Intelligent welding system for electric pole steel reinforcement framework

Abstract

The invention relates to the technical field of welding automation, and discloses an intelligent welding system for an electric pole steel reinforcement framework. The system comprises a welding execution module, a multi-mode sensing module, an edge calculation module, a real-time regulation and control module and a system control center. Synchronous collection and real-time analysis of images, temperature and sound signals of a welding pool area are achieved, so that defects such as air holes, slag inclusion and incomplete fusion can be recognized online in a key time window when the welding pool is not solidified, and time delay caused by a traditional off-line detection mode is thoroughly eliminated. Through a real-time regulation and control module, on the basis of a multi-target optimization decision and parameter self-adaptive adjustment mechanism, the system can dynamically adjust key process parameters such as welding current, arc voltage and welding speed at the moment of identifying defect germination or process deviation, and real-time intervention and closed-loop control on welding quality are achieved; therefore, the final formation of defect welding seams is avoided fundamentally, and the rework rate and the material waste are remarkably reduced.

Description

Technical Field

[0001] This invention belongs to the field of welding automation technology, specifically relating to an intelligent welding system for the steel reinforcement cage of utility poles. Background Technology

[0002] In the field of power infrastructure construction, concrete poles serve as key supporting structures for transmission and distribution lines. The welding quality of their internal steel reinforcement cage directly determines the overall mechanical performance and long-term service safety of the pole. The steel reinforcement cage is typically formed by welding main bars and spiral stirrups together. The stability and reliability of the welding process are crucial to ensuring the structural strength of the pole.

[0003] The automated welding system for the rebar cage of utility poles aims to improve welding efficiency and consistency by using a robotic arm, welding power source, and wire feeding device to automatically form welds. The system's fundamental goal is to replace traditional manual welding methods by completing the welding of rebar joints under pre-programmed control.

[0004] Current technologies generally employ offline quality inspection after welding, identifying welding defects through ultrasonic, X-ray, or visual inspection after the weld has completely cooled. However, this offline inspection mode has significant limitations: defects such as porosity, slag inclusions, or lack of fusion generated during welding can only be effectively observed during the cooling stage, resulting in a significant time delay between defect identification and welding operations; real-time intervention and process parameter adjustments cannot be made within the critical time window before the molten pool solidifies; and the discovery of any defect means that the already welded framework must be cut off, reassembled, and re-welded, causing material waste, project delays, and a significant increase in rework costs. Therefore, how to achieve online defect identification and real-time control during the welding process has become a pressing technical challenge in the field of intelligent welding of pole reinforcement frameworks. Summary of the Invention

[0005] The present invention aims to provide an intelligent welding system for the steel reinforcement cage of utility poles, in order to solve the problems of lagging welding defect identification, inability to intervene in real time during the welding process, and high rework costs in the existing technology.

[0006] The technical solution of this invention is that the system includes a welding execution module, a multimodal sensing module, an edge computing module, a real-time control module, and a system control center. The welding execution module consists of a multi-joint robotic arm, a digital welding power supply, and a high-precision wire feeding mechanism, and is used to complete the automated welding operation of the main reinforcement and spiral stirrup nodes of the pole's steel reinforcement skeleton under the drive of a preset program.

[0007] The multimodal sensing module is integrated into the end effector of the welding execution module. It includes a high-speed vision sensor, an infrared thermal imager, and an arc acoustic emission sensor, which are used to acquire image sequences, temperature field distribution data, and arc sound signals of the weld pool area, respectively.

[0008] The edge computing module is deployed at the welding equipment site. It connects to the multimodal sensing module via a high-speed data bus to perform real-time fusion analysis and feature extraction on the acquired multimodal sensing data. The real-time control module communicates bidirectionally with both the edge computing module and the welding execution module to dynamically adjust the process parameters of the welding execution module based on the analysis results output by the edge computing module. The system control center, as the scheduling core of the entire system, coordinates the data flow and control commands between the modules and maintains the stable operation of the welding process.

