface adopts an industrial standard interface, and a data transmission rate is set to 50 Mbps, to ensure real-time transmission of thermal radiation signals. The received continuous thermal radiation signals are fragmented and intercepted at a preset time interval of 10 ms, each signal fragment includes 12 frames of thermal image data, corresponding to a collection frequency of 120 fps, and a signal sequence formed after interception includes a complete thermal radiation change process from arc striking to a stable stage of the welding arc. Then, S22 is entered, a thermal signal enhancement module built in the model is called, the module adopts an adaptive filtering algorithm, a filtering window size is dynamically adjusted according to signal noise intensity, the range is 3*3 to 7*7, high-frequency noise generated by environmental interference is filtered out through multiple rounds of iteration, thermal radiation characteristic components in a 8-14 μm wave band that are directly related to the welding arc and the molten pool are retained, and a signal-to-noise ratio of the processed preliminary thermal signal is improved to more than 30 dB. Next, S23 is executed, a feature correlation module binds the preliminary thermal signal with the arc feature data transmitted in step S1 through a time stamp synchronization mechanism, a binding error is controlled to be within 5 ms, a mapping relationship between thermal radiation signal intensity and arc intensity fluctuation range and arc frequency is established, a feature group including 200 groups of correlation data is generated, and each group of data corresponds to thermal field and arc features at the same time point. Finally, S24 is performed, a feature dimension reduction module adopts a principal component analysis algorithm, a variance contribution rate threshold is set to 95%, the first 50 core features whose cumulative variance contribution rate reaches the threshold are screened out, the original 200-dimensional correlation feature group is compressed into a 50-dimensional thermal spectrum feature vector, key features such as thermal radiation peak value and temperature gradient are retained in the compression process, and the generated feature vector is transmitted to a feature receiving buffer of the federated edge multi-modal consensus network at a rate of 20 Mbps through an internal data bus.

[0044] Preferably, step S3 includes the following sub-steps: S31: a dynamic eddy current visual cooperative detection algorithm receives molten pool morphology image data transmitted by an ESTUN Digital Robot Welding workstation vision acquisition module, and divides the image into a molten pool core area, a transition area and a background area through an image segmentation module; S32: an eddy current detection submodule is started to perform eddy current field simulation operation on the molten pool core area image, to extract eddy current feature points and distribution rules generated by flow of substances in the molten pool; S33: a visual detection submodule is started to perform boundary recognition on the molten pool transition area, to obtain molten pool boundary contour coordinates and contour smoothness data by using a contour tracking algorithm; and S34: the eddy current feature point distribution rules, the boundary contour coordinates and the contour smoothness data are fused, to calculate a molten pool size change rate and a flow uniformity parameter, to form molten pool feature parameters and synchronize the molten pool feature parameters to the federated edge multi-modal consensus network.

[0045] Specifically, after step S3 is started, S31 is first executed, the dynamic eddy current visual collaborative detection algorithm receives the molten pool shape image data through the gigabit Ethernet interface, and the interface communication delay is controlled within 10 ms. The image segmentation module adopts an algorithm combining threshold segmentation and region growing. First, the gray threshold is set to 128, and the image is preliminarily divided into foreground and background. Then, the seed point of the molten pool core area is taken as the starting point, and region growing is performed according to the standard that the pixel gray difference is not more than 20. Finally, the image is accurately divided into the molten pool core area, the transition area and the background area, wherein the molten pool core area covers a circular area with a diameter of 100 mm at the center of the image, and the width of the transition area is set to 20 mm. Then, S32 is entered, the eddy current detection submodule sends a 5A, 10 kHz excitation signal to the eddy current probe, the probe and the welding area maintain a fixed distance of 50 mm, the received eddy current response signal is converted into 1920*1080 eddy current field distribution data after analog-digital conversion, not less than 50 eddy current feature points are identified through the eddy current field feature extraction algorithm, the intensity value and the coordinates of each feature point are recorded, and the distribution density and the clustering rule of the feature points are analyzed. Next, S33 is executed, the visual detection submodule adopts a Canny operator to perform edge detection on the transition area image, the gradient threshold is set to 30, the edge pixels are traversed in a clockwise direction through a contour tracking algorithm, more than 100 boundary contour feature point coordinates are obtained, the distance and the angle change of adjacent feature points are calculated, the contour smoothness parameter is obtained, the smoothness is represented by the contour fitting error, and the error value is controlled within 0.5 mm. Finally, S34 is performed, a weighted fusion algorithm is used to process the eddy current features and the visual features, the weight proportion of the eddy current features is 40%, the weight proportion of the visual features is 60%, the change rate of the molten pool length and width is calculated, the change rate is represented by the size change amount per 100 ms, the accuracy is controlled within 0.01 mm / 100 ms, the flow uniformity parameter of the molten pool internal material is obtained through the optical flow analysis, the uniformity is represented by the speed standard deviation, the standard deviation is less than 0.5 mm / s, and finally 150-dimensional molten pool feature parameters are formed, which are synchronized to the feature fusion module of the federal edge multimodal consensus network.

