Titanium-steel clad plate welding strength prediction and parameter control method based on neural network

Abstract

The application provides a neural network-based titanium steel composite plate welding strength prediction and parameter control method, relates to the technical field of artificial intelligence, and comprises the following steps: acquiring welding process parameters, a weld surface image and ultrasonic detection data, constructing a double-branch heterogeneous neural network, integrating a lightweight RNN and an improved LSTM structure of an attention mechanism, and realizing rapid trend analysis and accurate prediction of welding strength. Based on a prediction error and a reinforcement learning framework, the coupling degree between process parameters is calculated, adaptive control weights are generated, real-time collaborative adjustment of welding current, voltage and speed is realized, and the accurate control ability of welding quality is improved.

Description

Technical Field

[0001] This invention relates to artificial intelligence technology, and more particularly to a method for predicting the welding strength and controlling parameters of titanium-steel composite plates based on neural networks. Background Technology

[0002] Traditional welding process control for titanium-steel composite plates relies primarily on empirical parameters and offline quality inspection, resulting in issues such as low prediction accuracy, response lag, and inconsistent parameter optimization. Existing technologies, based on physical models, struggle to accurately describe the multi-physics coupling phenomena in complex welding processes, are computationally complex, and suffer from poor real-time performance. While prediction methods based on single neural networks can capture the nonlinear relationship between welding parameters and strength to some extent, they face a trade-off between accuracy and speed, and lack sufficient consideration for temporal characteristics. Regarding parameter control, traditional PID control and fuzzy control cannot effectively handle the complex coupling relationships between welding parameters, leading to mutual interference in parameter adjustments and significant fluctuations in welding quality. Some welding control methods employing reinforcement learning primarily focus on optimizing single parameters, neglecting the synergistic effects between parameters and lacking stability constraint mechanisms, easily leading to over- or under-adjustment of parameters. Furthermore, existing technologies generally lack the ability to integrate and analyze welding images, process parameters, and inspection data, making it difficult to comprehensively assess the welding state. This restricts the improvement of the stability of titanium-steel composite plate welding quality and fails to meet the stringent performance requirements of high-end equipment manufacturing for welded joints. Summary of the Invention

[0003] To address the shortcomings of existing technologies, this invention provides a neural network-based method for predicting the welding strength and controlling parameters of titanium-steel composite plates, which can solve the problems in existing technologies.

[0004] A first aspect of the present invention provides a method for predicting the welding strength and controlling parameters of titanium-steel composite plates based on neural networks, comprising:

[0005] Welding process parameters, weld surface images, and ultrasonic testing data of the weld are obtained during the welding process of titanium-steel composite plates. The welding process parameters include welding current, welding voltage, and welding speed. A standardized dataset is obtained through preprocessing.

[0006] A dual-branch heterogeneous neural network prediction mechanism is constructed. The first branch adopts a lightweight RNN structure with an attention mechanism to quickly extract features from the standardized dataset and output the welding strength change trend. The second branch adopts an improved LSTM structure to accurately model the standardized dataset and output welding strength prediction data. The welding strength change trend and the welding strength prediction data are adaptively fused to obtain a comprehensive prediction result.

[0007] Calculate the prediction error between the comprehensive prediction result and the actual welding strength data acquired in real time;

[0008] A collaborative optimization control scheme is constructed based on reinforcement learning. The welding process parameters, weld surface image, and prediction error are used as state inputs. The parameter coupling degree between welding current, welding voltage, and welding speed is calculated. The control weights of welding current, welding voltage, and welding speed are generated based on the parameter coupling degree, and the optimal adjustment strategy data is generated. Based on the optimal adjustment strategy data, the welding current, welding voltage, and welding speed are adjusted in real time.

[0009] Optional,

[0010] The first branch employs a lightweight RNN structure with an attention mechanism to quickly extract features from the standardized dataset, and the steps for outputting the welding strength change trend include:

[0011] A multi-scale temporal feature extraction module is constructed. The multi-scale temporal feature extraction module extracts features from the standardized dataset using sliding windows of different sizes. It calculates temporal attention weights based on the hidden state at the current time step and the state at the previous time step, and performs adaptive weighted fusion of the features extracted at different scales to obtain multi-scale fused features.

[0012] The multi-scale fusion features are processed by a bidirectional gated loop unit. The bidirectional gated loop unit calculates an adaptive gate threshold based on the mean and standard deviation of the input data and dynamically modulates the outputs of the update gate and the reset gate to obtain the optimized gate state.

[0013] Construct a forward propagation path and a backward propagation path. The forward propagation path generates a forward candidate state based on the modulation result of the optimized reset gate state and the hidden layer state at the previous time step and the multi-scale fusion feature. The backward propagation path generates a backward candidate state based on the modulation result of the optimized reset gate state and the hidden layer state at the next time step and the multi-scale fusion feature.

[0014] The forward candidate state, the reverse candidate state, and the hidden layer state of the previous time step are residually connected, and the residual connection result is selectively updated through the optimized update gate state to obtain the hidden layer state at the current time step; the hidden layer state at the current time step is subjected to feature extraction and fusion to generate welding strength change trend data.

[0015] Optional,

[0016] The second branch uses an improved LSTM structure to accurately model the standardized dataset and outputs welding strength prediction data. The steps include:

[0017] The standardized dataset is decomposed into features, and the importance weights of the decomposed features are calculated based on the multilayer perceptron. The importance weights are then combined with the corresponding decomposed features to obtain weighted fusion features.

[0018] A dynamic head-adaptive multi-head attention module is constructed. The multi-head attention module adaptively determines the number of attention heads based on the current state of the weighted fusion feature and maps the weighted fusion feature into a query matrix, a key matrix, and a value matrix. A position-aware matrix is ​​generated based on sine and cosine position encoding. The attention weights are obtained by combining the product of the position-aware matrix with the query matrix and the key matrix.

[0019] A channel attention enhancement module is constructed. The channel attention enhancement module performs global average pooling on the attention weights to obtain channel features, calculates the weight coefficients of the channel features based on a multilayer perceptron, and performs channel-by-channel product operation on the weight coefficients and the attention weights to obtain enhanced features.

[0020] An improved Long Short-Term Memory (LSTM) unit is constructed, and the enhanced features are used as input. The forget gate of the LSM unit is adaptively adjusted based on the state of the memory unit at the previous time step, and the input gate is adaptively adjusted based on the hyperbolic tangent transform of the state of the memory unit at the previous time step to obtain the hidden layer state at the current time step.

[0021] Dynamic coefficients are calculated based on the feature similarity of the hidden state of adjacent layers. The dynamic coefficients are used to perform residual connections on the hidden state of adjacent layers. Dense connections are then performed on the hidden state of multiple layers to obtain interlayer fusion features. Welding strength prediction results are generated based on the interlayer fusion features.

[0022] Optional,

[0023] The steps for adaptively fusing the welding strength variation trend and the welding strength prediction data to obtain a comprehensive prediction result include:

[0024] A time-series consistency score is calculated based on the temporal gradient characteristics of the welding strength change trend and the fluctuation characteristics of the welding strength prediction data. The time-series consistency score is then combined with the real-time status data of the welding process parameters to generate a dynamic fusion weight.

[0025] The dynamic fusion weights are adaptively modulated based on historical prediction accuracy to generate optimized fusion weight coefficients. The optimized fusion weight coefficients are then weighted and combined with the welding strength change trend and the welding strength prediction data to obtain an initial fusion result. The residual compensation value is calculated based on the deviation between the initial fusion result and the historical prediction data. The residual compensation value is then adaptively corrected with the initial fusion result to obtain a comprehensive prediction result.

[0026] Optional,

[0027] The steps for constructing a collaborative optimization control scheme based on reinforcement learning, using the welding process parameters, weld surface image, and prediction error as state inputs, calculating the parameter coupling degree between welding current, welding voltage, and welding speed, generating control weights for welding current, welding voltage, and welding speed based on the parameter coupling degree, and generating optimal adjustment strategy data include:

[0028] Welding process parameters, weld surface images, and prediction errors are combined to obtain fusion state characteristics; based on the fusion state characteristics, the Pearson correlation coefficients between welding current, welding voltage, and welding speed are calculated to generate a parameter coupling matrix;

[0029] The initial weight of each welding process parameter is calculated based on the coupling strength between the parameters in the parameter coupling matrix. The initial weight is then combined with the weight adjustment gradient of the reward function calculated by weighting the welding quality index and the process economic index to obtain the optimized dynamic control weight.