[0009] Furthermore, the edge computing module internally includes an online welding defect identification unit and a process status assessment unit. The online welding defect identification unit is built on a convolutional neural network architecture. Its input layer receives image, temperature, and sound data streams synchronously acquired by the multimodal sensing module. The network contains multiple convolutional and pooling layers to extract the spatial and temporal features of the sensing data step by step, and outputs the initial probability distribution of porosity, inclusions, and lack of fusion defects through a fully connected layer.

[0010] The process status assessment unit adopts a dynamic assessment model based on long short-term memory network. Its inputs are the high-dimensional feature sequence extracted by the online welding defect identification unit and the real-time process parameters fed back by the welding execution module. The model captures the dynamic evolution law of the welding process through the state of memory cells and outputs a comprehensive assessment index that characterizes the stability of the welding process.

[0011] Furthermore, the real-time control module includes a multi-objective optimization decision-making unit and a parameter adaptive adjustment unit. The multi-objective optimization decision-making unit uses welding quality, welding efficiency, and energy consumption as optimization objectives, establishing a comprehensive evaluation system that includes a weld formation quality scoring function, a welding cycle function, and an energy consumption function. This unit receives the comprehensive evaluation index output by the process status evaluation unit and, based on preset constraints, solves for the optimal process parameter adjustment amount that optimizes the comprehensive evaluation system. The parameter adaptive adjustment unit generates specific control commands based on the process parameter adjustment amount solved by the multi-objective optimization decision-making unit and sends them to the welding execution module. These control commands include amplitude adjustment of the welding current, stability compensation of the arc voltage, and linear correction of the welding speed.

[0012] Furthermore, the system control center integrates a welding knowledge base and a self-learning engine. The welding knowledge base stores historical welding process parameters, corresponding multimodal sensor data, and final welding quality assessment results, forming a welding process atlas. The self-learning engine periodically accesses the welding knowledge base, using newly accumulated welding process data to incrementally train the convolutional neural network and long short-term memory network in the edge computing module, achieving online updates of model parameters and continuous improvement of recognition and evaluation accuracy.

[0013] Furthermore, the data synchronization mechanism of the multimodal sensing module is implemented through hardware trigger signals. The high-speed visual sensor, infrared thermal imager, and arc acoustic emission sensor share a global synchronization clock signal issued by the system control center, ensuring that each frame of image, each set of temperature data, and each segment of sound signal has a strictly consistent timestamp, providing a basis for timing alignment for the multimodal data fusion of the edge computing module.

[0014] Furthermore, the convolutional neural network of the online welding defect identification unit adopts an encoder-decoder structure. The encoder part compresses the multimodal input data into a high-dimensional feature vector through convolution and pooling operations; the decoder part reconstructs the feature vector into a defect probability distribution map that matches the spatial resolution of the original input data through deconvolution and upsampling operations, thereby achieving pixel-level localization of welding defects in the molten pool area.

[0015] Furthermore, the long short-term memory network of the process status evaluation unit includes forget gates, input gates, and output gates. The forget gate determines which information to discard from the cell state based on the current input and the hidden state of the previous time step; the input gate controls which new information will be stored in the cell state; and the output gate determines the hidden state output at this moment based on the current cell state and the input information. Through these three gating mechanisms, the network can effectively capture the long-term dependencies of the process status during welding.

[0016] Furthermore, the parameter adaptive adjustment unit employs a fuzzy proportional-integral-derivative (PID) control algorithm to regulate the welding current. This algorithm uses the current adjustment output by the multi-objective optimization decision unit as the setpoint and the actual current value fed back from the welding power source as the process variable. By calculating the deviation between the setpoint and the process variable, as well as the rate of change of the deviation, it dynamically adjusts the proportional coefficient, integral time, and derivative time to achieve rapid and accurate tracking of the welding current.

[0017] Furthermore, the system control center also constructs a digital twin model of the welding process. This model receives real-time status data from the welding execution module, data collected by the multimodal sensing module, and analysis results from the edge computing module, synchronously mapping the complete state of the physical welding process in virtual space, providing a visualization platform for offline simulation of the welding process, parameter preset, and abnormal condition simulation.