[0046] Preferably, step S4 comprises the following sub-steps: S41: the nodes of the federated edge multi-modal consensus network receive the thermographic feature vector output by the local edge thermographic lightweight perception model, the molten pool feature parameter output by the dynamic eddy current visual cooperative detection algorithm, and the welding process parameter collected by the workstation, and perform local feature standardization processing; S42: each node sends the standardized local feature data to the consensus coordination node through an encrypted communication link, and the consensus coordination node performs consistency checking on the received multi-node feature data; S43: the federated average consensus strategy is used to weight and fuse the multi-node feature data that passes the checking, adjust the feature spaces of the nodes to a unified dimension, and complete the feature space alignment; S44: the aligned fused feature data is fed back to each edge node to achieve global consistency of the feature data in the federated edge multi-modal consensus network.

[0047] Specifically, after step S4 is started, first, S41 is executed, the nodes of the federated edge multi-modal consensus network perform local feature standardization processing on the received thermographic feature vector, molten pool feature parameter, and welding process parameter, and adopt the Z-score standardization method, with the mean value of the thermographic feature vector set to 0.5 and the standard deviation set to 0.2; the mean value of the molten pool feature parameter set to 0.4 and the standard deviation set to 0.15; the mean value of the current in the welding process parameter set to 250A and the standard deviation set to 50A, the mean value of the voltage set to 30V and the standard deviation set to 5V, the mean value of the walking speed set to 5mm / s and the standard deviation set to 1mm / s, and after standardization, all feature values are mapped to the interval of -3 to 3. Then, S42 is entered, each node is connected to the consensus coordination node through an SSL / TLS encrypted communication link, the link bandwidth is set to 100Mbps, the transmission delay is controlled within 50ms, the sent local feature data carries the node identifier and the time stamp, and after receiving the data, the consensus coordination node stores it according to the node identifier and performs consistency checking, including data integrity and time synchronization, with the data integrity reaching more than 99% and the time stamp deviation not exceeding 10ms. Next, S43 is executed, the federated average consensus strategy is adopted, the weights are assigned according to the data quality scores of the nodes, the weight of a node with a score of 90 or above is set to 0.8, the weight of a node with a score of 70-90 is set to 0.2, the multi-node feature data is weighted and summed, and then the feature vectors of different nodes are converted to a unified dimension through a feature space mapping algorithm, with the feature space deviation controlled within 5% to complete the feature space alignment. Finally, S44 is performed, the aligned fused feature data is fed back to each edge node according to the node identifier, the feedback frequency is consistent with the data collection frequency, i.e. 120 times / s, the local feature library of each node is updated after receiving the data, the feature data error of all nodes in the network is ensured to be not more than 3%, and global consistency is achieved.

[0048] Preferably, step S5 comprises the following sub-steps: S51: calling a feature matrix initialization module, constructing an empty welding quality feature matrix framework according to the power transmission line steel structure welding quality evaluation dimensions, and the framework dimensions are determined by the number of thermal spectrum features, molten pool features and process parameter features; S52: extracting the aligned thermal spectrum feature vector from the federal edge multi-modal consensus network, converting the thermal spectrum feature vector into a thermal spectrum feature column vector corresponding to the matrix framework through a feature mapping function, and filling the matrix at the specified position; S53: extracting the aligned molten pool feature parameters and welding process parameters, respectively converting them into corresponding column vectors in the same feature mapping manner, and sequentially filling them into the corresponding columns of the matrix framework; S54: scaling the elements of the filled matrix through a matrix normalization module to ensure that each feature column vector is in the same numerical range, and forming a final welding quality feature matrix.