[0030] A deep Q-network structure is constructed, and the fused state features and the optimized dynamic control weights are input into the deep Q-network structure to generate a state-action value estimate; an ε-greedy strategy is used to select the adjustment action based on the state-action value estimate.

[0031] The temporal difference error is calculated based on the immediate reward and the maximum value estimate of the next state, and the parameters of the deep Q network structure are optimized and updated using the temporal difference error.

[0032] The optimized dynamic control weights are weighted and modulated with the adjustment actions selected based on the deep Q-network structure to generate welding current adjustment, welding voltage adjustment, and welding speed adjustment, thereby obtaining the optimal adjustment strategy.

[0033] Optional,

[0034] The steps for calculating the initial weight of each welding process parameter based on the coupling strength between parameters in the parameter coupling matrix, and combining the initial weight with the weight adjustment gradient of the reward function calculated based on welding quality indicators and process economic indicators to obtain the optimized dynamic control weights include:

[0035] The basic coupling degree is obtained by calculating the Pearson correlation coefficient between each pair of welding current, welding voltage, and welding speed; wavelet transform is performed on the welding current, welding voltage, and welding speed, and the time-frequency coupling characteristics under different frequency bands are calculated based on the cross wavelet transform coefficients; the causal relationship strength between the welding current, welding voltage, and welding speed is calculated based on the Granger causality test method; the basic coupling degree, the time-frequency coupling characteristics, and the causal relationship strength are weighted and fused to generate a fused coupling degree matrix.

[0036] The initial weights of welding current, welding voltage, and welding speed are calculated based on the fusion coupling degree matrix; a weld formation consistency score is calculated based on weld contour characteristics; a mechanical property score is calculated based on tensile strength, yield strength, and elongation; and a process stability score is calculated based on arc characteristic fluctuations and molten pool dynamic behavior. The weld formation consistency score, mechanical property score, and process stability score are weighted and combined to obtain a welding quality evaluation value; a process economic evaluation value is calculated based on energy consumption indicators and material utilization rate indicators; and a comprehensive reward value is obtained by weighting the welding quality evaluation value and the process economic evaluation value.

[0037] Calculate the gradient of the comprehensive reward value with respect to the initial weight, combine the gradient with the initial weight, and introduce the momentum term of the difference between two adjacent weights to generate dynamic weights; calculate the compensation adjustment amount based on the historical adjustment effect; combine the compensation adjustment amount with the dynamic weights to obtain the optimized dynamic control weights.

[0038] Optional,

[0039] The steps of weighted modulation of the optimized dynamic control weights with the adjustment actions selected based on the deep Q-network structure to generate welding current adjustment, welding voltage adjustment, and welding speed adjustment, and to obtain the optimal adjustment strategy, include:

[0040] The optimized dynamic control weights are standardized and multiplied with the adjustment actions output by the deep Q network to obtain the basic current adjustment, basic voltage adjustment, and basic speed adjustment. A compensation value is calculated based on the deviation between the target parameter value and the current parameter value. The compensation value is combined with the corresponding basic adjustment to obtain the total current adjustment, total voltage adjustment, and total speed adjustment.

[0041] Calculate the stability constraint values ​​for the total current regulation, total voltage regulation, and total speed regulation. The stability constraint value is the ratio of the absolute value of the total regulation to the product of the current parameter value and the stability constraint coefficient. Based on the stability constraint value, perform a first adjustment on the total current regulation, total voltage regulation, and total speed regulation according to a preset ratio. Calculate the pairwise coupling constraint values ​​between each of the total current regulation, total voltage regulation, and total speed regulation after the first adjustment. Iteratively optimize the total current regulation, total voltage regulation, and total speed regulation based on the coupling constraint values ​​until the coupling constraint values ​​meet a preset coupling threshold. Output the optimized total current regulation, total voltage regulation, and total speed regulation as the optimal regulation strategy.

[0042] In a second aspect, a computer-readable storage medium is provided, having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0043] This invention achieves high-precision prediction of welding strength through a dual-branch heterogeneous neural network prediction mechanism, balancing prediction speed and accuracy. The first branch combines a lightweight RNN structure with an attention mechanism to quickly capture the trend of welding strength changes; the second branch uses an improved LSTM structure to accurately model the welding process and output high-precision prediction data. An adaptive fusion mechanism intelligently adjusts the weights of the two branches according to different working conditions, combining their advantages to improve overall prediction performance. A reinforcement learning-based collaborative optimization control scheme calculates parameter coupling degree to deeply understand the complex interaction between welding current, voltage, and speed, generating dynamic control weights to achieve collaborative optimization adjustment of welding parameters. A multi-constraint optimization mechanism ensures a smooth and reliable parameter adjustment process, avoiding parameter conflicts. The overall scheme constructs a closed-loop self-optimizing system, significantly improving the welding quality and consistency of titanium-steel composite plates, reducing energy consumption and material waste, providing effective technical support for intelligent welding of high-performance composite materials, and promoting the development of welding manufacturing technology towards intelligence and precision. Attached Figure Description

[0044] Figure 1 This is a flowchart illustrating the method for predicting welding strength and controlling parameters of titanium-steel composite plates based on neural networks, as described in an embodiment of the present invention.

[0045] Figure 2 This is a flowchart of the collaborative optimization control of welding parameters based on reinforcement learning, according to an embodiment of the present invention. Detailed Implementation

[0046] The technical solutions of the present invention will be described below with reference to the accompanying drawings. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0047] Figure 1 This is a flowchart illustrating the method for predicting welding strength and controlling parameters of titanium-steel composite plates based on neural networks, as described in this invention. Figure 1 As shown, the method includes:

[0048] Welding process parameters, weld surface images, and ultrasonic testing data of the weld are obtained during the welding process of titanium-steel composite plates. The welding process parameters include welding current, welding voltage, and welding speed. A standardized dataset is obtained through preprocessing.

[0049] A dual-branch heterogeneous neural network prediction mechanism is constructed. The first branch adopts a lightweight RNN structure with an attention mechanism to quickly extract features from the standardized dataset and output the welding strength change trend. The second branch adopts an improved LSTM structure to accurately model the standardized dataset and output welding strength prediction data. The welding strength change trend and the welding strength prediction data are adaptively fused to obtain a comprehensive prediction result.

[0050] Calculate the prediction error between the comprehensive prediction result and the actual welding strength data acquired in real time;

[0051] A collaborative optimization control scheme is constructed based on reinforcement learning. The welding process parameters, weld surface image, and prediction error are used as state inputs. The parameter coupling degree between welding current, welding voltage, and welding speed is calculated. The control weights of welding current, welding voltage, and welding speed are generated based on the parameter coupling degree, and the optimal adjustment strategy data is generated. Based on the optimal adjustment strategy data, the welding current, welding voltage, and welding speed are adjusted in real time.

[0052] For example, a data acquisition module is set up to acquire process parameters, weld surface images, and ultrasonic testing data during the welding process. Process parameters include welding current, welding voltage, and welding speed, with a sampling frequency of 100Hz; weld surface images are acquired in real-time using an industrial camera with a resolution of 1920×1080 pixels and a frame rate of 30fps; ultrasonic testing data is acquired using a probe array at a frequency of 5MHz. The acquired data is preprocessed: process parameters are normalized to the 0-1 range, image data is converted to grayscale and standardized in size, and ultrasonic data undergoes noise reduction and feature extraction to construct a standardized dataset.

[0053] The prediction error is calculated using a combination of relative error and root mean square error. Actual welding strength data is acquired online, and tensile strength values ​​are measured through tensile tests and compared with the prediction results. When the relative error exceeds 5% or the root mean square error exceeds 30 MPa, the online fine-tuning mechanism of the prediction model is triggered.

[0054] Optional,

[0055] The first branch employs a lightweight RNN structure with an attention mechanism to quickly extract features from the standardized dataset, and the steps for outputting the welding strength change trend include:

[0056] A multi-scale temporal feature extraction module is constructed. The multi-scale temporal feature extraction module extracts features from the standardized dataset using sliding windows of different sizes. It calculates temporal attention weights based on the hidden layer state at the current time step and the state at the previous time step, and performs adaptive weighted fusion of the features extracted at different scales to obtain multi-scale fused features. The standardized dataset includes welding process parameters, weld surface images, and weld ultrasonic detection data.

[0057] The multi-scale fusion features are processed by a bidirectional gated loop unit. The bidirectional gated loop unit calculates an adaptive gate threshold based on the mean and standard deviation of the input data and dynamically modulates the outputs of the update gate and the reset gate to obtain the optimized gate state.

