Parameter self-optimization method for welding robot cooperatively driven by IoTDB and AI

Abstract

The invention provides an IoTDB and AI collaborative driving welding robot parameter self-optimization method, which comprises the following steps: performing data alignment processing on an obtained original data set of a welding robot, generating time-space alignment features, and storing the time-space alignment features in an IoTDB time sequence database; performing cross-modal feature fusion and quality prediction through a pre-trained neural network based on all space-time alignment features in the I oTDB time sequence database, generating a quality prediction value, and writing the quality prediction value into the I oTDB time sequence database; and on the basis of the current space-time alignment features and the corresponding quality predicted values in the IoTDB time sequence database, optimized welding parameters are generated through an incremental reinforcement learning algorithm, and the optimized welding parameters are used for indicating the welding robot to conduct parameter adjustment. By adopting the method, the dynamic response capability and the process stability of intelligent manufacturing can be improved through long-term feature extraction of the time series data and collaborative optimization of an incremental learning mechanism.

Description

Technical Field

[0001] The present invention belongs to the technical field of welding robots, and in particular relates to a welding robot parameter self-optimization method driven collaboratively by IoTDB and AI. Background Art

[0002] With the in-depth application of industrial intelligent technology in the welding field, welding robots are gradually evolving from programmed control to data-driven decision-making. Traditionally, the setting of welding parameters relies heavily on the on-site experience of engineers: manually adjusting core parameters such as current, voltage, and welding speed by observing real-time working conditions such as arc status and weld pool morphology. This approach is manageable under stable working conditions, but in dynamic scenarios such as sudden changes in material thickness and ambient temperature fluctuations, manual response times are often insufficient to match millisecond-level process changes, often leading to frequent defects such as weld porosity and lack of fusion. Furthermore, parameter tuning cycles are long and iteration efficiency is low.

[0003] In recent years, some intelligent welding technologies have attempted to introduce data collection and artificial intelligence analysis, such as assisting parameter optimization through offline training models. However, the welding process generates tens of thousands of high-dimensional multi-source time series data (such as current pulses, voltage waveforms, and weld pool visual flows) per second, and traditional database systems are unable to support efficient storage and real-time feature extraction. At the same time, existing AI models are mostly trained in a static environment, and the optimization parameters cannot be fed back to the execution end in real time. When faced with new materials or new working conditions, massive data training models need to be re-collected, and the adaptive capabilities are seriously insufficient. These bottlenecks make it difficult for existing technologies to achieve closed-loop dynamic optimization of welding quality, which restricts the further improvement of the level of intelligent manufacturing. Summary of the Invention

[0004] Based on this, it is necessary to provide a welding robot parameter self-optimization method driven by IoTDB and AI to address the above technical problems. It can improve the dynamic response capability and process stability of intelligent manufacturing through the collaborative optimization of long-term feature extraction of time series data and incremental learning mechanism.

[0005] In the first aspect, the present application provides a welding robot parameter self-optimization method driven by IoTDB and AI, including:

[0006] Perform data alignment on the acquired welding robot raw data set, generate spatiotemporal alignment features, and store them in the IoTDB time series database;

[0007] Based on all the spatiotemporal alignment features in the IoTDB time series database, a pre-trained neural network is used to perform cross-modal feature fusion and quality prediction, generate quality prediction values, and write the quality prediction values ​​into the IoTDB time series database.

[0008] Based on the current spatiotemporal alignment features and corresponding quality prediction values ​​in the IoTDB time series database, the optimized welding parameters are generated through an incremental reinforcement learning algorithm. The optimized welding parameters are used to instruct the welding robot to adjust the parameters.

[0009] In one embodiment, data alignment processing is performed on the acquired raw data set of the welding robot to generate spatiotemporal alignment features and store them in the IoTDB time series database, including:

[0010] Acquire a raw data set collected by a multi-source sensor, wherein the raw data set includes current data, voltage data, and molten pool visual data;

[0011] The dynamic time warping algorithm is used to align the time axis of current data, voltage data and weld pool visual data to generate a time-synchronized feature set;

[0012] Based on the spatial coordinate transformation matrix of the welding gun head, the time synchronization feature set is mapped into three dimensions to generate spatiotemporal alignment features.

[0013] In one embodiment, based on all spatiotemporal alignment features in the IoTDB time series database, cross-modal feature fusion and quality prediction are performed through a pre-trained neural network to generate a quality prediction value and write the quality prediction value into the IoTDB time series database, including:

[0014] Extract historical spatiotemporal alignment features from the IoTDB time series database and calculate the dynamic time warping distance between the historical spatiotemporal alignment features and the current spatiotemporal alignment features;

[0015] Filter historical spatiotemporal alignment features whose dynamic time warping distance is less than a set threshold, input them into the long short-term memory network for feature extraction, and generate the working condition context feature vector;

[0016] Based on the working condition context feature vector, the current feature, voltage feature and molten pool visual feature in the current spatiotemporal alignment feature are fused with gated channel attention to generate a cross-modal fusion feature vector.

[0017] The cross-modal fusion feature vector is input into the fully connected regression layer for processing to generate the quality prediction value.

[0018] In one embodiment, based on the current spatiotemporal alignment features and corresponding quality prediction values ​​in the IoTDB time series database, an incremental reinforcement learning algorithm is used to generate optimized welding parameters, including:

[0019] Based on the welding process knowledge graph, the historical optimization case parameter set that matches the current spatiotemporal alignment features is retrieved, and the parameter value range of the historical optimization case parameter set is extracted as the initial action space boundary;

[0020] Generate exploration actions based on the initial action space boundary through the actor network, where the exploration actions are adjustment vectors of welding current, voltage and speed;

[0021] Constructing a real-time reward function based on the quality prediction value, where the value of the real-time reward function is positively correlated with the quality prediction value;

[0022] Based on the exploration actions and real-time reward function, the policy gradient of the actor network is calculated through the proximal policy optimization algorithm;

[0023] The parameters of the actor network are elastically updated according to the policy gradient to generate optimized welding parameters.

[0024] In one embodiment, based on the working condition context feature vector, gated channel attention fusion is performed on the current feature, voltage feature, and molten pool visual feature in the current spatiotemporal alignment feature to generate a cross-modal fusion feature vector, including:

[0025] Perform one-dimensional convolution coding on the current characteristics and voltage characteristics to generate electrical signal coding characteristics;

[0026] Perform three-dimensional convolution coding on the visual features of the melt pool to generate visual coding features;

[0027] Calculate the electrical signal channel weight matrix and the visual channel weight matrix through the working condition context feature vector;

[0028] The electrical signal encoding features are weighted according to the electrical signal channel weight matrix, and the visual encoding features are weighted according to the visual channel weight matrix. The weighted results are concatenated to generate a cross-modal fusion feature vector.