[0049] Specifically, after step S5 is started, first, S51 is executed, a feature matrix initialization module is called, the module loads 30 evaluation sub-dimension configurations of the welding quality of the power transmission line steel structure, including 10 forming quality sub-dimensions such as weld width consistency, excess height uniformity, 10 internal defect sub-dimensions such as pore density and crack length, and 10 mechanical property correlation sub-dimensions such as hardness distribution and tensile strength correlation value, thereby constructing a 353-row and 30-column empty feature matrix framework, the rows correspond to the total feature dimensions of the thermal spectrum, the molten pool and the process parameters, and the columns correspond to the evaluation sub-dimensions. Then, S52 is entered, the aligned thermal spectrum feature vector is extracted from the feature storage unit of the federal edge multi-modal consensus network, and when the extraction is performed, it is verified whether the deviation of the timestamp of the verification data from the current evaluation period is within 20 ms, the 200-dimensional thermal spectrum feature vector is converted into a 200x30 column vector matrix through a linear mapping function, and according to the correlation between the features and the evaluation sub-dimensions, the corresponding feature values are filled into the first 200 rows of the matrix, for example, the thermal spectrum features related to the weld width consistency are filled into the first column, and the pore density related features are filled into the 11th column. Next, S53 is executed, the same linear mapping method is used to convert the 150-dimensional molten pool feature parameters into a 150x30 column vector matrix, which is filled into the 201st-350th rows of the matrix; the 3-dimensional welding process parameters are converted into a 3x30 column vector matrix, which is filled into the 351st-353rd rows, wherein the welding current corresponds to the 1st, 11th and 21st columns, the welding voltage corresponds to the 2nd, 12th and 22nd columns, and the walking speed corresponds to the 3rd, 13th and 23rd columns, to ensure that the parameters accurately correspond to the sub-dimensions. Finally, S54 is performed, the filled matrix is processed through a min-max normalization module, and the numerical value of each element is mapped to the 0-1 interval according to the column, and the mapping is based on the historical maximum and minimum values of each column feature, for example, the maximum value of the 1st column (weld width consistency) is set to 1.0, and the minimum value is set to 0.2. The extreme value parameters of each column are recorded during the normalization process and stored in the local configuration file for subsequent reverse calculation of the evaluation results. Finally, a standardized welding quality feature matrix is formed and stored in the SSD storage device of the edge node, and the storage response time is controlled to be within 1 ms.

[0050] The federal edge multi-modal consensus network of the application is a distributed data processing and fusion architecture deployed in the welding site edge node, which is specially adapted to the multi-source data collaboration demand of the steel structure welding scene of the power transmission line. Its implementation first completes the deployment of the edge node, selects an embedded module with computing power that can process not less than 100MB of data per second, configures the number of nodes to 3-5 to cover different monitoring points, completes identity authentication, encryption protocol configuration and consensus trigger threshold setting (start when the data amount reaches 50MB) during deployment. In operation, after receiving the thermospectrum feature vector, molten pool characteristic parameter and welding process parameter, each node first processes the data by Z-score standardization method, maps the feature value to the specified interval according to the preset mean and standard deviation; then transmits the data with identifier and timestamp to the coordination node through the SSL / TLS encryption link (bandwidth 100Mbps, delay ≤50ms), the coordination node checks the data integrity (≥99%) and time synchronization (deviation ≤10ms), then uses the federal average strategy to assign weights according to the node quality score (weight 0.8 for score above 90, 0.2 for score between 70 and 90), performs weighted fusion, aligns through the feature space mapping algorithm (deviation ≤5%), and finally feeds back the fusion data to each node (frequency 120 times / s, error ≤3%). The function of the network is to realize the real-time collection, feature alignment and fusion of multi-source data on the edge side, avoid the delay problem of centralized processing, deeply mine the correlation of arc, thermospectrum, molten pool and other data. Break through the spatial limitation and collaboration bottleneck of traditional data processing, provide a comprehensive and consistent feature basis for welding quality evaluation, promote the evaluation from local sampling to global coverage upgrade, and adapt to the real-time demand in strong interference environment.