[0058] Construct a forward propagation path and a backward propagation path. The forward propagation path generates a forward candidate state based on the modulation result of the optimized reset gate state and the hidden layer state at the previous time step and the multi-scale fusion feature. The backward propagation path generates a backward candidate state based on the modulation result of the optimized reset gate state and the hidden layer state at the next time step and the multi-scale fusion feature.

[0059] The forward candidate state, the reverse candidate state, and the hidden layer state of the previous time step are residually connected, and the residual connection result is selectively updated through the optimized update gate state to obtain the hidden layer state at the current time step; the hidden layer state at the current time step is subjected to feature extraction and fusion to generate welding strength change trend data.

[0060] For example, in the implementation of the multi-scale temporal feature extraction module, sliding windows of different sizes are used for feature extraction on the standardized dataset. Specifically, sliding windows of lengths 3, 5, and 7 are used to scan and extract welding process parameters, weld surface images, and weld ultrasonic testing data. Welding process parameters include welding current (120-180A), voltage (18-24V), and welding speed (5-12mm / s). The weld surface images are preprocessed to obtain grayscale images of size 128×128 pixels, and spatial features are extracted through convolutional layers. The weld ultrasonic testing data includes waveform signals and attenuation coefficients. For each sliding window, preliminary feature mapping is performed through a single-layer perceptron (input dimension is the same as the window size, output dimension is 64). After obtaining the hidden state at the current time step (dimension is 128), its dot product with the state at the previous time step is calculated, and then converted into temporal attention weights through the Softmax function. The temporal attention weights range from 0 to 1, weighting the features extracted at different scales. A higher weight indicates greater importance of the feature at that moment. The weighted features from different scales are then concatenated and dimensionality-reduced through a linear layer (input dimension 64×3, output dimension 128) to obtain multi-scale fused features.

[0061] The bidirectional gated loop unit improves upon the standard gated loop unit structure when processing multi-scale fusion features. A conventional gated loop unit includes an update gate and a reset gate, controlling information retention and forgetting, respectively. In this embodiment, an adaptive gating threshold is calculated based on the mean and standard deviation of the input data. Batch statistics are performed on the multi-scale fusion features (128 dimensions) to calculate the mean μ (scalar) and standard deviation σ (scalar). The adaptive gating threshold τ is calculated as μ ± k × σ, where k is an adjustable parameter, initially set to 1.5. When the output values ​​of the update and reset gates are less than the lower threshold μ - k × σ, they are multiplied by a decay coefficient of 0.8; when the output values ​​are greater than the upper threshold μ + k × σ, they are multiplied by an enhancement coefficient of 1.2; otherwise, they remain unchanged. This dynamic modulation mechanism enables the gating unit to adaptively respond to different data distribution characteristics, resulting in an optimized gating state.

[0062] The forward and backward propagation paths are constructed to fully utilize the bidirectional dependencies of temporal data. In the forward propagation path, the optimized reset gate state (dimension 128) is multiplied element-wise with the previous hidden state (dimension 128) to obtain the modulation result. The modulation result is concatenated with multi-scale fusion features and used to generate forward candidate states through a two-layer perceptron (layer 1: input dimension 256, output dimension 128, activation function ReLU; layer 2: input dimension 128, output dimension 128, no activation function). Similarly, in the backward propagation path, the optimized reset gate state is multiplied element-wise with the next hidden state to obtain the modulation result, which is then concatenated with multi-scale fusion features and used to generate backward candidate states through a two-layer perceptron with the same structure. The two propagation paths share parameters but process data in different directions: forward from past to present, and backward from future to present, thus capturing the complete temporal dependencies of the data.

[0063] The forward candidate state, the reverse candidate state, and the previous hidden state are weighted and summed, with a weight of 0.4 for the forward candidate state, 0.4 for the reverse candidate state, and 0.2 for the previous hidden state. The residual connection result is selectively updated using the optimized update gate state: the optimized update gate state (128 dimensions) is multiplied element-wise with the residual connection result, and then the complement of the optimized update gate state (1 - update gate state) is multiplied element-wise with the previous hidden state. The two are then added together to obtain the current hidden state. The current hidden state undergoes final processing: features are extracted through a fully connected layer (128 input dimensions, 64 output dimensions), and then another fully connected layer (64 input dimensions, 1 output dimension) generates welding strength change trend data. This data represents the direction and magnitude of welding strength changes over a short period, typically ranging from -10% to +10%, with positive values ​​indicating an increasing trend and negative values ​​indicating a decreasing trend.

[0064] This invention combines multi-scale temporal feature extraction with bidirectional gated recurrent units to efficiently capture the complex relationship between welding parameters and strength changes; the adaptive gate threshold mechanism improves the model's adaptability to different data distributions; and the bidirectional propagation path design makes full use of the dependencies between temporal information.

[0065] Optional,

[0066] The second branch uses an improved LSTM structure to accurately model the standardized dataset and outputs welding strength prediction data. The steps include:

[0067] The standardized dataset is decomposed into features, and the importance weights of the decomposed features are calculated based on the multilayer perceptron. The importance weights are then combined with the corresponding decomposed features to obtain weighted fusion features.

[0068] A dynamic head-adaptive multi-head attention module is constructed. The multi-head attention module adaptively determines the number of attention heads based on the current state of the weighted fusion feature and maps the weighted fusion feature into a query matrix, a key matrix, and a value matrix. A position-aware matrix is ​​generated based on sine and cosine position encoding. The attention weights are obtained by combining the product of the position-aware matrix with the query matrix and the key matrix.

[0069] A channel attention enhancement module is constructed. The channel attention enhancement module performs global average pooling on the attention weights to obtain channel features, calculates the weight coefficients of the channel features based on a multilayer perceptron, and performs channel-by-channel product operation on the weight coefficients and the attention weights to obtain enhanced features.

[0070] An improved Long Short-Term Memory (LSTM) unit is constructed, and the enhanced features are used as input. The forget gate of the LSM unit is adaptively adjusted based on the state of the memory unit at the previous time step, and the input gate is adaptively adjusted based on the hyperbolic tangent transform of the state of the memory unit at the previous time step to obtain the hidden layer state at the current time step.

[0071] Dynamic coefficients are calculated based on the feature similarity of the hidden state of adjacent layers. The dynamic coefficients are used to perform residual connections on the hidden state of adjacent layers. Dense connections are then performed on the hidden state of multiple layers to obtain interlayer fusion features. Welding strength prediction results are generated based on the interlayer fusion features.

[0072] For example, when performing feature decomposition on a standardized dataset, the empirical mode decomposition method is used to decompose time-series data such as welding current, welding voltage, and welding speed into multiple intrinsic mode functions (IMFs). Taking welding current as an example, the standardized current data is decomposed into four IMFs, each corresponding to different frequency features: the first IMF mainly contains high-frequency fluctuation information (period less than 0.5 seconds), the second contains mid-frequency information (period 0.5-2 seconds), the third contains low-frequency information (period 2-10 seconds), and the fourth contains trend information (period greater than 10 seconds). For weld surface image data, spatial features are extracted using a pre-trained feature extractor to obtain a 128-dimensional feature vector; for ultrasonic detection data, waveform features are extracted to obtain a 64-dimensional feature vector. When calculating the importance weights of the decomposed features based on a multilayer perceptron, a three-layer neural network is constructed, with the number of input layer nodes being the same as the number of decomposed features, the hidden layer containing 64 neurons, and the number of output layer nodes being the same as the number of decomposed features. For the four intrinsic mode functions of welding current, the calculated importance weights during the welding process of titanium-steel composite plates are 0.15, 0.25, 0.35, and 0.25, respectively, indicating that low- and mid-frequency features are more important for predicting welding strength. These importance weights are then weighted and combined with the corresponding decomposition features to obtain a weighted fusion feature. The dimension is the same as the original feature dimension, but each frequency component is either enhanced or suppressed according to its importance.