[0018] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention integrates a multimodal sensing module and an edge computing module to achieve synchronous acquisition and real-time analysis of images, temperature and sound signals of the weld pool area. This enables defects such as porosity, slag inclusions and lack of fusion to be identified online within the critical time window before the weld pool solidifies, completely eliminating the time delay caused by traditional offline detection modes.

[0019] 2. Through the real-time control module based on multi-objective optimization decision-making and parameter adaptive adjustment mechanism, the system can dynamically adjust key process parameters such as welding current, arc voltage and welding speed at the moment when defects are detected or process deviations are detected, so as to realize real-time intervention and closed-loop control of welding quality, thereby avoiding the final formation of defective welds at the source and significantly reducing rework rate and material waste.

[0020] 3. The integrated welding knowledge base and self-learning engine in the system control center enable the system to continuously learn and optimize from historical welding data, thereby continuously improving the accuracy of defect identification and the precision of process control, forming an intelligent welding system with continuous evolution capabilities. The introduction of a digital twin model of the welding process provides an efficient digital tool for the optimization and verification of welding processes, further enhancing the system's intelligence level and engineering practical value. Attached Figure Description

[0021] Figure 1 This is a schematic diagram of the overall technical solution architecture of the intelligent welding system for the steel reinforcement cage of utility poles proposed in this invention; Figure 2 This is a schematic diagram of the core principle framework for real-time fusion analysis and feature extraction of multimodal sensing data in this invention; Figure 3 This is a logical flowchart of the online identification of welding defects and process status assessment in this invention; Figure 4 This is a flowchart illustrating the logical process of multi-objective optimization decision-making and adaptive parameter adjustment in this invention. Figure 5 This is a schematic diagram of the multi-level interaction relationship and data flow between the welding knowledge base and the self-learning engine in this invention; Figure 6 This is a schematic diagram illustrating the principle framework of the digital twin model of the welding process in this invention. Detailed Implementation

[0022] Example 1: Please refer to the appendix Figure 1-6This embodiment provides a complete implementation plan for an intelligent welding system for the steel reinforcement cage of utility poles. The system consists of five core components: a welding execution module, a multimodal sensing module, an edge computing module, a real-time control module, and a system control center. Each module is interconnected with a real-time data bus via a high-speed industrial Ethernet, forming a closed-loop control architecture. The welding execution module specifically includes a six-degree-of-freedom multi-joint robotic arm, a fully digital inverter welding power supply, and a high-precision wire feeding mechanism driven by a servo motor.

[0023] The multi-joint robotic arm uses an absolute encoder to provide feedback on joint angles, achieving a repeatability accuracy of ±0.05 mm. The digital welding power supply supports multiple output modes including pulse, DC, and AC, with a current adjustment range of 50 to 500 amps and a voltage adjustment range of 15 to 40 volts. The high-precision wire feeding mechanism employs a dual-drive wheel design, allowing for continuous adjustment of the wire feeding speed between 2 and 20 meters per minute, ensuring stable wire delivery. Driven by motion trajectory and process parameter commands issued by the system control center, the welding execution module automatically completes all-position welding operations at the intersections of main reinforcement bars and spiral stirrups in the pole's steel reinforcement cage.

[0024] The multimodal sensing module is directly integrated into the end effector of the welding execution module. Its physical encapsulation utilizes a high-temperature resistant alloy shell and forced air cooling design to ensure long-term stable operation under the high-temperature radiation environment of the welding arc. This module includes three types of sensors: The system comprises a high-speed visual sensor with a resolution of 2 megapixels, an infrared thermal imager with a thermal sensitivity of 0.05 degrees Celsius, and an arc acoustic emission sensor with a frequency response range of 20 Hz to 20 kHz. The high-speed visual sensor acquires image sequences of the weld pool area at a rate of 1000 frames per second, with the image format being 12-bit grayscale. The infrared thermal imager acquires two-dimensional temperature field distribution data of the weld pool and its surrounding heat-affected zone at a sampling frequency of 100 Hz, with a temperature measurement range of 300 to 3000 degrees Celsius. The arc acoustic emission sensor collects arc sound signals during the welding process using piezoelectric ceramic elements, with a signal dynamic range of 40 to 120 dB. The data synchronization mechanism of the multimodal sensing module is implemented through hardware triggering. The system control center generates a global synchronization clock signal, which is transmitted to each sensor at a 5-volt transistor logic level, ensuring that each image frame, each set of temperature data, and each segment of sound signal carries a timestamp with nanosecond precision, providing a foundation for multimodal data timing alignment.