[0029] In one embodiment, the cross-modal fusion feature vector is input into a fully connected regression layer for processing to generate a quality prediction value, further comprising:

[0030] The cross-modal fusion feature vector is nonlinearly transformed using the following formula to generate the initial prediction value:

[0031]

[0032] in, is the initial forecast value at time t, f t is the cross-modal fusion feature vector at time t, W1 is the weight matrix of the first fully connected layer, b1 is the bias vector of the first fully connected layer, σ is the ReLU activation function, W2 is the weight matrix of the second fully connected layer, and b2 is the bias vector of the second fully connected layer;

[0033] Based on the initial prediction value, the fully connected regression layer is fine-tuned online using the following loss function:

[0034]

[0035] Among them, y i is the actual welding quality label, N is the training batch size, and λ is the L2 regularization coefficient.

[0036] In one embodiment, a real-time reward function is constructed based on the quality prediction value, including:

[0037] The real-time reward function is constructed using the following formula:

[0038] r t =α·tanh(q t )-β·‖Δp t ‖2

[0039] Among them, r t is the real-time reward value at time t, q t is the quality prediction value at time t, Δp t is the difference vector between the current welding parameters and the previous welding parameters, α and β are preset weighting coefficients, and tanh is the hyperbolic tangent function.

[0040] Secondly, this application also provides a welding robot parameter self-optimization device driven by IoTDB and AI, including:

[0041] The data alignment module is used to align the acquired raw data set of the welding robot, generate spatiotemporal alignment features, and store them in the IoTDB time series database;

[0042] The quality prediction module is used to perform cross-modal feature fusion and quality prediction based on all spatiotemporal alignment features in the IoTDB time series database through a pre-trained neural network, generate quality prediction values, and write the quality prediction values ​​into the IoTDB time series database;

[0043] The parameter optimization module is used to generate optimized welding parameters through an incremental reinforcement learning algorithm based on the current spatiotemporal alignment features and corresponding quality prediction values ​​in the IoTDB time series database. The optimized welding parameters are used to instruct the welding robot to adjust the parameters.

[0044] On the third aspect, the present application also provides a computer device including a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, it implements the above-mentioned IoTDB and AI collaboratively driven welding robot parameter self-optimization method.

[0045] Fourthly, the present application also provides a computer-readable storage medium on which a computer program is stored. When the computer program is executed by a processor, it implements the above-mentioned IoTDB and AI-coordinated self-optimization method for welding robot parameters.

[0046] The above-mentioned IoTDB and AI-driven welding robot parameter self-optimization method generates spatiotemporal alignment features by aligning the welding robot's original data set, and stores them in the IoTDB time series database to achieve real-time and efficient management of high-dimensional data. Then, based on all the spatiotemporal alignment features in the database, a pre-trained neural network is used to perform cross-modal feature fusion and implicit quality prediction, generating quality prediction values ​​reflecting the welding status and writing them back to the database. Combined with the current spatiotemporal alignment features and their corresponding quality prediction values, an incremental reinforcement learning algorithm is used to dynamically generate optimized welding parameters and provide real-time feedback to the welding robot execution end. The above-mentioned technical solution realizes intelligent adaptive adjustment of welding parameters, effectively solving the problems of low data storage efficiency, delayed quality assessment, and slow parameter response in traditional welding, and improving the dynamic adaptability and quality stability of the welding process. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] In order to more clearly illustrate the technical solutions in the embodiments of the present application or related technologies, the following briefly introduces the drawings required for use in the embodiments or related technical descriptions. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0048] Figure 1 A flowchart of a welding robot parameter self-optimization method driven by IoTDB and AI in collaboration, provided by an embodiment of the present invention;

[0049] Figure 2 A schematic flow chart of a welding robot quality prediction method provided by an embodiment of the present invention;

[0050] Figure 3 A schematic diagram of the structure of a welding robot parameter self-optimization device driven by IoTDB and AI in an embodiment of the present invention. DETAILED DESCRIPTION

[0051] In order to make the purpose, technical solutions and advantages of this application more clear, the following further describes this application in detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain this application and are not intended to limit this application.

[0052] First, a brief introduction is given to the terms involved in the embodiments of this application.

[0053] The IoTDB time-series database is a high-performance data management system developed specifically for industrial IoT scenarios. Its core capability lies in the efficient storage, organization, and retrieval of massive amounts of equipment monitoring data that changes dynamically over time. Through a columnar storage engine, time-series data compression algorithms, and a distributed architecture, the database enables millisecond-level writes and real-time aggregate queries of high-frequency, heterogeneous time-series streams (such as temperature, pressure, and current) generated by terminals such as sensors and controllers. This provides a reliable data foundation for status monitoring, anomaly diagnosis, and predictive analysis of industrial equipment. In the field of intelligent manufacturing, IoTDB significantly reduces the processing complexity of multi-source heterogeneous data by uniformly managing time-series information across the entire life cycle of equipment, supporting the critical infrastructure for real-time decision-making optimization.

[0054] A welding robot is a mechatronic industrial equipment that integrates sensing, control, and execution functions. Its core consists of a multi-joint robotic arm, a welding power supply system, and an intelligent controller. Driven by preset programs or real-time decisions, it can autonomously complete welding operations such as arc welding and laser welding on complex spatial trajectories. Unlike traditional manual welding, this equipment uses high-precision posture feedback and dynamic parameter adjustment modules to achieve closed-loop control of key process elements such as molten pool morphology and heat input, thereby improving weld consistency and operational safety. As the core execution unit of the smart factory, the welding robot integrates multidisciplinary technologies such as materials science, automatic control, and machine vision, supporting highly flexible and reliable intelligent welding process systems in fields such as automotive manufacturing and aerospace.

[0055] The incremental reinforcement learning algorithm is a continuous optimization decision-making method for dynamic environments. Its core mechanism is to progressively update the agent's strategy through real-time interactive data streams, rather than relying on batch training with static datasets. Based on the Markov decision process, this algorithm is implemented in the welding robot control scenario as follows: the agent (parameter optimization module) generates actions (current / voltage adjustment instructions) based on the current welding state (such as the molten pool morphology and temperature field distribution), obtains immediate rewards (changes in quality prediction values) from environmental feedback after execution, and uses temporal difference errors to fine-tune the strategy network parameters online. Compared with traditional reinforcement learning, its incremental nature gives it two key capabilities: first, it avoids sudden policy changes through an elastic gradient update mechanism to ensure the stability of the industrial process; second, it uses incremental data streams to continuously adapt to new materials and new working conditions, significantly reducing the cost of model retraining, thereby providing highly robust real-time decision support for complex manufacturing systems.