[0073] When constructing a multi-head attention module with adaptive dynamic head count, an adaptive head count adjustment mechanism was designed. The basic head count was set to 4, and dynamically adjusted according to the information entropy of the weighted fusion features. In the specific implementation, the distribution entropy value of each dimension of the weighted fusion features was calculated. When the entropy value was greater than a preset threshold (e.g., 1.8), it indicated that the features contained rich information, and the head count was increased to 6; when the entropy value was less than another threshold (e.g., 1.2), it indicated that the features were redundant, and the head count was reduced to 2. In the actual welding process, during the transition stage when the welding current changed from 185A to 205A, the feature entropy value usually exceeded 1.8. At this time, 6 attention heads were used to capture richer feature relationships; while in the stable welding stage, the entropy value was usually between 1.3 and 1.5, and 4 attention heads were used to balance computational efficiency and expressive power. When mapping the weighted fusion features to a query matrix, key matrix, and value matrix, three independent fully connected layers were used. The weight matrix dimension of each layer was the feature dimension (e.g., 256) multiplied by the projection dimension (64). A position-aware matrix is ​​generated based on sine and cosine position coding. Position codes are calculated using sine and cosine functions of different frequencies, with the position code dimension being the same as the query / key / value matrix. When combining the position-aware matrix with the product of the query and key matrices, the position codes are added to both matrices, their dot product is calculated, and then divided by a scaling factor of 8 (the square root of 64). Finally, the attention weights are obtained by normalization using the softmax function. During the welding process of titanium-steel composite plates, the attention mechanism automatically focuses on important moments. For example, when the welding current increases from 195A to 210A, the attention weights significantly increase near the point of change (from an average of 0.05 to approximately 0.15).

[0074] When constructing the channel attention enhancement module, global average pooling is performed along the sequence dimension of the attention weight matrix to obtain the average activation value of each channel, forming a channel feature vector. For an attention weight matrix of dimension 256, pooling yields a 256-dimensional channel feature vector. The weight coefficients of the channel features are calculated based on a multilayer perceptron, employing a "bottleneck" structure. The first layer reduces the number of channels to 1 / 16 of the original (i.e., 16 neurons) and uses the ReLU activation function; the second layer restores the original number of channels (256) and uses the Sigmoid activation function to output weight coefficients between 0 and 1. This bottleneck structure effectively reduces the number of parameters and improves computational efficiency. In practical applications, the channel attention module assigns different weights to different feature channels. For example, when the welding speed changes significantly (e.g., from 32 cm / min to 28 cm / min), the weights of speed-related channels automatically increase by about 30%, while the weights of other channels decrease accordingly. The enhanced features are obtained by multiplying the channel weight coefficients and attention weights channel by channel, i.e., multiplying the attention weight of each channel by the corresponding channel coefficient, thus strengthening the feature representation of important channels.

[0075] When constructing the improved Long Short-Term Memory (LSTM) unit, two key improvements were made to the standard LSTM structure. The forget gate adaptively adjusts based on the previous memory unit state. Specifically, this is achieved by adding an adjustment factor dependent on the previous memory unit state to the standard forget gate calculation. This adjustment factor is multiplied by the original forget gate output through a single-layer neural network (input is the previous memory unit state, output is the adjustment factor) to obtain the adjusted forget gate value. In practical applications, when the memory unit state value is large (absolute value greater than 0.8), the adjustment factor is typically greater than 1 (approximately 1.2), enhancing forgetting ability and preventing state saturation; when the memory unit state value is moderate (absolute value between 0.3 and 0.6), the adjustment factor is close to 1, maintaining standard behavior; when the memory unit state value is small (absolute value less than 0.2), the adjustment factor is less than 1 (approximately 0.8), reducing forgetting and retaining useful information. The input gate adaptively adjusts based on the hyperbolic tangent transform of the memory cell state at the previous time step. This is achieved by applying the hyperbolic tangent function to the memory cell state at the previous time step, passing the result through a neural network to calculate an adjustment factor, and multiplying this factor by the original input gate output. This mechanism allows the input gate to adaptively adjust according to the saturation level of the memory cells.

[0076] When calculating the dynamic coefficients based on the feature similarity of the hidden states of adjacent layers, cosine similarity is used to measure the similarity between the hidden states of two layers. For the improved LSTM network with three stacked layers, the cosine similarity of the hidden states of adjacent layers (first and second layers, second and third layers) is calculated. When the similarity is high (greater than 0.8), the dynamic coefficient is set to a smaller value (e.g., 0.3) to reduce information redundancy; when the similarity is low (less than 0.5), the dynamic coefficient is set to a larger value (e.g., 0.7) to preserve the difference information. The dynamic coefficients are used to perform residual connections on the hidden states of adjacent layers, specifically by adding the current hidden state to the previous hidden state multiplied by the dynamic coefficient. This adaptive residual connection method can effectively alleviate the gradient vanishing problem and improve the training stability of deep networks. When performing dense connections on multiple layers of hidden states, the hidden states of all layers are concatenated and then fused through dimensionality reduction using fully connected layers. In a three-layer network, if the hidden state dimension of each layer is 128, the concatenated state has 384 dimensions. This is then reduced to 128 dimensions through a fully connected layer, resulting in inter-layer fusion features. Test results show that, compared to using the top-level hidden state alone, inter-layer fusion features improve the average accuracy of welding strength prediction by 5.8 percentage points.

[0077] When generating welding strength prediction results based on interlayer fusion features, a mapping is achieved through a two-layer fully connected network. The first layer maps the 128-dimensional fusion features to 64 dimensions using the ReLU activation function; the second layer maps the 64-dimensional features to 1 dimension and outputs the welding strength prediction value.

[0078] The multi-scale temporal feature extraction and dynamic head number adaptive attention mechanism of this invention effectively capture key features in the welding process. The channel attention enhancement module and adaptive gating mechanism significantly improve the model's adaptability to parameter fluctuations, while the inter-layer residual connection and dense connection strategy solves the problem of training deep networks. The overall scheme significantly reduces computational latency while ensuring prediction accuracy, achieving real-time high-precision prediction of welding strength and providing a reliable foundation for intelligent welding control.

[0079] Optional,

[0080] The steps for adaptively fusing the welding strength variation trend and the welding strength prediction data to obtain a comprehensive prediction result include:

[0081] A time-series consistency score is calculated based on the temporal gradient characteristics of the welding strength change trend and the fluctuation characteristics of the welding strength prediction data. The time-series consistency score is then combined with the real-time status data of the welding process parameters to generate a dynamic fusion weight.

[0082] The dynamic fusion weights are adaptively modulated based on historical prediction accuracy to generate optimized fusion weight coefficients. The optimized fusion weight coefficients are then weighted and combined with the welding strength change trend and the welding strength prediction data to obtain an initial fusion result. The residual compensation value is calculated based on the deviation between the initial fusion result and the historical prediction data. The residual compensation value is then adaptively corrected with the initial fusion result to obtain a comprehensive prediction result.

[0083] For example, when calculating the temporal consistency score based on the temporal gradient features of welding strength change trends and the fluctuation features of welding strength prediction data, it is necessary to extract feature representations from both types of data. For welding strength change trend data, the first-order difference values ​​at 20 consecutive time points are calculated as the temporal gradient feature, representing the rate of strength change. For welding strength prediction data, the ratio of the standard deviation to the mean at nearly 30 time points is extracted as the fluctuation feature, representing the stability of the strength prediction. This ratio is typically small during the stable welding phase, but increases significantly during the parameter adjustment phase. The temporal consistency score is obtained by calculating the cross-correlation coefficient between the temporal gradient feature and the fluctuation feature. The larger the absolute value of the cross-correlation coefficient, the more consistent the change patterns of the two prediction results. For example, when the two prediction methods respond consistently to changes in welding parameters, the cross-correlation coefficient is greater than 0.7; when the two methods respond inconsistently, the cross-correlation coefficient drops below 0.3.

[0084] When generating dynamic fusion weights by combining the time-series consistency score with real-time status data of welding process parameters, the current values ​​and rates of change of welding current, voltage, and speed are considered. Specifically, the time-series consistency score is concatenated with the normalized rates of change of welding current, voltage, and speed to form a fusion feature vector. When the welding current changes significantly, its rate of change affects the calculation of the fusion weights. For example, when the current rapidly increases from 200 amps to 220 amps, the rate of change is large, and the system tends to adopt the prediction result that is more sensitive to rapid changes. A three-layer neural network is used to process the fusion feature vector. The hidden layers contain 16 and 8 neurons respectively, and the output layer contains 2 neurons, corresponding to the fusion weights of the two prediction branches. The output layer uses a softmax activation function to ensure that the sum of the two weights is 1. During the welding process of titanium-steel composite plates, when the welding parameters are stable, the weight of the accurate prediction branch is usually high, around 0.7; when the parameters change rapidly, the weight of the trend branch increases to around 0.6.