[0025] The edge computing module is deployed in a shockproof and dustproof cabinet at the welding equipment site. Its hardware platform is built on the NVIDIA Zwinga architecture edge computing unit, equipped with an 8-core 64-bit processor and a 384-core graphics processor, achieving a floating-point operation capability of 20 trillion operations per second. This module connects to the multimodal sensing module via a 4-channel 10 Gigabit Ethernet interface, receiving and caching multimodal sensing data in real time. Internally, the edge computing module houses two core processing engines: an online welding defect identification unit and a process status evaluation unit. The online welding defect identification unit is built on a convolutional neural network architecture. The network input layer receives image sequences acquired by a high-speed visual sensor, temperature field data acquired by an infrared thermal imager, and sound signals acquired by an arc acoustic emission sensor, which are then preprocessed to form a multi-channel tensor.

[0026] The preprocessing workflow includes gamma correction and histogram equalization of image data, noise filtering and normalization of temperature data, and fast Fourier transform and Mel-frequency cepstral coefficient feature extraction of sound signals. The convolutional neural network employs an encoder-decoder structure. The encoder consists of five convolutional layers and three max-pooling layers, all with a 3×3 kernel size, a stride of 1, and uniform padding. Each convolutional layer is followed by a batch normalization layer and a modified linear unit activation function. Through convolution and pooling operations, the multimodal input data is progressively compressed into a 512-dimensional feature vector. The decoder consists of three deconvolutional layers and two upsampling layers, with a 3×3 kernel size and a stride of 2. The final output layer generates a defect probability distribution map with the same spatial resolution as the original input image through 1×1 convolution and a soft maximum activation function. Each pixel in the map corresponds to the probability value of three defects: porosity, inclusion, and lack of fusion. During the training phase, the network uses a cross-entropy loss function and an adaptive moment estimation algorithm optimizer, with an initial learning rate of 0.001 and a batch size of 32.

[0027] The process status evaluation unit employs a dynamic evaluation model based on a Long Short-Term Memory (LSTM) network. Its input consists of a 512-dimensional feature sequence extracted by the online welding defect identification unit and a vector of process parameters fed back in real-time from the welding execution module, including welding current, arc voltage, welding speed, and wire feed speed. The LSTM network comprises two hidden layers, each with 256 hidden units. The network incorporates a triple gating mechanism: a forget gate, an input gate, and an output gate. The forget gate calculates a forgetting factor based on the current input and the hidden state from the previous time step, determining which historical information to discard from the cell state. The input gate calculates the storage weights for new information using an input activation function and candidate cell states. The output gate calculates its activation value based on the current cell state and input information, controlling the output content of the hidden state at that moment. Through this gating mechanism, the network effectively captures the long-term dependencies of the process status during welding. The model output is a comprehensive evaluation index, ranging from 0 to 1, with higher values ​​indicating a more stable welding process. The comprehensive evaluation index is calculated by multiplying the L2 norm of the hidden state vector by the weight matrix of the fully connected layer and then activating it using a logistic function.