[0051] The edge thermogram lightweight perception model of the application is a lightweight intelligent algorithm model running on an edge node, focusing on feature extraction and correlation processing of the welding area thermal radiation signal. When implemented, 64 convolution kernels of 3*3 size (corresponding to different thermal feature modes) are loaded first, the thermal radiation signal in the 8-14 mu m band is received through an industrial-grade infrared sensing interface (rate 50 Mbps), and the signal sequence is formed by intercepting at 10 ms intervals (corresponding to a 120 fps acquisition frequency). After adaptive filtering (window 3*3 to 7*7) to filter high-frequency noise, the signal-to-noise ratio is improved to more than 30 dB, and the thermal features related to the arc and molten pool are retained; then the arc feature data is bound through timestamp synchronization (deviation ≤5 ms), the mapping relationship between thermal radiation intensity and arc parameters is established, and 200 groups of correlated features are generated; finally, 50 core features are selected by principal component analysis algorithm (variance contribution rate threshold 95%), compressed into a thermogram feature vector and transmitted at a rate of 20 Mbps. Its role is to accurately capture the dynamic change characteristics of the welding thermal field, convert the original thermal radiation signal into a quantifiable feature vector, and enhance the feature representation ability by correlating with the arc features. It solves the problems of large signal noise interference and isolated features in traditional thermal detection, adapts to edge computing with lightweight design, provides key thermal field dimension data support for quality evaluation, and improves the recognition basis for defects in the thermal affected zone.

[0052] The dynamic eddy current visual collaborative detection algorithm of the application is a composite detection technology that combines eddy current detection and visual analysis, used to comprehensively extract the physical state characteristics of the molten pool. The parameters are configured first: eddy current excitation 5A, 10 kHz, probe distance from welding area 50 mm, visual gray threshold 128, and gradient threshold 30. After receiving the image through the gigabit Ethernet interface (delay ≤10 ms), the molten pool core area (diameter 100 mm), transition area (width 20 mm), and background area are divided by threshold segmentation and region growing algorithm; the eddy current sub-module converts the response signal into 1920*1080 eddy current field data, identifies ≥50 feature points and analyzes the distribution rule; the visual sub-module extracts the transition area edge with the Canny operator, obtains ≥100 feature point coordinates through contour tracking, and calculates the smoothness parameter with a fitting error ≤0.5 mm; finally, the two types of features are fused with a weight of 4:6, the size change rate with an accuracy of 0.01 mm / 100 ms and the flow uniformity with a standard deviation <0.5 mm / s are calculated, and 150-dimensional parameters are formed. The role is to synchronously capture the surface morphology and internal material state characteristics of the molten pool, realize the quantitative extraction of the multi-dimensional physical parameters of the molten pool, break through the limitations of single detection technology, accurately represent the dynamic changes of the molten pool through the collaborative complementation of eddy current and vision, provide comprehensive data for early identification of welding defects, and improve the fine degree of evaluation.

[0053] The ESTUN Digital Robot Welding workstation of the present application is an integrated device integrating welding execution, data collection and quality evaluation functions, and is a hardware carrier and execution terminal of the entire evaluation method. When implemented, the built-in arc sensor (200Hz collection), high-speed industrial camera (120fps, 1920x1080 resolution), current sensor (0-500A, ±1A accuracy), voltage sensor (0-50V, ±0.1V accuracy) and displacement sensor (0-10mm / s, ±0.01mm / s accuracy) are connected to the edge node and each algorithm model through the interface. In welding, the arc, molten pool and process parameters are synchronously collected and transmitted to the edge network; in the evaluation stage, the edge processing unit reads the quality feature matrix at a rate of 200MB / s, loads the 30x1 evaluation weight matrix (sum is 1) and bias term (0.01-0.03), and after weighted summation and Sigmoid mapping (0-1 interval), it is compared with 1000 standard features (similarity ≥90% and proportion ≥70% are high quality), generates evaluation results and transmits them to the terminal in JSON format at a rate of 100Mbps (delay ≤100ms). The role is to provide a stable welding execution environment, realize synchronous collection and transmission of multi-source data, complete the final analysis and output of the evaluation results. The welding execution and quality evaluation processes are integrated to build a closed-loop system from welding operation to quality judgment, avoid the problem of poor coordination between devices, provide hardware support for real-time management and control of power transmission steel structure welding quality, and ensure the implementation of the evaluation method.