[0056] According to the above explanation of terms, the implementation environment of a welding robot parameter self-optimization method driven by IoTDB and AI provided in an embodiment of the present application is explained. Schematically, the implementation environment includes: a terminal, a sensor, a processor, and a memory. The terminal is connected to the processor, sensor, and memory signal through a network; the sensor includes but is not limited to a Hall current sensor, a voltage transformer, a high-speed industrial camera, an infrared thermal imager, a laser displacement sensor, a multi-axis inertial measurement unit, and an arc acoustic emission sensor; the processor can be a central processing unit (CPU), a graphics processing unit (GPU), or an artificial intelligence chip (such as an NPU or TPU); the memory can be a distributed cloud storage system or a local server cluster, which is not limited here.

[0057] In combination with the above explanations of terms and implementation environment, the application scenarios of the embodiments of this application are explained. The welding robot parameter self-optimization method provided in the embodiments of this application, which is driven by IoTDB and AI in a collaborative manner, can be applied to, but not limited to, the following scenarios:

[0058] In the welding of large, complex components in fields such as aerospace and energy equipment, material combinations are diverse (such as dissimilar connections between titanium alloys and composite materials), spatial trajectories are complex, and there is a need to address the risk of deformation caused by local heat accumulation. This solution uses IoTDB to store welding current, temperature field distribution, and visual tracking data in real time. The AI ​​model integrates multi-source time series features to predict thermal stress distribution trends and drives reinforcement learning to dynamically adjust welding speed and pulse frequency. This method can improve the quality of dissimilar material fusion and the accuracy of structural deformation control by continuously iterating parameter strategies during continuous operation.

[0059] On flexible automotive body-in-white welding lines, vehicle model changes often lead to frequent changes in joint configurations. IoTDB centrally manages historical process data for different vehicle models and workstations. AI models incrementally learn new joint characteristics and instantly generate adaptive current-speed curves. This system iterates rapidly, eliminating the need for manual retraining to cover new vehicle models, shortening process introduction cycles and ensuring batch consistency.

[0060] Illustratively, the welding robot parameter self-optimization method driven by IoTDB and AI provided in the embodiment of the present application can also be applied to other application scenarios. It is only used as an example here and is not limited to the specific application scenario.

[0061] In an exemplary embodiment, Figure 1As shown, a method for self-optimizing welding robot parameters driven by IoTDB and AI is provided. This embodiment uses the method applied to a terminal in the aforementioned implementation environment as an example. It is understandable that the method can also be applied to a server, and can also be applied to a system including a terminal and a server, and implemented through the interaction between the terminal and the server. In this embodiment, the method includes the following steps 101 to 103:

[0062] Step 101: perform data alignment processing on the acquired original data set of the welding robot, generate spatiotemporal alignment features and store them in the IoTDB time series database.

[0063] Specifically, during operation, the welding robot can synchronously capture raw data such as current pulse waveform, visible light image of the molten pool, infrared temperature field distribution, and wire feed speed through arc sensors, high-speed industrial cameras, infrared temperature measurement arrays, and current and voltage Hall probes. These heterogeneous signals are stamped with a unified timestamp under the triggering of the nanosecond hardware clock. Furthermore, the sampling rates of each channel can be aligned to the same frequency grid through Kalman filtering and cubic spline interpolation, and the spatial coordinates can be shifted with the arc starting moment of the weld as the zero point to obtain a spatiotemporal alignment feature vector containing time series dimensions and spatial dimensions. This vector is written to the weld_raw table of the IoTDB time series database in a columnar storage format, providing a complete and traceable data source for subsequent steps. The timestamp field can be saved as a 64-bit long integer, and the spatial coordinate and physical quantity fields are compressed and encoded using float. Continuous writing of hundreds of thousands of points per second can be achieved on a single node, and millisecond-level range queries are supported.

[0064] Step 102: Based on all the spatiotemporal alignment features in the IoTDB time series database, cross-modal feature fusion and quality prediction are performed through a pre-trained neural network to generate a quality prediction value and write the quality prediction value into the IoTDB time series database.

[0065] Specifically, after accumulating a preset number of spatiotemporal alignment features in IoTDB, the pre-trained multi-stream neural network can be called to perform cross-modal fusion and quality prediction. For example: the first layer of the network can be composed of three parallel spatiotemporal convolution sub-networks, which process the current waveform, the molten pool image sequence and the temperature field tensor respectively. The three outputs are weighted by the attention gating mechanism and input into the cross-modal Transformer encoder. The global feature vector output by the encoder is sent to the quality regression head to generate a dimensionless prediction value characterizing the quality of the weld formation; the prediction value is written back to the weld_pred table of IoTDB together with the timestamp of the corresponding feature to provide instant labels for subsequent incremental reinforcement learning.

[0066] Step 103: Based on the current spatiotemporal alignment features and corresponding quality prediction values ​​in the IoTDB time series database, an incremental reinforcement learning algorithm is used to generate optimized welding parameters. The optimized welding parameters are used to instruct the welding robot to adjust parameters.

[0067] Furthermore, at the end of each welding cycle, the method reads the latest set of spatiotemporal alignment features and their quality prediction values ​​from the weld_pred table, splices them into an environmental state vector, and uses a deep deterministic policy gradient algorithm to search for the optimal increments of current, voltage, and welding speed in the continuous action space; the policy network adopts a double-delay update structure, and the objective function adds a physical constraint penalty term based on the molten pool heat input analytical model in addition to the conventional mean square error to ensure that the output parameters are always within the process feasible domain; the newly generated optimization parameters are sent to the robot control cabinet through the OPC UA protocol and take effect in the next weld segment, thereby converting the AI ​​derivation results into physical actions in real time to complete closed-loop control.

[0068] For example, when a sudden increase in workpiece thickness results in insufficient penetration and a decrease in quality prediction, the above process can detect the anomaly within one weld length cycle and automatically increase the base current and reduce the welding speed, so that the penetration quickly returns to the target range; if the visual channel is lost due to strong light interference, the network can rely on the complementary information of the current and temperature channels to continue to output reliable predictions, avoiding system failure.

[0069] The above-mentioned IoTDB and AI-driven welding robot parameter self-optimization method generates spatiotemporal alignment features by aligning the welding robot's original data set, and stores them in the IoTDB time series database to achieve real-time and efficient management of high-dimensional data. Then, based on all the spatiotemporal alignment features in the database, a pre-trained neural network is used to perform cross-modal feature fusion and implicit quality prediction, generating quality prediction values ​​reflecting the welding status and writing them back to the database. Combined with the current spatiotemporal alignment features and their corresponding quality prediction values, an incremental reinforcement learning algorithm is used to dynamically generate optimized welding parameters and provide real-time feedback to the welding robot execution end. The above-mentioned technical solution realizes intelligent adaptive adjustment of welding parameters, effectively solving the problems of low data storage efficiency, delayed quality assessment, and slow parameter response in traditional welding, and improving the dynamic adaptability and quality stability of the welding process.