[0028] The real-time control module establishes bidirectional communication with the edge computing module and welding execution module via a process fieldbus. This module includes a multi-objective optimization decision-making unit and a parameter adaptive adjustment unit. The multi-objective optimization decision-making unit establishes a comprehensive evaluation system with welding quality, welding efficiency, and energy consumption as core objectives. Welding quality is quantified using a weld formation quality scoring function, which comprehensively considers four geometric characteristics: weld width, reinforcement height, penetration depth, and surface smoothness. Welding efficiency is characterized by a welding period function, defined as the time required for a single welding node. Energy consumption is described by an energy consumption function, calculating the total electrical energy and shielding gas consumption during the welding process.

[0029] The multi-objective optimization decision unit receives the comprehensive evaluation index output by the process status evaluation unit. When the index falls below a threshold of 0.8, optimization calculations are triggered. The optimization problem is formulated as finding the optimal combination of parameters for the comprehensive evaluation function within the feasible region of welding current, arc voltage, and welding speed. The feasible region boundary is preset based on the welding material type and specifications, typically ranging from 200 to 400 amperes for current, 22 to 32 volts for voltage, and 3 to 10 millimeters per second for speed. The optimization algorithm employs a second-generation non-dominated sorting genetic algorithm with a population size of 100 and a maximum of 200 iterations.

[0030] The parameter adaptive adjustment unit generates specific control commands based on the process parameter adjustment amounts output by the multi-objective optimization decision unit. The welding current is adjusted using a fuzzy proportional-integral-derivative (FID) control algorithm, which takes the deviation and rate of change of the deviation between the current setpoint output by the multi-objective optimization decision unit and the actual current value fed back by the welding power source as input. The fuzzification process converts precise quantities into seven linguistic variables: negative large, negative medium, negative small, zero, positive small, positive medium, and positive large. The fuzzy inference rule base contains 40 rules, and the centroid method is used for defuzzification. The fuzzy adjustment rules for the proportional coefficient, integral time, and derivative time are formulated based on expert experience, ultimately outputting precise control quantities that act on the current regulation circuit of the digital welding power source. A feedforward control strategy is used for arc voltage stability compensation, calculating the voltage compensation amount in real time based on welding speed changes. Linear correction of the welding speed is achieved through replanning the spatial trajectory of the robotic arm joints, and speed adjustment commands are sent to the robotic arm controller through the robot operating system middleware.

[0031] The system control center adopts a collaborative architecture of industrial-grade programmable logic controllers (PLCs) and industrial control computers (ICCs). The PLCs are responsible for real-time control task scheduling, while the ICCs run advanced application algorithms. The system control center integrates a welding knowledge base and a self-learning engine. The welding knowledge base uses a hybrid storage scheme of time-series and relational databases to record complete process data for each welding node, including raw multimodal sensor data, edge computing feature extraction results, real-time control parameter adjustment records, and the final weld quality inspection report.

[0032] The data storage period is 3 years, supporting multi-dimensional retrieval by time range, process parameter interval, and quality level. The self-learning engine automatically initiates an incremental training process every 24 hours, extracting the latest 1000 welding process records from the welding knowledge base to fine-tune the convolutional neural network and long short-term memory network in the edge computing module. Training data is divided into training, validation, and test sets in a 7:2:1 ratio. An early stopping strategy is employed during training to prevent overfitting; training terminates if the validation set loss does not decrease for 10 consecutive cycles. A hot-swap mechanism is used for model updates to ensure uninterrupted system operation.

[0033] The system control center also constructs a digital twin model of the welding process, which uses a unified modeling language and solid mapping technology to virtually map the physical welding process. The digital twin model receives real-time data from the robotic arm joint angles and welding torch pose of the welding execution module, the molten pool image, temperature field, and sound characteristics of the multimodal sensing module, and the defect probability distribution and comprehensive evaluation index of the edge computing module. The dynamic morphology of the weld pool, temperature field cloud map, and sound pressure level distribution are simultaneously rendered in the virtual scene, and the defect probability is visualized using color coding. The digital twin model supports offline simulation of process parameters, allowing users to adjust welding current, voltage, and speed parameters in the virtual environment to predict weld quality and defect probability. The abnormal operating condition simulation function can simulate fault scenarios such as wire feeding blockage and arc instability to assess the system's fault tolerance. The visualization platform is developed based on web-based graphics library technology, supporting multi-terminal access and real-time data dashboard display.