[0054] As Figure 2As shown, the power transmission line steel structure welding quality evaluation method based on edge computing is characterized in that the method is realized through different units, including: a federal edge multi-modal consensus network node deployment unit, which is directly connected with a data acquisition interface of an ESTUN Digital Robot Welding workstation, and is used to receive multi-dimensional welding process data transmitted by the workstation; an edge thermal spectrum lightweight perception signal processing unit, which is connected with the federal edge multi-modal consensus network node deployment unit through a data bus, and is used to extract features of thermal radiation signals of a welding area and transmit the features to the consensus network; a dynamic eddy current visual cooperative detection operation unit, which is connected with the federal edge multi-modal consensus network node deployment unit through a communication interface, and is used to analyze molten pool shape data and output characteristic parameters; a multi-modal feature consensus alignment unit, which is bidirectionally connected with the edge thermal spectrum lightweight perception signal processing unit, the dynamic eddy current visual cooperative detection operation unit and the federal edge multi-modal consensus network node deployment unit, respectively, and is used to perform consistency processing on multi-source features; a welding quality feature matrix construction unit, which is connected with an output end of the multi-modal feature consensus alignment unit, and is used to construct a quality feature matrix based on aligned features; and an edge evaluation result output unit, which is connected with an output end of the welding quality feature matrix construction unit, and is in communication connection with an edge processing unit of the ESTUN Digital Robot Welding workstation, and is used to analyze the matrix and output an evaluation result.

[0055] The power transmission line steel structure welding quality evaluation method based on edge computing has obvious advantages in data processing and feature fusion, and effectively overcomes the defects of insufficient collaboration of existing technologies. By deploying a federal edge multi-modal consensus network on an edge node, the traditional mode of relying on a centralized platform is abandoned, and welding current, voltage and other process parameters and thermal radiation, molten pool shape and other detection data are directly summarized on the edge side. Through the consensus mechanism built in the network, the data of each node is efficiently aligned, not only greatly reducing the data transmission delay, but also deeply mining the potential association between multi-source data, constructing a comprehensive and accurate quality feature representation, and completely solving the problems of scattered data processing and low fusion efficiency in existing technologies.

[0056] In terms of perception detection accuracy, the method realizes breakthrough through the coordinated linkage of models and algorithms, and makes up for the one-sidedness defects of existing detection. The edge thermal spectrum lightweight perception model and the dynamic eddy current visual cooperative detection algorithm are no longer independently operated, but establish a deeply related processing flow. The thermal field features captured by the thermal spectrum perception model and the molten pool profile and flow state features extracted by the eddy current visual algorithm form complementary verification, accurately capture the linkage law of thermal field changes and molten pool flow in the welding process, and significantly improve the early identification ability of subtle defects, effectively solving the drawbacks that existing single detection methods cannot fully reflect the welding quality.

[0057] The method also has the practical advantage of adapting to the scene requirements, further strengthening the overcoming effect of the short board of the prior art. The overall architecture design of the edge side makes the network, model and algorithm close to the welding site operation, and the data acquisition, processing and evaluation result output form an efficient closed loop, ensuring the real-time of the evaluation. This design not only adapts to the characteristics of strong interference and dynamic change in the transmission line steel structure welding environment, but also through the deep integration of multiple technical elements, changes the quality evaluation from partial sampling inspection to whole process accurate monitoring, fully meets the strict requirements of the scene on the timeliness and refinement of the evaluation.

[0058] In the description of the present application, it should be noted that unless otherwise specified and limited, the terms "set", "install", "connect", "connect", "fix" should be understood broadly, for example, it can be fixedly connected, or it can be detachably connected, or integrally connected; it can be mechanically connected, or it can be electrically connected; it can be directly connected, or it can be indirectly connected through an intermediate medium, or it can be the communication inside two elements. For those skilled in the art, the specific meaning of the above terms in the present application can be understood according to the specific circumstances.

[0059] Although the embodiments of the present application have been shown and described, it can be understood by those skilled in the art that various equivalent changes, modifications, replacements and variations of the embodiments can be made without departing from the principles and spirits of the present application, and the scope of the present application is defined by the appended claims and their equivalent scope.