[0070] In one embodiment, data alignment processing is performed on the acquired raw data set of the welding robot to generate spatiotemporal alignment features and store them in the IoTDB time series database, including:

[0071] The original data set collected by the multi-source sensor is obtained, where the original data set includes current data, voltage data and molten pool visual data.

[0072] Specifically, the raw data set can be collected in real time by the embedded sensing unit of the welding robot, where the current sensor captures the dynamic impedance changes of the arc in a high-frequency sampling mode, the voltage sensor synchronously records the ripple characteristics of the power supply output, and the molten pool visual data obtains the spatiotemporal evolution sequence of the droplet transition through a high-speed industrial camera. For example, the current and voltage signals use a differential transmission protocol that is resistant to electromagnetic interference, and the visual data is transmitted via optical fiber to avoid motion blur. Furthermore, multi-source data are timestamped at the millisecond level when they converge at the edge computing node to form a raw data queue with a time stamp. This step ensures that the raw data set completely covers the electric-thermal-visual multi-dimensional physical field information of the welding process, providing high-fidelity input for subsequent alignment processing.

[0073] The dynamic time warping algorithm is used to align the time axes of current data, voltage data and weld pool visual data to generate a time-synchronized feature set.

[0074] Specifically, the dynamic time warping algorithm uses the pulse leading edge of the current waveform as the reference sequence, and elastically stretches and compresses the voltage sampling sequence and the visual frame sequence on the time axis through the minimum cumulative distance path search, so that the three achieve strict time synchronization, thereby generating a time synchronization feature set.

[0075] Based on the spatial coordinate transformation matrix of the welding gun head, the time synchronization feature set is mapped into three dimensions to generate spatiotemporal alignment features.

[0076] Specifically, the time synchronization feature set is further mapped to the spatial coordinate transformation matrix of the welding gun head in three dimensions: the laser displacement sensor measures the six-dimensional posture of the welding gun end in the robot base coordinate system in real time, and the computing unit constructs a homogeneous transformation matrix based on the posture, and back-projects each current, voltage value and visual pixel coordinate in the time synchronization feature set to the local coordinate system of the weld, so that all physical quantities carry precise spatial coordinates at the same time, generating spatiotemporal alignment features. This feature is compressed and stored in IoTDB in the form of a triple with timestamp as the primary key, spatial coordinates as the index, and physical quantity as the value range, achieving millisecond-level random reading and writing and efficient compression.

[0077] like Figure 2 As shown, in one embodiment, based on all the spatiotemporal alignment features in the IoTDB time series database, cross-modal feature fusion and quality prediction are performed through a pre-trained neural network to generate a quality prediction value and write the quality prediction value into the IoTDB time series database, including:

[0078] Step 201: extract historical spatiotemporal alignment features from the IoTDB time series database, and calculate the dynamic time warping distance between the historical spatiotemporal alignment features and the current spatiotemporal alignment features.

[0079] Specifically, all historical spatiotemporal alignment features stored in the IoTDB time series database are derived from the current waveform, voltage waveform and molten pool visual frame sequence collected in real time by the welding robot during the previous operation cycle and written after spatiotemporal alignment processing; when it is necessary to call historical information for the current weld segment, a query operation is performed to pull all records overlapping with the time window of the current weld segment from the database, and the current waveform of the current weld segment is used as the reference sequence. Dynamic time warping is performed on the current waveform in each historical record, and the cumulative distance of the warped path is calculated. This distance is defined as the sum of the Euclidean distances of all corresponding sampling points between the two waveforms divided by the sequence length. Only historical records with a distance less than the set threshold are retained, thereby constructing a local sample set that is highly similar to the current working condition; the threshold setting can be obtained through offline statistics, that is, taking the specified quantile of the distance distribution in the historical data. The purpose of this step is to eliminate working condition drift samples, reduce the variance of subsequent network inputs, and improve feature correlation.

[0080] Step 202 : Filter the historical spatiotemporal alignment features whose dynamic time warping distance is less than a set threshold, input them into the long short-term memory network for feature extraction, and generate a working condition context feature vector.

[0081] Exemplarily, when the long short-term memory network processes the filtered local sample set, the input tensor shape is [number of samples, time step, channel], where the channels correspond to the normalized values ​​of current, voltage, and vision respectively, the network hidden state dimension is the same as the number of channels, and the gate control unit updates the hidden state through the forget gate, input gate, and output gate, and finally outputs a fixed-dimensional working condition context feature vector; this vector is essentially a compressed representation of the entire historical weld formation process by the network, which contains comprehensive information on the evolution of the molten pool temperature, arc stability, and penetration depth, providing transferable prior knowledge for the current weld and avoiding prediction bias caused by data sparsity.

[0082] In step 203 , based on the working condition context feature vector, gated channel attention fusion is performed on the current feature, voltage feature, and molten pool visual feature in the current spatiotemporal alignment feature to generate a cross-modal fusion feature vector.

[0083] For example, when the gated channel attention module fuses the working condition context feature vector with the current spatiotemporal alignment feature, the two are spliced ​​in the channel dimension to obtain a joint feature of the shape of [channel × 2]. Subsequently, a global description of each channel is obtained through global average pooling, which is mapped into a channel weight vector through a two-layer fully connected network. The first layer uses ReLU activation and the second layer uses Sigmoid activation. The weight vector is multiplied channel by channel with the current spatiotemporal alignment feature to achieve adaptive weighting of cross-modal information. When the visual channel is polluted by arc light and the signal-to-noise ratio decreases, this weighting mechanism automatically reduces the visual weight and increases the current channel weight, thereby suppressing abnormal modal interference, ensuring that the network focuses on the most reliable information source, and improving the discriminability of the fused features.

[0084] In step 204 , the cross-modal fusion feature vector is input into a fully connected regression layer for processing to generate a quality prediction value.

[0085] Exemplarily, the cross-modal fusion feature vector is input into a single hidden layer fully connected regression network, the number of hidden layer nodes is equal to the fusion feature dimension, the activation function uses ReLU, the number of output layer nodes is 1, there is no activation function, and the network weights are obtained by offline training by minimizing the mean square error loss. The L2 regularization term is added to the loss function to prevent overfitting, and the output value is mapped to a dimensionless weld quality prediction value after denormalization; the prediction value and its timestamp are written back to IoTDB to provide real-time supervision labels for subsequent incremental reinforcement learning, realizing millisecond-level quality estimation and parameter closed-loop optimization, thereby improving the welding process's adaptability to material thickness, gap fluctuations and heat accumulation changes, and ensuring the stability and consistency of weld quality.