[0085] When adaptively modulating the dynamic fusion weights based on historical prediction accuracy, a sliding window is maintained to record the historical accuracy of the two prediction branches. The window size is set to 100 time points, recording the root mean square error (RMSE) between the predicted and actual measured values ​​for each branch. In practical applications, under normal welding conditions, the RMSE of the lightweight RNN branch is approximately 5% of the predicted value, while the RMSE of the improved LSTM branch is approximately 3%. Modulation coefficients are calculated based on historical accuracy: if the historical accuracy of one branch is significantly higher than that of another branch (error difference exceeding 50%), the weight of that branch is increased by multiplying the original weight by a scaling factor. For example, when the error of the LSTM branch is 70% lower than that of the RNN branch, the original weight of the LSTM branch (0.6) is adjusted to 0.75. After generating the optimized fusion weight coefficients, the weights are renormalized to ensure a sum of 1.

[0086] When the optimized fusion weight coefficients are weighted and combined with the welding strength change trend and welding strength prediction data respectively, the weights of each component are directly multiplied by their corresponding predicted values, and then summed to obtain the initial fusion result. For example, when the current increases from 190 Amperes to 210 Amperes, the voltage remains at 26 Volts, and the welding speed decreases from 30 cm / min to 28 cm / min, the lightweight RNN branch predicts an 8.5% increase in strength, while the improved LSTM branch predicts a specific strength value that increases from 520 MPa to 550 MPa (an increase of 5.8%). If the weight of the lightweight RNN branch is 0.4 and the weight of the improved LSTM branch is 0.6, then the initial fusion result is a 6.9% increase in strength, which translates to an increase from 520 MPa to 556 MPa.

[0087] When calculating residual compensation values ​​based on the deviation between the initial fusion results and historical prediction data, the systematic deviation between recent predicted values ​​and actual measurements is analyzed. A sliding window of length 50 is maintained to record prediction deviation values. An autoregressive moving average model is applied to these deviation values ​​to predict possible systematic deviations at the current moment. During the welding process, if the predicted values ​​at multiple consecutive time points are approximately 2% higher than the actual values, this systematic deviation is identified and a corresponding compensation value is generated. The calculation of residual compensation values ​​considers the persistence and trend of the deviation. When the deviation remains stable, the compensation value is close to the average deviation; when the deviation fluctuates significantly, the compensation value decreases to avoid overcorrection.

[0088] When adaptively correcting the residual compensation value with the initial fusion result, the correction coefficient is determined based on the current welding state. During the parameter stability phase, the correction coefficient is close to 1, fully utilizing residual compensation; during the rapid parameter change phase, the correction coefficient is smaller to reduce the impact of compensation. The correction coefficient ranges from 0.2 to 1.0, inversely proportional to the rate of change of welding parameters. The adaptive correction is implemented by multiplying the residual compensation value by the correction coefficient and then adding it to the initial fusion result to obtain the final comprehensive prediction result. During the welding of titanium-steel composite plates, the consistency between the predicted result after residual compensation and the actual measured value is significantly improved. Especially during long-term welding, the compensation mechanism effectively eliminates systematic prediction biases and maintains prediction accuracy.

[0089] In practical applications, when the welding process of titanium-steel composite plates shifts from one set of parameters to another, the adaptive fusion mechanism intelligently balances rapid response and accurate prediction. For example, when welding suddenly transitions from a steady state to a parameter change phase, the system increases the weight of the lightweight RNN branch to quickly capture the changing trend; once the parameters stabilize again, the system gradually increases the weight of the improved LSTM branch, providing more accurate prediction results. This dynamic adjustment mechanism enables the integrated prediction system to respond quickly to changes in process parameters while providing high-precision predictions in the stable phase, meeting the dual requirements of titanium-steel composite plate welding process control.

[0090] The comprehensive prediction results are calculated 20 times per second to ensure that the predictions can reflect changes in welding strength in a timely manner. A parallel computing architecture is adopted, with lightweight RNN branches and improved LSTM branches performing calculations simultaneously. The fusion module processes the outputs of the two branches to achieve high-efficiency real-time prediction.

[0091] The adaptive fusion method of this invention effectively combines the fast response capability of lightweight RNNs and the accurate prediction advantages of improved LSTMs through temporal consistency scoring and dynamic weight modulation, achieving efficient and accurate prediction of the welding strength of titanium-steel composite plates. The historical accuracy feedback mechanism automatically adjusts the weights of each branch, and the residual compensation technique effectively eliminates systematic biases, significantly improving prediction stability. This method maintains a fast response during drastic parameter changes and provides high-precision predictions in the steady-state phase, balancing speed and accuracy. It provides reliable data support for welding quality control and significantly improves the intelligence level of the titanium-steel composite plate welding process.

[0092] Optional,

[0093] The steps for constructing a collaborative optimization control scheme based on reinforcement learning, using the welding process parameters, weld surface image, and prediction error as state inputs, calculating the parameter coupling degree between welding current, welding voltage, and welding speed, generating control weights for welding current, welding voltage, and welding speed based on the parameter coupling degree, and generating optimal adjustment strategy data include:

[0094] Welding process parameters, weld surface images, and prediction errors are combined to obtain fusion state characteristics; based on the fusion state characteristics, the Pearson correlation coefficients between welding current, welding voltage, and welding speed are calculated to generate a parameter coupling matrix;

[0095] The initial weight of each welding process parameter is calculated based on the coupling strength between the parameters in the parameter coupling matrix. The initial weight is then combined with the weight adjustment gradient of the reward function calculated by weighting the welding quality index and the process economic index to obtain the optimized dynamic control weight.

[0096] A deep Q-network structure is constructed, and the fused state features and the optimized dynamic control weights are input into the deep Q-network structure to generate a state-action value estimate; an ε-greedy strategy is used to select the adjustment action based on the state-action value estimate.

[0097] The temporal difference error is calculated based on the immediate reward and the maximum value estimate of the next state, and the parameters of the deep Q network structure are optimized and updated using the temporal difference error.

[0098] The optimized dynamic control weights are weighted and modulated with the adjustment actions selected based on the deep Q-network structure to generate welding current adjustment, welding voltage adjustment, and welding speed adjustment, thereby obtaining the optimal adjustment strategy.

[0099] Combination Figure 2The flowchart illustrates the collaborative optimization control process for welding parameters based on reinforcement learning. When constructing a collaborative optimization control scheme based on reinforcement learning, the welding process parameters, weld surface image, and prediction error are first combined to obtain fused state features. Welding process parameters include welding current, welding voltage, and welding speed. These parameters are standardized and converted to the range [0,1]. Taking titanium-steel composite plate welding as an example, the current range is typically 150-250 amperes, the voltage range is 20-30 volts, and the welding speed range is 20-40 cm / min. For the weld surface image, a pre-trained convolutional neural network is used to extract features. The image is first adjusted to 224×224 pixels, and spatial features are extracted through 5 convolutional layers, ultimately obtaining a 128-dimensional image feature vector. The prediction error includes the absolute value of the intensity prediction error and its trend, expressed as the mean and standard deviation of the error over the most recent 20 time steps. These features are concatenated to obtain a fused state feature vector with a dimension of 256.

[0100] Based on the fusion state characteristics, Pearson correlation coefficients were calculated between each pair of welding current, welding voltage, and welding speed to generate a parameter coupling degree matrix. During the calculation, parameter data from the most recent 100 time steps were extracted from the fusion state characteristics, and Pearson correlation coefficients were calculated between current and voltage, current and speed, and voltage and speed, respectively. The correlation coefficient values ​​range from -1 to 1, with the absolute value closer to 1 indicating stronger coupling between parameters. In the welding process of titanium-steel composite plates, the correlation coefficient between current and voltage is typically between 0.65 and 0.75, indicating a strong positive correlation; the correlation coefficient between current and speed is typically between -0.4 and -0.6, indicating a moderate negative correlation; and the correlation coefficient between voltage and speed is typically between -0.2 and -0.3, indicating a weak negative correlation. These correlation coefficients form a 3×3 parameter coupling degree matrix, with diagonal elements being 1, indicating that the parameter is perfectly correlated with itself.