[0034] In this embodiment, the installation position of the multimodal sensing module is precisely calibrated to ensure that the angle between the optical axis of the high-speed vision sensor and the welding torch axis is 45 degrees, and the working distance remains constant at 15 cm. The infrared thermal imager's field of view is set to 30 degrees, and the focal length is automatically adjusted. The arc acoustic emission sensor is installed on the side wall of the welding torch clamping device, 8 cm from the center of the arc. The data processing pipeline of the edge computing module adopts a double buffering mechanism to ensure the parallel execution of data acquisition and analysis. The control cycle of the real-time control module is 10 milliseconds, and the parameter adjustment command transmission delay is less than 1 millisecond. The system control center provides a human-machine interface through the operator station, displaying the real-time curve of the welding process, defect alarm information, and system operating status. This intelligent welding system is suitable for welding the reinforcing steel skeleton of ring-shaped reinforced concrete poles with diameters of 200 to 600 mm and lengths of 6 to 15 meters. The typical number of welding nodes is 200 to 500, and the system's continuous operation reliability reaches 99.5%.

[0035] Example 2: Please refer to the appendix Figure 2 This embodiment provides an alternative implementation scheme for an intelligent welding system for the steel reinforcement cage of utility poles, focusing on the optimized design of the multimodal sensor data fusion and analysis process. In this embodiment, the edge computing module adopts a hierarchical feature fusion strategy, fusing visual features, thermal features, and acoustic features at different abstraction levels. Visual feature extraction uses a residual network 50 architecture, performing transfer learning based on pre-trained weights from an image classification task. Thermal features are extracted using a 3D convolutional network to extract the spatiotemporal variation pattern of the temperature field, with a convolutional kernel size of 3×3×3. Acoustic features are extracted using a depth waveform encoder, which consists of a stack of one-dimensional convolutional layers and gated recurrent units.

[0036] The feature vectors of the three modalities are concatenated before the fully connected layer, resulting in a 1024-dimensional feature set. The fused features are weighted using an attention mechanism, with the attention weights dynamically calculated by a feature importance scoring network. This scoring network is a two-layer fully connected network; its input is the statistics of each modal feature, and its output is a weight coefficient between 0 and 1. The weighted fused features are then input to the classification head of the online welding defect identification unit. This classification head consists of two fully connected layers and a random deactivation layer with a deactivation rate of 0.3.

[0037] In this embodiment, the process status assessment unit uses a spatiotemporal graph convolutional network instead of a long short-term memory network. This network models the welding process as a dynamic graph structure, where nodes represent key physical quantities in the welding process, and edges represent the interactions between these physical quantities. Node features include 12 dimensions such as molten pool area, average temperature, sound pressure level, and current fluctuation coefficient. The graph structure is updated every 50 milliseconds, capturing local correlations between nodes through graph convolution operations. The spatiotemporal graph convolutional network contains three spatiotemporal convolutional modules, each consisting of a spatial graph convolution and a temporal convolution. The spatial graph convolution uses a Chebyshev polynomial approximation with an order of 3. The temporal convolution uses a one-dimensional convolutional layer with a kernel size of 5. Finally, a comprehensive evaluation index is output through global average pooling and a fully connected layer. This architecture can explicitly model the multi-physics coupling relationships in the welding process, improving the accuracy of the status assessment.

[0038] In this embodiment, the multi-objective optimization decision-making unit of the real-time control module employs a Bayesian optimization algorithm to solve for the process parameter adjustment. This algorithm constructs a Gaussian process regression model of welding quality and process parameters, guiding parameter space exploration through a data acquisition function. The data acquisition function is selected based on desired improvement, balancing the conflict between exploration and utilization. The optimization process iterates 20 times, evaluating a set of process parameter combinations in each iteration. The parameter adaptive adjustment unit introduces a model predictive control strategy, constructing a linear time-varying state-space model of the welding process. The model state variables include weld pool width, weld depth, and heat input, while the control variables are welding current, arc voltage, and welding speed. The controller solves for the optimal control sequence within a finite time domain in each control cycle, implementing only the first control variable. This strategy can explicitly handle system constraints, enhancing the robustness of the control process.