[0086] In one embodiment, based on the current spatiotemporal alignment features and corresponding quality prediction values ​​in the IoTDB time series database, an incremental reinforcement learning algorithm is used to generate optimized welding parameters, including:

[0087] Based on the welding process knowledge graph, the historical optimization case parameter set that matches the current spatiotemporal alignment features is retrieved, and the parameter value range of the historical optimization case parameter set is extracted as the initial action space boundary.

[0088] Specifically, the welding process knowledge graph uses welding materials, plate thickness, groove form, and shielding gas type as entity nodes, and the edge relationship records the spatiotemporal alignment features of all successful welds and the final current, voltage, and speed triples adopted. Through the graph embedding model, the current feature is mapped to a low-dimensional vector, and the cosine similarity with the graph node vector is calculated to obtain several historical cases with the highest similarity. The extreme values ​​of their parameters are extracted to form the initial action space boundary, thereby limiting the search scope to the proven feasible process domain under unknown working conditions, avoiding blind exploration leading to welding defects. For example, for the stainless steel sheet lap welding scenario, the historical parameter boundaries of the current range [120A, 150A] and voltage range [22V, 25V] are retrieved and used as the initial exploration area of ​​the reinforcement learning action space, effectively solving the blind search problem in the cold start phase.

[0089] An actor network is used to generate exploration actions based on the initial action space boundary, where the exploration actions are the adjustment vectors of welding current, voltage and speed.

[0090] The actor network is a policy function approximator within the reinforcement learning framework. Its core function is to generate optimization decisions (such as current, voltage, and speed adjustments) within a continuous action space based on environmental conditions (real-time sensor data from the welding robot). This network maps high-dimensional state features to action vectors via a parameterized policy function (such as a Gaussian probability distribution). In welding control scenarios, this network receives spatiotemporally aligned feature vectors and outputs parameter adjustment instructions that meet process safety boundaries. Compared to traditional PID controllers, the actor network adaptively adjusts policy parameters through an online learning mechanism, achieving progressive optimization of welding quality while ensuring arc stability, becoming the intelligent core of the autonomous decision-making system.

[0091] Specifically, the actor network adopts a fully connected structure. Its input is a state vector concatenated from the current spatiotemporal alignment features and the quality prediction value. The output layer has three dimensions, corresponding to the increments of current, voltage, and speed. The network generates exploratory actions within the initial action space boundary using a Gaussian sampling strategy. This applies truncated normal noise to the increments to ensure that the output never crosses the boundary. This action is then sent to the welding power supply and servo drive via the fieldbus, performing a slight adjustment to achieve closed-loop parameter injection at the millisecond level. Furthermore, the actor network can use a Gaussian distribution strategy to generate exploratory actions: using the median of the initial boundary as the reference point, the current adjustment, voltage adjustment, and speed adjustment are randomly sampled according to a normal distribution with a standard deviation of σ = 0.1 to form a three-dimensional continuous action vector. This design ensures that the initial exploration is carried out within the process safety zone, avoiding welding defects caused by parameter out-of-limit.

[0092] A real-time reward function is constructed based on the quality prediction value, where the value of the real-time reward function is positively correlated with the quality prediction value.

[0093] Furthermore, the real-time reward function uses the current quality prediction value as the only independent variable and adopts a piecewise linear form. When the prediction value is higher than the upper limit of the target interval, the reward is a positive constant, and when it is lower than the lower limit, the reward is a negative constant. When it is within the interval, the reward increases linearly with the prediction value, thereby clearly guiding the strategy to converge to higher quality areas during the reinforcement learning process; the reward value is calculated instantly by the edge computing node based on the latest quality prediction value written into IoTDB at the moment the weld is completed.

[0094] Based on the explored actions and the real-time reward function, the policy gradient of the actor network is calculated by a proximal policy optimization algorithm.

[0095] The parameters of the actor network are elastically updated according to the policy gradient to generate optimized welding parameters.

[0096] Specifically, after each weld is completed, the proximal policy optimization algorithm uses the sampled exploration actions and corresponding rewards to construct an advantage function, and limits the policy update amplitude through the importance sampling ratio to prevent violent policy fluctuations; the policy gradient is processed by the Adam optimizer to perform elastic updates on the actor network parameters, where the learning rate decays exponentially with the weld number, taking into account both convergence speed and stability; the updated network parameters are immediately applied to the next weld segment, and new optimized welding parameters are output to achieve incremental closed-loop optimization.

[0097] In one embodiment, based on the working condition context feature vector, gated channel attention fusion is performed on the current feature, voltage feature, and molten pool visual feature in the current spatiotemporal alignment feature to generate a cross-modal fusion feature vector, including:

[0098] Perform one-dimensional convolution coding on the current characteristics and voltage characteristics to generate electrical signal coding characteristics.

[0099] Specifically, both the current characteristics and the voltage characteristics exist in the form of one-dimensional sequences with the weld travel direction as the time axis and the sampling sequence as the length. In order to uniformly represent and compress redundant information, a one-dimensional causal convolution with a length equal to the convolution kernel size is used to encode the two sequences separately; the convolution kernel slides in the time dimension and performs an inner product operation with the local sequence at each position. The output channel dimension is equal to the number of convolution kernels, thereby mapping the original electrical signal into an electrical signal coding feature. This feature significantly compresses channel redundancy while maintaining the same temporal resolution as the original sequence, facilitating subsequent alignment with the visual modality.

[0100] The molten pool visual features are processed by three-dimensional convolution coding to generate visual coding features.

[0101] Specifically, the visual feature of the melt pool is a three-dimensional tensor, whose dimensions correspond to the image height, width and number of channels in turn. A three-dimensional convolution kernel is used to perform space-time joint encoding on the above tensor: the convolution kernel slides simultaneously in the height, width and time dimensions, and a single operation captures the spatial texture and inter-frame changes at the same time, and outputs the visual coding feature. This feature reduces the spatial resolution while retaining the melt pool contour and dynamic evolution information, reduces the amount of calculation, and creates conditions for splicing with the electrical signal coding feature in the channel dimension; the causal constraint of three-dimensional convolution ensures that the encoding process does not introduce future information, meeting the needs of online reasoning.

[0102] The electrical signal channel weight matrix and the visual channel weight matrix are calculated using the working condition context feature vector.

[0103] The electrical signal encoding features are weighted according to the electrical signal channel weight matrix, and the visual encoding features are weighted according to the visual channel weight matrix. The weighted results are concatenated to generate a cross-modal fusion feature vector.