Claims

1. A method for evaluating the welding quality of steel structures in power transmission and transformation lines based on edge computing, characterized in that, include: Step S1: Deploy the federated edge multimodal consensus network at the edge node of the ESTUN Digital Robot Welding workstation. Collect data on arc characteristics, molten pool morphology, welding current, welding voltage, and walking speed during the welding process of power transmission line steel structures through the built-in sensor module of the workstation, and transmit the data to the federated edge multimodal consensus network node. Step S2: Activate the edge thermal spectrum lightweight sensing model to sense and capture the thermal radiation signal of the welding area received by the edge node, generate a thermal spectrum feature vector by combining the arc feature data, and transmit it to the federated edge multimodal consensus network. Step S3: Call the dynamic eddy current visual collaborative detection algorithm to collaboratively process the molten pool morphology data, extract the molten pool boundary contour, size change, and flow state feature parameters, and synchronize them to the federated edge multimodal consensus network; Step S4: The federated edge multimodal consensus network performs multimodal data fusion on the received thermal spectrum feature vector, molten pool feature parameters, welding current, welding voltage, and walking speed data, and completes the data feature alignment between nodes through the consensus mechanism; Step S5: Based on the aligned fused feature data, and combining the output results of the edge thermal spectrum lightweight perception model and the dynamic eddy current visual collaborative detection algorithm, construct the welding quality feature matrix; Step S6: The edge processing unit of the ESTUN Digital Robot Welding workstation analyzes the welding quality feature matrix and outputs the welding quality assessment results of the steel structure of the power transmission line.

2. The edge computing-based method for evaluating the welding quality of steel structures in power transmission and transformation lines according to claim 1, characterized in that, The feature fusion expression of the federated edge multimodal consensus network is as follows: ,in Represents multimodal fusion features, Represents the number of federated edge nodes. Representing the The weight factor of each node, Representing the The parameter matrix is ​​composed of welding current, voltage, and travel speed collected by each node. Representing the The thermal spectral feature vector output by the lightweight sensing model of node edge thermal spectrum. Representing the The molten pool feature parameters output by the node dynamic eddy current visual collaborative detection algorithm Representative feature splicing operation, The alignment matrix represents the consensus mechanism's generated values. This represents the matrix dot product operation.

3. The edge computing-based method for evaluating the welding quality of steel structures in power transmission lines according to claim 1, characterized in that, The thermal spectral feature extraction expression of the edge thermal spectrum lightweight sensing model is as follows: ,in Represents the thermal spectral eigenvector. Represents the activation function. Represents a two-dimensional convolution operation. Represents the thermal sensing convolution kernel. Represents convolution operation. Represents the thermal radiation signal of the welding area. This represents the thermal spectrum sensing bias term. Represents element-wise multiplication, and Pool represents pooling operations. Represents the characteristic mapping function of electric arc. This represents the characteristic data of the electric arc.

4. The edge computing-based method for evaluating the welding quality of steel structures in power transmission and transformation lines according to claim 1, characterized in that, The expression for the molten pool feature extraction of the dynamic eddy current visual collaborative detection algorithm is as follows: ,in Represents the characteristic parameters of the molten pool. Represents gradient operation, Edge represents the edge detection operator, and E represents the vortex field feature extraction function. Data representing the morphology of the molten pool. Representative feature intersection operation. Represents the flow state analysis function. Represents the visual feature transformation function. Represents the visual collaborative weights of eddy currents.

5. The edge computing-based method for evaluating the welding quality of steel structures in power transmission lines according to claim 1, characterized in that, The expression for constructing the welding quality feature matrix is ​​as follows: ,in Represents the welding quality characteristic matrix. Represents the thermal spectral eigenvector. Represents the thermal spectrum feature weight matrix. Represents the characteristic parameters of the molten pool. Represents the feature weight matrix of the molten pool. A matrix representing welding current, voltage, and travel speed parameters. Represents the weight matrix of welding process parameters. Represents the federal consensus alignment transformation matrix. This represents the matrix transpose operation.