[0039] In this embodiment, the self-learning engine of the system control center integrates a meta-learning framework, enabling the system to quickly adapt to new steel reinforcement materials or welding positions. The meta-learner employs a model-independent meta-learning algorithm to learn model initialization parameters on multiple welding tasks. When encountering a new task, the model can achieve good performance through gradient updates using a small number of samples. The welding knowledge base expands to store material chemical composition, mechanical properties, and weldability evaluation data, forming a material welding process knowledge graph. The digital twin model adds multi-physics coupled simulation capabilities, calculating the welding thermal process, stress-strain field, and microstructure evolution based on the finite element method. Simulation results are compared with real-time sensor data to verify and correct the empirical model.

[0040] This embodiment deploys a system on a mobile welding workstation, which integrates a diesel generator set and hydraulic leveling outriggers, making it suitable for field construction sites. The welding execution module uses a lightweight collaborative robotic arm with a load capacity of 10 kg and a repeatability accuracy of ±0.1 mm. The multimodal sensing module employs a wireless transmission scheme, transmitting sensor data via 60 Hz millimeter-wave communication with a transmission latency of less than 5 milliseconds. The edge computing module uses a field-programmable gate array to accelerate feature extraction algorithms, reducing processing latency to 2 milliseconds. The system control center is equipped with a five-level network security protection system, including firewall, intrusion detection, data encryption, access control, and security auditing, ensuring data security in the industrial internet environment. This implementation significantly improves the system's mobility and environmental adaptability while ensuring welding quality, making it particularly suitable for power grid emergency repairs and the construction of power pole production lines in remote areas.

[0041] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A pole reinforcement framework intelligent welding system, characterized in that, The system comprises: a welding execution module for completing the automatic welding work of the main reinforcement and spiral stirrup joints of the steel reinforcement framework of the electric pole under the preset program driving; a multi-modal sensing module integrated on the end effector of the welding execution module, which contains a high-speed visual sensor, an infrared thermal imager and an electric arc acoustic emission sensor, respectively used for collecting image sequences, temperature field distribution data and electric arc sound signals in the welding pool area; an edge computing module deployed on the welding equipment site, connected with the multi-modal sensing module through a high-speed data bus, used for real-time fusion analysis and feature extraction of the collected multi-modal sensing data; a real-time control module in bidirectional communication with the edge computing module and the welding execution module, used for dynamically adjusting the process parameters of the welding execution module based on the analysis results output by the edge computing module; a system control center as the scheduling core of the whole system, coordinating the data flow and control instructions among the modules and maintaining the stable operation of the welding process; the edge computing module is internally provided with a welding defect online identification unit and a process state evaluation unit; the welding defect online identification unit is based on a convolutional neural network architecture, the input layer of which receives the image, temperature and sound data streams synchronously collected by the multi-modal sensing module, the network contains multiple convolutional layers and pooling layers for progressively extracting spatial and temporal features of the sensing data, and outputs the initial probability distribution of porosity, slag inclusion and incomplete fusion defects through a fully connected layer; the process state evaluation unit adopts a dynamic evaluation model based on a long short-term memory network, the input of which is the high-dimensional feature sequence extracted by the welding defect online identification unit and the real-time process parameters fed back by the welding execution module, the model captures the dynamic evolution law of the welding process through the memory cell state, and outputs a comprehensive evaluation index representing the stability of the welding process.