[0101] The initial weight of each welding process parameter is calculated based on the coupling strength between parameters in the parameter coupling matrix. Specifically, for each parameter, the sum of the absolute values ​​of its correlation coefficients with other parameters is calculated, and then normalized to obtain the initial weight. For example, the initial weight of current is calculated by dividing the sum of the absolute values ​​of the correlation coefficients of current with all other parameters (including itself) by the sum of the absolute values ​​of all correlation coefficients. Under typical welding conditions, the initial weights of current, voltage, and speed are approximately 0.45, 0.35, and 0.20, respectively, reflecting the dominant role of current in the welding process. Welding quality indicators include weld formation consistency, joint strength, and porosity; process economic indicators include energy consumption, material utilization, and production efficiency. These indicators are mapped to reward values ​​using a multilayer perceptron network. The network contains two hidden layers with 32 and 16 nodes respectively, and an output layer with one node representing the comprehensive reward value. The gradient of the reward function with respect to the initial weights is calculated, i.e., the rate of change of the reward value relative to the weights of each parameter. In the welding of titanium-steel composite plates, when the current weight increases by 0.05, if the reward value increases by 0.02, the gradient of the current weight is 0.4; similarly, the gradients of the voltage and speed weights are calculated. The initial weights are combined with the weight adjustment gradients, and the adjustment magnitude is controlled by the learning rate, which is set to 0.1. If the gradient of a parameter is positive, its weight is increased; if the gradient is negative, its weight is decreased. The optimized dynamic control weights are obtained by renormalizing the adjusted weights to ensure that the sum of the weights of the three parameters is 1.

[0102] When constructing the deep Q-network structure, a dual-network architecture is adopted, including an evaluation network and a target network. The evaluation network is used to generate the action value in the current state, and the target network is used to calculate the value estimate for the next state. The network structure contains three fully connected layers. The number of nodes in the input layer is the same as the dimension of the fused state features (256), the number of nodes in the hidden layer are 128 and 64 respectively, and the number of nodes in the output layer is the size of the action space. In this embodiment, each welding parameter (current, voltage, speed) has 5 possible adjustment actions (significant increase, slight increase, no change, slight decrease, significant decrease), for a total of 5×5×5=125 combined actions. The fused state features and optimized dynamic control weights are input into the deep Q-network to generate state action value estimates. The dynamic control weights introduce parameter importance information by modulating the feature representation inside the network. Specifically, after the output of the first hidden layer, the feature vector is grouped according to the parameter type and multiplied by the weight of the corresponding parameter. For example, the feature component related to current is multiplied by the dynamic control weight of current. An ε-greedy strategy is used to select the adjustment action based on the state action value estimate. The initial value of ε is set to 0.9, which means that there is a 90% probability of randomly selecting an action and a 10% probability of selecting the action with the highest estimated value. As training progresses, the value of ε decays linearly, eventually dropping to 0.1, which increases the probability of selecting the optimal action.

[0103] The immediate reward is jointly determined by welding quality and process economic indicators. Welding quality evaluation is based on weld formation consistency (40%), joint strength (40%), and weld surface porosity (20%); process economic evaluation is based on energy consumption (50%), material utilization (30%), and production efficiency (20%). The quality evaluation and economic evaluation constitute the comprehensive reward in a 7:3 ratio. In titanium-steel composite plate welding, the comprehensive reward under ideal conditions is approximately 0.9, which may drop below 0.3 under poor conditions. The maximum value estimate of the next state is calculated through the target network, i.e., the next state after executing the selected action is input into the target network, and the maximum Q value among all possible actions is taken. The temporal difference error is the immediate reward plus a discount factor multiplied by the maximum value estimate of the next state, and then subtracted from the current state action value estimate. The discount factor is set to 0.95 to balance the importance of the current reward and future rewards. The parameters of the deep Q-network structure are optimized and updated using the temporal difference error, employing the stochastic gradient descent method with a learning rate of 0.001 and a batch size of 64. Simultaneously, an experience replay mechanism is used to maintain an experience pool with a capacity of 10,000, randomly sampling historical experiences for learning to reduce sample correlation. The target network parameters are updated every 100 steps, copied from the evaluation network, to enhance training stability.

[0104] The optimized dynamic control weights are weighted and modulated with the adjustment actions selected based on a deep Q-network structure to generate welding current, welding voltage, and welding speed adjustments. The adjustment actions are represented as discrete values, such as +2 for a large increase, +1 for a small increase, 0 for no change, -1 for a small decrease, and -2 for a large decrease. These discrete values ​​are multiplied by the corresponding step size factors (5 amperes for current, 0.5 volts for voltage, and 2 cm / min for speed), and then multiplied by the dynamic control weight of the corresponding parameter to obtain the actual adjustment amount. For example, if the dynamic control weight of the current is 0.5 and the selected adjustment action is a small increase (+1), then the current adjustment amount is +1 × 5 × 0.5 = +2.5 amperes. Through this weighted modulation mechanism, parameters with higher importance will have a larger adjustment effect, while the adjustment range of parameters with lower importance will be correspondingly smaller. The final generated adjustment strategy considers the coupling relationship between parameters, avoids conflicts in parameter adjustments, and achieves collaborative optimization control of welding process parameters.

[0105] For example, when weld strength is detected to be below the target value and porosity is present, the current (increased by 4 amperes), voltage (increased by 0.6 volts), and speed (decreased by 1.5 cm / min) are adjusted simultaneously based on parameter coupling and the current state. This coordinated adjustment scheme can more effectively improve welding quality while keeping energy consumption within a reasonable range compared to adjusting a single parameter.

[0106] This invention achieves intelligent collaborative control of welding process parameters through parameter coupling analysis and dynamic control weight generation. This method fully considers the complex interrelationships between parameters, avoiding the limitations of traditional single-parameter adjustment methods. The deep Q-network structure combined with an optimized reward function balances welding quality and economic requirements, achieving multi-objective optimization. The weighted modulation mechanism of dynamic weights and adjustment actions ensures the rationality and effectiveness of parameter adjustment, significantly improving the stability and welding strength of titanium-steel composite plates while reducing energy consumption and material waste, providing strong support for intelligent manufacturing.

[0107] Optional,

[0108] The steps for calculating the initial weight of each welding process parameter based on the coupling strength between parameters in the parameter coupling matrix, and combining the initial weight with the weight adjustment gradient of the reward function calculated based on welding quality indicators and process economic indicators to obtain the optimized dynamic control weights include:

[0109] The basic coupling degree is obtained by calculating the Pearson correlation coefficient between each pair of welding current, welding voltage, and welding speed; wavelet transform is performed on the welding current, welding voltage, and welding speed, and the time-frequency coupling characteristics under different frequency bands are calculated based on the cross wavelet transform coefficients; the causal relationship strength between the welding current, welding voltage, and welding speed is calculated based on the Granger causality test method; the basic coupling degree, the time-frequency coupling characteristics, and the causal relationship strength are weighted and fused to generate a fused coupling degree matrix.

[0110] The initial weights of welding current, welding voltage, and welding speed are calculated based on the fusion coupling degree matrix; a weld formation consistency score is calculated based on weld contour characteristics; a mechanical property score is calculated based on tensile strength, yield strength, and elongation; and a process stability score is calculated based on arc characteristic fluctuations and molten pool dynamic behavior. The weld formation consistency score, mechanical property score, and process stability score are weighted and combined to obtain a welding quality evaluation value; a process economic evaluation value is calculated based on energy consumption indicators and material utilization rate indicators; and a comprehensive reward value is obtained by weighting the welding quality evaluation value and the process economic evaluation value.

[0111] Calculate the gradient of the comprehensive reward value with respect to the initial weight, combine the gradient with the initial weight, and introduce the momentum term of the difference between two adjacent weights to generate dynamic weights; calculate the compensation adjustment amount based on the historical adjustment effect; combine the compensation adjustment amount with the dynamic weights to obtain the optimized dynamic control weights.

[0112] For example, the basic coupling degree is obtained by calculating the Pearson correlation coefficient between each pair of welding current, welding voltage, and welding speed. Parameter data from the most recent 500 time points during the welding process are collected, and the Pearson correlation coefficients between each parameter are calculated. During the welding of titanium-steel composite plates, current and voltage typically show a moderate positive correlation, with a correlation coefficient of approximately 0.68; current and welding speed typically show a weak negative correlation, with a correlation coefficient of approximately -0.42; and voltage and welding speed typically show a weak correlation, with a correlation coefficient of approximately -0.25. These correlation coefficients form a 3×3 basic coupling degree matrix, with diagonal elements of 1, indicating that the parameter is perfectly correlated with itself. This matrix reflects the linear correlation between parameters but cannot capture the coupling relationships at different time scales and potential nonlinear relationships.