[0104] Specifically, the working condition context feature vector is extracted by the long short-term memory network from the temporal evolution pattern of historically similar welds, and its dimension is equal to the number of weights required for channel attention; the vector is sent to two independent fully connected layers, and the output dimensions of the two groups of layers correspond to the number of channels of electrical signal encoding features and the number of channels of visual encoding features, respectively. The weights of the fully connected layers are mapped into electrical signal channel weight matrices and visual channel weight matrices with values ​​ranging from zero to one through Sigmoid activation; each element of the matrix represents the relative importance of the corresponding channel to the current weld quality, thereby injecting the working condition prior into the attention mechanism in a parameterized form, avoiding the weight drift caused by the lack of process knowledge in traditional attention. For example, the electrical signal channel weight matrix is ​​multiplied channel-by-channel with the electrical signal coding features, and the visual channel weight matrix is ​​multiplied channel-by-channel with the visual coding features to achieve channel-level adaptive weighting; the two weighted sets of features are directly spliced ​​in the channel dimension to obtain a cross-modal fusion feature vector, which fully integrates the visual spatial morphological information while retaining the high-frequency transient information of the electrical signal, and dynamically adjusts the contribution ratio of each modality according to the working condition context; when arc light interference causes the visual signal-to-noise ratio to drop sharply, the visual channel weight automatically approaches zero, and the electrical signal weight increases accordingly, ensuring that the fusion vector is still dominated by reliable modalities, thereby improving the robustness and accuracy of subsequent quality predictions.

[0105] In one embodiment, the cross-modal fusion feature vector is input into a fully connected regression layer for processing to generate a quality prediction value, further comprising:

[0106] The cross-modal fusion feature vector is nonlinearly transformed using the following formula to generate the initial prediction value:

[0107]

[0108] in, is the initial forecast value at time t, f t is the cross-modal fusion feature vector at time t, W1 is the weight matrix of the first fully connected layer, b1 is the bias vector of the first fully connected layer, σ is the ReLU activation function, W2 is the weight matrix of the second fully connected layer, and b2 is the bias vector of the second fully connected layer;

[0109] Based on the initial prediction value, the fully connected regression layer is fine-tuned online using the following loss function:

[0110]

[0111] Among them, y i is the actual welding quality label, N is the training batch size, and λ is the L2 regularization coefficient.

[0112] Specifically, the weight matrices W1 and W2, and the bias vectors b1 and b2 can be obtained through offline end-to-end training before deployment: the cross-modal fusion feature vectors f of several complete welds are fused into t and its corresponding manual metallographic rating y i A supervised data set is formed, and the Adam optimizer is used to iteratively minimize the loss function L until the mean square error on the validation set no longer decreases. At this time, the obtained W1, W2, b1, and b2 are burned into the edge computing node as pre-training parameters; after going online, these parameters are only incrementally corrected through online fine-tuning and are no longer randomly reinitialized, thus ensuring that the model has both prior knowledge and can adapt to new working conditions.

[0113] For example, the actual welding quality label y i Generated by the offline detection link after the weld cools down: After the robot completes a weld, the laser profiler scans along the weld toe to obtain the residual height and weld width, and the X-ray transmission device detects internal pores and lack of fusion. The defect level is converted into a dimensionless score and written into IoTDB. The score is used as y i and the initial prediction value at the corresponding moment Pairing forms the supervised pairs required for online fine-tuning; this step ensures that the labels are interpretable while avoiding the delay of human rating.

[0114] Furthermore, the L2 regularization term in the loss function Used to prevent weights from overfitting on sparse incremental data: λ is selected through cross-validation in the offline stage to minimize the product of the validation error and the weight norm; during online fine-tuning, λ remains fixed, and the gradient only updates the weights along the joint direction of error and regularization, suppressing the severe disturbance of abnormal welds to the model, thereby maintaining prediction stability.

[0115] For example, when arc length random fluctuation occurs in welding of aviation titanium alloy thin plates, the cross-modal fusion feature vector f t The high-frequency current component in the θ is significantly increased. The nonlinear transformation formula maps the high-frequency mode to the high-response area after ReLU activation through W1, and W2 further converts the response into a significantly decreased Online fine-tuning then uses the newly measured y i Small step corrections are made to W1 and W2 to increase the network's sensitivity to arc length fluctuations in subsequent weld segments. The closed-loop control unit reduces the current in advance accordingly, suppresses undercut defects, and achieves continuous progress in quality prediction and parameter optimization.

[0116] In one embodiment, a real-time reward function is constructed based on the quality prediction value, including:

[0117] The real-time reward function is constructed using the following formula:

[0118] r t =α·tanh(q t )-β·||Δp t ||2

[0119] Among them, r t is the real-time reward value at time t, q t is the quality prediction value at time t, Δp t is the difference vector between the current welding parameters and the previous welding parameters, α and β are preset weighting coefficients, and tanh is the hyperbolic tangent function.

[0120] Specifically, α and β are determined by process experts and data-driven optimization in the offline stage; the quality prediction value q t The output of the aforementioned fully connected regression layer in each sampling period is a dimensionless fraction; Δp t The three-dimensional vector composed of the current welding current, voltage, and speed is subtracted element by element from the corresponding vector of the previous sampling period. The difference vector represents the parameter adjustment amplitude; tanh converts q t Mapping to the (-1,1) interval compresses the excessive amplification of rewards by extreme prediction values ​​while preserving monotonicity and ensuring that the reward sign correctly reflects the quality. t The calculation can be performed in the edge computing node with the welding cycle as the beat, and the calculation results are immediately sent to the experience replay pool of the proximal strategy optimization algorithm; the algorithm is based on r t As the core input of advantage function estimation, it guides the actor network to generate better parameter increments in the subsequent weld segments, achieving closed-loop guidance of "high quality, high reward, large disturbance, large penalty"; when the weld quality decreases due to material mutation, q t The reduction leads to tanh(q t ) tends to be negative, r tThe strategy gradient direction changes rapidly, prompting the network to reduce the current or speed and suppress the expansion of defects. For example, in the high-speed welding scenario of automobile thin plates, if the arc length fluctuation causes q t Instantaneous drop, Δp t Keep it small, then r t As the negative value increases, the strategy immediately reduces the current increment; if q t Improve but Δp t Due to mechanical vibration, it increases significantly, then β·||Δp t || 2 items offset the positive reward to prevent over-regulation from causing burn-through; thus, r t The integration of quality and stability indicators enables reinforcement learning to quickly optimize while maintaining process stability under complex working conditions.