2. The intelligent welding system for a pole reinforcement framework according to claim 1, wherein The real-time control module contains a multi-objective optimization decision unit and a parameter adaptive adjustment unit; the multi-objective optimization decision unit takes the welding quality, welding efficiency and energy consumption as the optimization objectives, establishes a comprehensive evaluation system containing a weld formation quality score function, a welding cycle function and an energy consumption function, receives the comprehensive evaluation index output by the process state evaluation unit, and solves the process parameter adjustment amount that optimizes the comprehensive evaluation system based on the preset constraints; the parameter adaptive adjustment unit generates specific control instructions according to the process parameter adjustment amount solved by the multi-objective optimization decision unit and sends them to the welding execution module, including amplitude adjustment of the welding current, stability compensation of the arc voltage and linear correction of the welding speed.

3. The intelligent welding system for pole reinforcement framework according to claim 1, characterized in that, The system control center is integrated with a welding knowledge base and a self-learning engine; the welding knowledge base stores historical welding process parameters, corresponding multi-modal sensing data and final welding quality evaluation results, forming a welding process atlas; the self-learning engine periodically accesses the welding knowledge base, incrementally trains the convolutional neural network and long short-term memory network in the edge computing module using newly accumulated welding process data, realizes online updating of model parameters and continuous improvement of identification and evaluation accuracy.

4. The pole reinforcement framework intelligent welding system according to claim 1, characterized in that, The data synchronization mechanism of the multi-modal sensing module is realized through a hardware trigger signal. The high-speed vision sensor, the infrared thermal imager and the arc acoustic emission sensor share a global synchronization clock signal sent by the system control hub, ensuring that each frame of image, each set of temperature data and each piece of sound signal has a strict consistent timestamp, providing a timing alignment basis for the multi-modal data fusion of the edge computing module.

5. The pole reinforcement framework intelligent welding system according to claim 1, characterized in that, The convolutional neural network of the welding defect online identification unit adopts an encoder-decoder structure; The encoder part compresses the multi-modal input data into a high-dimensional feature vector through convolution and pooling operations; The decoder part reconstructs the feature vector into a defect probability distribution map matching the spatial resolution of the original input data through deconvolution and up-sampling operations, thereby realizing pixel-level positioning of the welding defects in the molten pool area.

6. The pole reinforcement framework intelligent welding system according to claim 1, characterized in that, The long short-term memory network of the process state evaluation unit contains a forget gate, an input gate and an output gate mechanism; The forget gate decides which information to discard from the cell state according to the current input and the hidden state at the last time; The input gate controls which new information will be stored in the cell state; The output gate decides the hidden state output at this time based on the current cell state and input information; Through these three gating mechanisms, the network can effectively capture the long-term dependencies of the process state in the welding process.

7. The pole reinforcement framework intelligent welding system according to claim 2, characterized in that, The parameter adaptive adjustment unit adopts a fuzzy proportional-integral-derivative control algorithm for welding current adjustment; This algorithm takes the current adjustment amount output by the multi-objective optimization decision unit as the set value and the actual current value fed back by the welding power supply as the process variable, dynamically adjusts the proportional coefficient, integral time and differential time by calculating the deviation and the rate of change of the deviation, and realizes fast and accurate tracking of the welding current.

8. The pole reinforcement framework intelligent welding system according to claim 1, characterized in that, The system control hub is built with a welding process digital twin model; This model synchronously maps the complete state of the physical welding process in the virtual space by receiving the state data of the welding execution module, the collected data of the multi-modal sensing module and the analysis results of the edge computing module in real time, providing a visual platform for offline simulation, parameter presetting and abnormal condition deduction of the welding process.

9. The pole reinforcement framework intelligent welding system according to claim 2, characterized in that, The multi-objective optimization decision unit adopts the non-dominated sorting genetic algorithm II for optimization calculation; This algorithm takes the welding current, arc voltage and welding speed as decision variables, searches for the optimal parameter combination in the preset feasible region to make the comprehensive evaluation function optimal, sets the population size to 100 and the maximum iteration number to 200.

10. The pole reinforcement framework intelligent welding system according to claim 3, characterized in that, The self-learning engine adopts an early stopping strategy to prevent overfitting; When the validation set loss does not decrease for 10 consecutive periods, the training is terminated, and the training data is divided into a training set, a validation set and a test set in the ratio of 7:2:1.

Images