[0113] Wavelet transforms were performed on welding current, welding voltage, and welding speed, using Morlet wavelets as the mother wavelet. Continuous wavelet transforms were applied to each parameter, decomposing them into multiple frequency bands. Typical frequency band divisions are: high frequency (0.5-2Hz), corresponding to rapid fluctuations; mid-frequency (0.1-0.5Hz), corresponding to medium-time-scale changes; and low frequency (0.01-0.1Hz), corresponding to slow-moving trends. The correlation between parameters in each frequency band was calculated using cross-wavelet transforms. In the welding of titanium-steel composite plates, the coupling coefficient between current and voltage is typically 0.72 in the high-frequency band, indicating a high correlation in instantaneous changes; 0.65 in the mid-frequency band; and 0.58 in the low-frequency band. The coupling coefficient between current and speed is -0.38 in the high-frequency band, -0.45 in the mid-frequency band, and -0.50 in the low-frequency band, indicating a stronger negative correlation in long-term trends. The coupling coefficients between voltage and speed in each frequency band are -0.22, -0.28, and -0.30, respectively. The coupling characteristics of these multi-band frequencies form a time-frequency coupling characteristic matrix, which provides a more comprehensive description of the dynamic relationship between parameters.

[0114] For each pair of parameters, a Granger causality test with a lag order of 5 was performed, and the F-statistic and corresponding p-value were calculated. A larger F-statistic and a smaller p-value indicate a more significant causal relationship. During the welding process of titanium-steel composite plates, the Granger causality F-statistic for the relationship between current and voltage was approximately 12.6, with a p-value less than 0.01, indicating that current changes have a significant predictive ability on voltage. The F-statistic for voltage and current was approximately 8.3, with a p-value less than 0.05, indicating a bidirectional causal relationship, but with current having a stronger influence on voltage. The F-statistic for current and velocity was approximately 9.7, with a p-value less than 0.01; the F-statistic for velocity and current was approximately 3.2, with a p-value greater than 0.05, indicating that current primarily has a unidirectional influence on velocity. The F-statistic for voltage and velocity was approximately 5.4, with a p-value less than 0.05; the F-statistic for velocity and voltage was approximately 2.8, with a p-value greater than 0.05. These statistics form a causality strength matrix, reflecting the predictive influence between the parameters.

[0115] The basic coupling degree matrix, time-frequency coupling characteristic matrix, and causal relationship strength matrix were normalized and then weighted and averaged according to their respective weight coefficients. In the titanium-steel composite plate welding application, the weight of the basic coupling degree was 0.3, the weight of the time-frequency coupling characteristic matrix was 0.4, and the weight of the causal relationship strength matrix was 0.3. The fused coupling degree matrix more comprehensively reflects the correlation between parameters. For example, the fused coupling degree of current and voltage was 0.71, the fused coupling degree of current and velocity was -0.48, and the fused coupling degree of voltage and velocity was -0.29. These values ​​indicate that the coupling relationship between current and voltage is the strongest, followed by the negative correlation between current and velocity.

[0116] The initial weight calculation method is as follows: for each parameter, calculate the sum of the absolute values ​​of its coupling degree with all parameters (including itself), and then normalize to obtain the initial weight. During the welding process of titanium-steel composite plates, the initial weights for current, voltage, and speed are approximately 0.48, 0.32, and 0.20, respectively, reflecting the dominant role of current in the welding process. When calculating the mechanical property score based on tensile strength, yield strength, and elongation, for titanium-steel composite plates, the standard target value for tensile strength is 550 MPa, the target value for yield strength is 450 MPa, and the target value for elongation is 18%. The ratio of the actual measured value to the target value is used as the score rate for each indicator, with an upper limit of 1. The score rates for tensile strength, yield strength, and elongation are weighted and averaged in a ratio of 5:3:2 to obtain the mechanical property score, with a maximum score of 100 points.

[0117] Arc characteristic fluctuations were assessed using the fluctuation coefficients of current and voltage (standard deviation divided by mean), while the dynamic behavior of the molten pool was obtained through analysis of molten pool image sequences captured by a high-speed camera. Current fluctuation coefficients below 0.05 and voltage fluctuation coefficients below 0.07 were considered stable, as was the molten pool area change rate below 0.10. The arc stability and molten pool stability scores were weighted and averaged in a 5:5 ratio to obtain the process stability score, with a maximum score of 100.

[0118] The welding quality evaluation value is obtained by weighting and combining the mechanical property score and the process stability score. In the welding of titanium-steel composite plates, the weights of these two indicators are 0.7 and 0.3, respectively, reflecting the core position of mechanical properties in welding quality. The weighted average welding quality evaluation value ranges from 0 to 100 points, and is usually normalized to the range of 0 to 1. The process economic evaluation value is calculated based on energy consumption and material utilization indicators. The energy consumption indicator is calculated by the energy consumption per unit weld length, with a standard value of 25 kJ / cm; the material utilization rate is evaluated by the spatter rate and the excess height ratio, with an ideal spatter rate of less than 1% and an excess height ratio of less than 5%. The energy consumption indicator and the material utilization rate indicator are weighted and averaged in a 7:3 ratio to obtain the process economic evaluation value, ranging from 0 to 100 points, and usually normalized to the range of 0 to 1. The comprehensive reward value is obtained by weighting and calculating the welding quality evaluation value and the process economic evaluation value. In practical applications, the weight ratio of quality to economy is 8:2, indicating that quality is the primary consideration. The comprehensive reward value ranges from 0 to 1, with a higher value indicating a better welding effect.

[0119] When calculating the gradient of the overall reward value with respect to the initial weights, a perturbation method is used: a small perturbation (e.g., 0.05) is added to each parameter weight, the change in the overall reward value is calculated, and the ratio of the two is used as the gradient estimate for that parameter weight. In the welding of titanium-steel composite plates, if an increase of 0.05 in the current weight leads to an increase of 0.03 in the overall reward value, then the gradient of the current weight is 0.6; if an increase of 0.05 in the voltage weight leads to an increase of 0.02 in the overall reward value, then the gradient of the voltage weight is 0.4; if an increase of 0.05 in the velocity weight leads to an increase of 0.01 in the overall reward value, then the gradient of the velocity weight is 0.2. The gradients are combined with the initial weights, and the gradient ascent method is used to update the weights. With a learning rate of 0.1, the updated current weight is 0.48 + 0.1 × 0.6 = 0.54, the updated voltage weight is 0.32 + 0.1 × 0.4 = 0.36, and the updated velocity weight is 0.20 + 0.1 × 0.2 = 0.22.

[0120] A momentum term is introduced based on the difference between two adjacent weights, with a momentum coefficient set to 0.8. The current weight update considers not only the current gradient but also the previous update. For example, if the previous current weight update was 0.03, the current update is 0.1 × 0.6 + 0.8 × 0.03 = 0.084, and the updated current weight is 0.48 + 0.084 = 0.564. Similarly, the weights for voltage and velocity are updated. When calculating the compensation adjustment based on historical adjustment effects, the effects of the last 20 adjustments are analyzed, comparing the changes in the overall reward value before and after adjustment. If an increase in the weight of a specific parameter leads to a continuous decrease in the overall reward value, negative compensation is given; otherwise, positive compensation is given. In the welding of titanium-steel composite plates, if the current weight continuously increases but the increase in overall reward decreases, the compensation adjustment may be -0.02; if an increase in the voltage weight leads to a significant increase in reward, the compensation adjustment may be +0.03; if the change in the velocity weight has little impact on the reward, the compensation adjustment may be +0.01. The compensation adjustment is combined with the dynamic weights to obtain the optimized dynamic control weights. The final weight for current is 0.564 - 0.02 = 0.544, the final weight for voltage is 0.36 + 0.03 = 0.39, and the final weight for speed is 0.22 + 0.01 = 0.23. These weights are normalized to ensure a sum of 1, resulting in the final dynamic control weights: current weight 0.47, voltage weight 0.33, and speed weight 0.20.

[0121] In the welding application of titanium-steel composite plates, when the welding process transitions from a steady state to a transitional state (such as a change in the material thickness at the weld area), the optimized dynamic control weights described above can be automatically adjusted according to the new parameter coupling relationships. For example, in the transition region where the plate thickness increases from 8mm to 12mm, the current weight increases to 0.52, the voltage weight decreases to 0.30, and the speed weight remains at 0.18, reflecting the crucial role of the current parameter in responding to thickness changes.

[0122] This invention comprehensively evaluates the interactions between welding parameters by integrating multi-dimensional parameter coupling analysis methods, overcoming the limitations of traditional single correlation analysis. By combining a comprehensive evaluation system of welding mechanical properties and process stability, it achieves precise dynamic adjustment of control weights. The introduction of momentum terms and compensation adjustment mechanisms effectively improves the stability and adaptability of weight updates.