[0121] In summary, the present application provides a welding robot parameter self-optimization method driven by the collaboration of IoTDB and AI. By dynamically time-warping and spatial coordinate transformation of the multi-source sensor data of the welding robot, spatiotemporal alignment features with physical location labels are generated and stored in the IoTDB time series database, thereby solving the data island problem in traditional welding. Then, based on the similarity matching of historical and real-time features, the process context rules are extracted through the long short-term memory network, and the gated channel attention mechanism is driven to dynamically fuse the current, voltage and molten pool visual features to construct a cross-modal fusion representation. The fused features are input into a two-layer fully connected network to generate real-time quality prediction values, and combined with regularization constraints to achieve online fine-tuning, thereby overcoming the bottleneck that the fixed model cannot adapt to new working conditions. The reward function is constructed with quality prediction and parameter adjustment amplitude as the dual objectives, and the optimization parameters are generated within the process safety boundary through incremental reinforcement learning to achieve autonomous decision-making optimization of the welding robot. The above technical solution can enable welding robots to have the ability to self-perceive process status, self-diagnose quality risks, and self-evolve parameter strategies. Based on the collaborative optimization of long-term feature extraction of time series data and incremental learning mechanism, it can enhance the dynamic response capability and process stability of intelligent manufacturing, and improve quality consistency and production continuity under complex dynamic working conditions.

[0122] It should be understood that, although the various steps in the flowcharts involved in the various embodiments described above are displayed in sequence according to the instructions of the arrows, these steps are not necessarily executed in sequence in the order indicated by the arrows. Unless otherwise specified herein, there is no strict order restriction on the execution of these steps, and these steps can be executed in other orders. Moreover, at least a portion of the steps in the flowcharts involved in the various embodiments described above can include multiple steps or multiple stages, and these steps or stages are not necessarily executed and completed at the same time, but can be executed at different times, and the execution order of these steps or stages is not necessarily to be carried out in sequence, but can be executed in turn or alternately with other steps or at least a portion of steps or stages in other steps.

[0123] Based on the same inventive concept, the embodiment of the present application also provides an IoTDB and AI-coordinated self-optimization device 10 for implementing the aforementioned method for self-optimization of welding robot parameters driven by the IoTDB and AI. The implementation solution provided by the device is similar to the implementation solution described in the aforementioned method. Therefore, the specific limitations of the embodiments of one or more IoTDB and AI-coordinated self-optimization devices 10 for welding robot parameters provided below can be found in the above-mentioned limitations of the method for self-optimization of welding robot parameters driven by the IoTDB and AI, and will not be repeated here.

[0124] In an exemplary embodiment, Figure 3 As shown, a welding robot parameter self-optimization device 10 driven by IoTDB and AI is provided, comprising:

[0125] The data alignment module 11 is used to perform data alignment processing on the acquired raw data set of the welding robot, generate spatiotemporal alignment features and store them in the IoTDB time series database;

[0126] The quality prediction module 12 is used to perform cross-modal feature fusion and quality prediction based on all spatiotemporal alignment features in the IoTDB time series database through a pre-trained neural network, generate a quality prediction value, and write the quality prediction value into the IoTDB time series database;

[0127] The parameter optimization module 13 is used to generate optimized welding parameters through an incremental reinforcement learning algorithm based on the current spatiotemporal alignment features and corresponding quality prediction values ​​in the IoTDB time series database. The optimized welding parameters are used to instruct the welding robot to adjust the parameters.

[0128] In one embodiment, the data alignment module 11 includes:

[0129] A sensor data acquisition unit is used to acquire a raw data set collected by a multi-source sensor, wherein the raw data set includes current data, voltage data and molten pool visual data;

[0130] The time warping unit is used to align the time axis of current data, voltage data and weld pool visual data through a dynamic time warping algorithm to generate a time-synchronized feature set;

[0131] The spatial mapping unit is used to perform three-dimensional spatial mapping processing on the time synchronization feature set based on the spatial coordinate transformation matrix of the welding gun head to generate spatiotemporal alignment features.

[0132] In one embodiment, the quality prediction module 12 includes:

[0133] The feature retrieval unit is used to extract historical spatiotemporal alignment features from the IoTDB time series database and calculate the dynamic time warping distance between the historical spatiotemporal alignment features and the current spatiotemporal alignment features;

[0134] Filter historical spatiotemporal alignment features whose dynamic time warping distance is less than a set threshold, input them into the long short-term memory network for feature extraction, and generate the working condition context feature vector;

[0135] The context fusion unit is used to perform gated channel attention fusion on the current features, voltage features, and molten pool visual features in the current spatiotemporal alignment features based on the working condition context feature vector to generate a cross-modal fusion feature vector;

[0136] The regression prediction unit is used to input the cross-modal fusion feature vector into the fully connected regression layer for processing to generate a quality prediction value.

[0137] In one embodiment, the parameter optimization module 13 includes:

[0138] A knowledge retrieval unit is used to retrieve a historical optimization case parameter set that matches the current spatiotemporal alignment feature based on the welding process knowledge graph, and extract the parameter value range of the historical optimization case parameter set as the initial action space boundary;

[0139] An action generation unit is used to generate exploration actions based on the initial action space boundary through an actor network, where the exploration actions are adjustment vectors of welding current, voltage and speed;

[0140] a reward construction unit, configured to construct a real-time reward function based on the quality prediction value, wherein a value of the real-time reward function is positively correlated with the quality prediction value;

[0141] Gradient calculation unit, used to calculate the policy gradient of the actor network through the proximal policy optimization algorithm based on the exploration action and the real-time reward function;

[0142] The parameter updating unit is used to elastically update the parameters of the actor network according to the policy gradient to generate optimized welding parameters.

[0143] In one embodiment, the context fusion unit further includes:

[0144] An electrical signal encoding unit, configured to perform one-dimensional convolution encoding processing on the current characteristics and the voltage characteristics to generate electrical signal encoding characteristics;

[0145] A visual coding unit is used to perform three-dimensional convolution coding processing on the visual features of the melt pool to generate visual coding features;

[0146] A weight calculation unit, used to calculate the electrical signal channel weight matrix and the visual channel weight matrix based on the working condition context feature vector;

[0147] The feature weighting unit is used to weight the electrical signal encoding features according to the electrical signal channel weight matrix, and at the same time weight the visual encoding features according to the visual channel weight matrix, and concatenate the weighted results to generate a cross-modal fusion feature vector.

[0148] In one embodiment, the regression prediction unit further includes:

[0149] The nonlinear transformation unit is used to perform nonlinear transformation on the cross-modal fusion feature vector to generate an initial prediction value using the following formula:

[0150]

[0151] in, is the initial forecast value at time t, f t is the cross-modal fusion feature vector at time t, W1 is the weight matrix of the first fully connected layer, b1 is the bias vector of the first fully connected layer, σ is the ReLU activation function, W2 is the weight matrix of the second fully connected layer, and b2 is the bias vector of the second fully connected layer;

[0152] The online fine-tuning unit is used to fine-tune the fully connected regression layer online based on the initial prediction value using the following loss function:

[0153]

[0154] Among them, y i is the actual welding quality label, N is the training batch size, and λ is the L2 regularization coefficient.