6. The edge computing-based method for evaluating the welding quality of steel structures in power transmission lines according to claim 1, characterized in that, The evaluation result output expression of the ESTUN Digital Robot Welding workstation is as follows: ,in The sigmoid function represents the sigmoid activation function and represents the welding quality assessment result. Represents the welding quality characteristic matrix. Represents the evaluation weight matrix. Represents the evaluation bias term. Represents the set of standard characteristics for welding quality. This represents the feature matching operation.

7. The edge computing-based method for evaluating the welding quality of steel structures in power transmission lines according to claim 1, characterized in that, Step S2 includes the following sub-steps: S21: The edge thermal spectrum lightweight sensing model receives real-time thermal radiation signals from the welding area through the infrared sensing interface of the ESTUN Digital Robot Welding workstation, and segments the signals at preset time intervals to obtain a continuous thermal radiation signal sequence. S22: Call the built-in thermal signal enhancement module of the model to filter noise in the intercepted signal sequence, retain the thermal radiation characteristic components related to the welding arc and molten pool area, and form a preliminary processed thermal signal; S23: Combining the arc characteristic data transmitted in step S1, establish a mapping relationship between thermal radiation signal and arc intensity and frequency characteristics through the feature association module, and generate associated feature groups. S24: The model's feature dimensionality reduction module compresses the dimensions of the associated feature groups, extracts the core heat spectrum features, constructs a heat spectrum feature vector, and sends it to the federated edge multimodal consensus network.

8. The edge computing-based method for evaluating the welding quality of steel structures in power transmission lines according to claim 1, characterized in that, Step S3 includes the following sub-steps: S31: The dynamic eddy current visual collaborative detection algorithm receives the molten pool morphology image data transmitted by the visual acquisition module of the ESTUN Digital Robot Welding workstation, and divides the image into three regions: the molten pool core area, the transition area, and the background area through the image segmentation module. S32: Start the eddy current detection submodule to perform eddy current field simulation calculations on the image of the core area of ​​the molten pool, and extract the eddy current feature points and distribution patterns generated by the material flow inside the molten pool. S33: The visual inspection submodule performs boundary recognition of the molten pool transition zone and uses a contour tracking algorithm to obtain the molten pool boundary contour coordinates and contour smoothness data. S34: The distribution pattern of eddy current feature points is fused with boundary contour coordinates and contour smoothness data to calculate the rate of change of molten pool size and flow uniformity parameters, forming molten pool feature parameters and synchronizing them to the federated edge multimodal consensus network.

9. The edge computing-based method for evaluating the welding quality of steel structures in power transmission lines according to claim 1, characterized in that, Step S4 includes the following sub-steps: S41: Each node of the federated edge multimodal consensus network receives the thermal spectrum feature vector output by the local edge thermal spectrum lightweight perception model, the molten pool feature parameters output by the dynamic eddy current visual collaborative detection algorithm, and the welding process parameters collected by the workstation, and performs local feature standardization processing. S42: Each node sends its standardized local feature data to the consensus coordination node through an encrypted communication link. The consensus coordination node performs consistency verification on the received multi-node feature data. S43: The federated average consensus strategy is used to weight and fuse the verified multi-node feature data, adjust the feature space of each node to a unified dimension, and complete the feature space alignment. S44: Feed back the aligned and fused feature data to each edge node to achieve global consistency of feature data within the federated edge multimodal consensus network.

10. The method for evaluating the welding quality of steel structures in transmission and transformation lines based on edge computing according to claim 1, characterized in that, Step S5 It includes the following steps: S51: Call the feature matrix initialization module to construct an empty welding quality feature matrix framework based on the evaluation dimensions of welding quality of steel structure of transmission and transformation lines. The framework dimensions are determined by the number of thermal spectrum features, molten pool features and process parameter features. S52: Extract aligned hot spectrum feature vectors from the federated edge multimodal consensus network, convert the hot spectrum feature vectors into hot spectrum feature column vectors corresponding to the matrix frame through the feature mapping function, and fill them into the matrix at the set positions; S53: Extract the aligned molten pool feature parameters and welding process parameters, convert them into corresponding column vectors using the same feature mapping method, and fill them into the corresponding columns of the matrix frame in sequence; S54: The matrix normalization module scales the elements of the filled matrix to ensure that each feature column vector is within the same numerical range, thus forming the final welding quality feature matrix.

Images