[0155] In one embodiment, the reward building unit includes:

[0156] Function generation unit, used to construct a real-time reward function using the following formula:

[0157] r t =α·tanh(q t )-β·||Δp t ||2

[0158] Among them, r t is the real-time reward value at time t, q t is the quality prediction value at time t, Δp t is the difference vector between the current welding parameters and the previous welding parameters, α and β are preset weighting coefficients, and tanh is the hyperbolic tangent function.

[0159] In one embodiment, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, the steps of the welding robot parameter self-optimization method driven by IoTDB and AI as described above are implemented.

[0160] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the steps in the above-mentioned method embodiments are implemented.

[0161] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to the partial description of the method embodiments. The device embodiments described above are merely illustrative, wherein the components described as separate parts may or may not be physically separated, and the parts displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the disclosed solution. A person of ordinary skill in the art can understand and implement it without expending creative work.

[0162] The above-described embodiments merely represent several implementation methods of the embodiments of the present application. While the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the patent application. It should be noted that a person skilled in the art may make various modifications and improvements without departing from the concept of the embodiments of the present application, and these modifications and improvements fall within the scope of protection of the embodiments of the present application.

Claims

1. A welding robot parameter self-optimization method driven by IoTDB and AI, characterized in that: The method comprises: Perform data alignment on the acquired welding robot raw data set, generate spatiotemporal alignment features, and store them in the IoTDB time series database; Based on all the spatiotemporal alignment features in the IoTDB time series database, cross-modal feature fusion and quality prediction are performed through a pre-trained neural network to generate a quality prediction value and write the quality prediction value into the IoTDB time series database; Based on the current spatiotemporal alignment features and the corresponding quality prediction values ​​in the IoTDB time series database, optimized welding parameters are generated through an incremental reinforcement learning algorithm, and the optimized welding parameters are used to instruct the welding robot to adjust parameters.

2. The method according to claim 1, characterized in that The data alignment process is performed on the acquired original data set of the welding robot to generate spatiotemporal alignment features and store them in the IoTDB time series database, including: Acquire the raw data set collected by the multi-source sensor, wherein the raw data set includes current data, voltage data, and molten pool visual data; Performing time axis alignment on the current data, the voltage data, and the molten pool visual data using a dynamic time warping algorithm to generate a time synchronization feature set; Based on the spatial coordinate transformation matrix of the welding gun head, the time synchronization feature set is subjected to three-dimensional spatial mapping processing to generate the time-space alignment feature.

3. The method according to claim 1, characterized in that The method of performing cross-modal feature fusion and quality prediction based on all the spatiotemporal alignment features in the IoTDB time series database through a pre-trained neural network, generating a quality prediction value, and writing the quality prediction value into the IoTDB time series database includes: Extract historical spatiotemporal alignment features from the IoTDB time series database, and calculate the dynamic time warping distance between the historical spatiotemporal alignment features and the current spatiotemporal alignment features; Screening the historical spatiotemporal alignment features whose dynamic time warping distance is less than a set threshold, and inputting them into a long short-term memory network for feature extraction to generate a working condition context feature vector; Based on the working condition context feature vector, gated channel attention fusion is performed on the current feature, voltage feature and molten pool visual feature in the current spatiotemporal alignment feature to generate a cross-modal fusion feature vector; The cross-modal fusion feature vector is input into a fully connected regression layer for processing to generate the quality prediction value.

4. The method according to claim 1, wherein The method of generating optimized welding parameters by using an incremental reinforcement learning algorithm based on the current spatiotemporal alignment features and the corresponding quality prediction values ​​in the IoTDB time series database includes: Retrieving a historical optimization case parameter set that matches the current spatiotemporal alignment feature based on the welding process knowledge graph, and extracting a parameter value range of the historical optimization case parameter set as an initial action space boundary; generating an exploration action based on the initial action space boundary by an actor network, wherein the exploration action is an adjustment vector of welding current, voltage and speed; constructing a real-time reward function based on the quality prediction value, wherein a value of the real-time reward function is positively correlated with the quality prediction value; Based on the exploration action and the real-time reward function, calculating the policy gradient of the actor network through a proximal policy optimization algorithm; Parameters of the actor network are elastically updated according to the policy gradient to generate the optimized welding parameters.

5. The method according to claim 3, characterized in that The gated channel attention fusion is performed on the current feature, the voltage feature, and the molten pool visual feature in the current spatiotemporal alignment feature based on the working condition context feature vector to generate a cross-modal fusion feature vector, including: performing one-dimensional convolution coding processing on the current feature and the voltage feature to generate an electrical signal coding feature; Performing three-dimensional convolution coding processing on the molten pool visual features to generate visual coding features; Calculating an electrical signal channel weight matrix and a visual channel weight matrix using the working condition context feature vector; The electrical signal coding features are weighted according to the electrical signal channel weight matrix, and the visual coding features are weighted according to the visual channel weight matrix, and the weighted results are concatenated to generate the cross-modal fusion feature vector.

6. The method according to claim 3, characterized in that The step of inputting the cross-modal fusion feature vector into a fully connected regression layer for processing to generate the quality prediction value further includes: The cross-modal fusion feature vector is nonlinearly transformed using the following formula to generate an initial prediction value: in, is the initial prediction value at time t, f t is the cross-modal fusion feature vector at time t, W1 is the weight matrix of the first fully connected layer, b1 is the bias vector of the first fully connected layer, σ is the ReLU activation function, W2 is the weight matrix of the second fully connected layer, and b2 is the bias vector of the second fully connected layer; Based on the initial prediction value, the fully connected regression layer is fine-tuned online using the following loss function: Among them, y i is the actual welding quality label, N is the training batch size, and λ is the L2 regularization coefficient.

7. The method according to claim 4, characterized in that The constructing of a real-time reward function based on the quality prediction value includes: The real-time reward function is constructed by the following formula: r t =α·tanh(q t )-β·||Δp t ||2 Among them, r t is the real-time reward value at time t, q t is the quality prediction value at time t, ΔP t is the difference vector between the current welding parameters and the previous welding parameters, α and β are preset weighting coefficients, and tanh is the hyperbolic tangent function.

8. A welding robot parameter self-optimization device driven by IoTDB and AI, characterized in that: The device comprises: The data alignment module is used to align the acquired raw data set of the welding robot, generate spatiotemporal alignment features, and store them in the IoTDB time series database; A quality prediction module, configured to perform cross-modal feature fusion and quality prediction based on all the spatiotemporal alignment features in the IoTDB time series database through a pre-trained neural network, generate a quality prediction value, and write the quality prediction value into the IoTDB time series database; A parameter optimization module is used to generate optimized welding parameters through an incremental reinforcement learning algorithm based on the current spatiotemporal alignment features and the corresponding quality prediction values ​​in the IoTDB time series database. The optimized welding parameters are used to instruct the welding robot to adjust parameters.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the method according to any one of claims 1 to 7 is implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the method according to any one of claims 1 to 7 is implemented.

Images