[0123] Optional,

[0124] The steps of weighted modulation of the optimized dynamic control weights with the adjustment actions selected based on the deep Q-network structure to generate welding current adjustment, welding voltage adjustment, and welding speed adjustment, and to obtain the optimal adjustment strategy, include:

[0125] The optimized dynamic control weights are standardized and multiplied with the adjustment actions output by the deep Q network to obtain the basic current adjustment, basic voltage adjustment, and basic speed adjustment. A compensation value is calculated based on the deviation between the target parameter value and the current parameter value. The compensation value is combined with the corresponding basic adjustment to obtain the total current adjustment, total voltage adjustment, and total speed adjustment.

[0126] Calculate the stability constraint values ​​for the total current regulation, total voltage regulation, and total speed regulation. The stability constraint value is the ratio of the absolute value of the total regulation to the product of the current parameter value and the stability constraint coefficient. Based on the stability constraint value, perform a first adjustment on the total current regulation, total voltage regulation, and total speed regulation according to a preset ratio. Calculate the pairwise coupling constraint values ​​between each of the total current regulation, total voltage regulation, and total speed regulation after the first adjustment. Iteratively optimize the total current regulation, total voltage regulation, and total speed regulation based on the coupling constraint values ​​until the coupling constraint values ​​meet a preset coupling threshold. Output the optimized total current regulation, total voltage regulation, and total speed regulation as the optimal regulation strategy.

[0127] For example, the optimized dynamic control weights are standardized and multiplied with the adjustment action output by the deep Q-network to obtain the base current adjustment, base voltage adjustment, and base speed adjustment. The adjustment action output by the deep Q-network is a discrete value. Taking current as an example, it is usually divided into 5 levels: large increase (+2), small increase (+1), unchanged (0), small decrease (-1), and large decrease (-2). During standardization, base step sizes are set for current, voltage, and speed respectively: the base current step size is 5 amperes, the base voltage step size is 0.5 volts, and the base speed step size is 2 cm / min. The discrete value output by the deep Q-network is multiplied by the corresponding base step size, and then multiplied by the dynamic control weight of the corresponding parameter to obtain the base adjustment amount. For example, if the dynamic control weight of the current is 0.45, and the current adjustment action output by the deep Q-network is a small increase (+1), then the base current adjustment amount is +1 × 5 × 0.45 = +2.25 amperes. Similarly, if the dynamic control weight of voltage is 0.35, the voltage regulation action of the deep Q network output is to remain unchanged (0), then the basic voltage regulation is 0; if the dynamic control weight of speed is 0.20, the speed regulation action of the deep Q network output is to decrease slightly (-1), then the basic speed regulation is -1×2×0.2=-0.4 cm / min.

[0128] The target parameter value is determined based on a combination of historically optimal process parameters and the current welding condition. For titanium-steel composite plate welding, when the joint strength is detected to be lower than the target value and there is slight porosity on the weld surface, the system may set the target current to be 8 amperes higher than the current value. The compensation value calculation adopts a proportional adjustment mechanism, with a compensation coefficient set to 0.3, meaning 30% of the parameter deviation is used as the compensation adjustment amount. For example, if the current current is 195 amperes and the target current is 203 amperes, then the current compensation value is (203-195)×0.3=+2.4 amperes. The total current adjustment amount is the sum of the basic adjustment amount and the compensation value, i.e., +2.25+2.4=+4.65 amperes. Similarly, if the current voltage is 26 volts and the target voltage is 26.5 volts, then the voltage compensation value is (26.5-26)×0.3=+0.15 volts, and the total voltage adjustment is 0+0.15=+0.15 volts; if the current speed is 30 cm / min and the target speed is 29 cm / min, then the speed compensation value is (29-30)×0.3=-0.3 cm / min, and the total speed adjustment is -0.4-0.3=-0.7 cm / min.

[0129] Calculate the stability constraint values ​​for the total current regulation, total voltage regulation, and total speed regulation. The stability constraint value is the ratio of the absolute value of the total regulation to the product of the current parameter value and the stability constraint coefficient. The stability constraint coefficient is set according to the welding process requirements: 0.05 for current, 0.04 for voltage, and 0.06 for speed. These coefficients represent the maximum allowable change ratio for a single adjustment. Taking current as an example, the stability constraint value is calculated as the absolute value of the total current regulation divided by the product of the current value and the stability constraint coefficient, i.e., |+4.65|÷(195×0.05)=0.477. Similarly, the stability constraint value for voltage is |+0.15|÷(26×0.04)=0.144; and the stability constraint value for speed is |-0.7|÷(30×0.06)=0.389. These constraint values ​​are used to determine whether the regulation exceeds the stability limit; a value less than 1 indicates it is within the allowable range, and a value greater than 1 indicates that adjustment is required.

[0130] The total current regulation, total voltage regulation, and total speed regulation are adjusted initially according to preset proportions based on stability constraint values. When the stability constraint value is greater than 1, the total regulation is reduced proportionally to ensure it does not exceed the stability limit; when the stability constraint value is less than or equal to 1, the total regulation remains unchanged. In this example, the stability constraint values ​​for all parameters are less than 1, therefore, no initial adjustment is needed. The total current regulation remains +4.65 amps, the total voltage regulation remains +0.15 volts, and the total speed regulation remains -0.7 cm / min. To illustrate the adjustment mechanism, let's assume the total current regulation is +12 amps. Then, the stability constraint value is |+12| ÷ (195 × 0.05) = 1.23, exceeding the stability limit. In this case, the total current regulation needs to be reduced to +12 ÷ 1.23 = +9.76 amps to ensure the adjustment range remains within the stable range.

[0131] The coupling constraint value is calculated based on the coupling relationship between parameters and is used to evaluate the rationality of simultaneous adjustment of multiple parameters. Taking current and voltage as an example, the coupling constraint value is calculated as the product of the ratio of the current adjustment to the current current and the ratio of the voltage adjustment to the current voltage, and the coupling coefficient between current and voltage. If the coupling coefficient between current and voltage is set to 0.7, then the coupling constraint value between current and voltage is (4.65 / 195) ÷ (0.15 / 26) × 0.7 = 4.12. Similarly, if the coupling coefficient between current and speed is set to -0.5, the coupling constraint value between current and speed is (4.65 / 195) ÷ (-0.7 / 30) × (-0.5) = 0.51; if the coupling coefficient between voltage and speed is set to -0.3, the coupling constraint value between voltage and speed is (0.15 / 26) ÷ (-0.7 / 30) × (-0.3) = 0.07. The preset coupling threshold is 1.5, meaning that the difference in adjustment ratio between parameters should not exceed 1.5 times when considering the coupling relationship.

[0132] Based on the coupling constraint values, the total current regulation, total voltage regulation, and total speed regulation after the first adjustment are iteratively optimized using the gradient descent method. Each iteration adjusts the excess amount by 20%. In this example, the coupling constraint value between current and voltage is 4.12, exceeding the preset threshold of 1.5, requiring adjustment. The excess is calculated as 4.12 - 1.5 = 2.62, with an adjustment range of 2.62 × 20% = 0.524. The current regulation decreases to 4.65 × (1 - 0.524) = 2.21 amperes, and the voltage regulation increases to 0.15 × (1 + 0.524) = 0.23 volts. The recalculated coupling constraint value between current and voltage is (2.21 / 195) ÷ (0.23 / 26) × 0.7 = 1.28, which meets the preset threshold requirement. Checking the coupling constraint values ​​of other parameter pairs, the coupling constraint value of current and speed becomes (2.21 / 195)÷(-0.7 / 30)×(-0.5)=0.24, and the coupling constraint value of voltage and speed becomes (0.23 / 26)÷(-0.7 / 30)×(-0.3)=0.11. Both meet the preset threshold requirements, and the iterative optimization is completed.

[0133] The optimized total current, voltage, and speed adjustments are output as the optimal control strategy. The final control strategy is: current adjustment +2.21 amps, voltage adjustment +0.23 volts, and speed adjustment -0.7 cm / min. These adjustments are applied to the current welding parameters, resulting in new welding parameters: current 197.21 amps, voltage 26.23 volts, and speed 29.3 cm / min. In actual control systems, these adjustments are typically rounded to an acceptable level for the equipment, such as adjusting the current to +2.2 amps, the voltage to +0.2 volts, and the speed to -0.7 cm / min.

[0134] For specific welding conditions, the system automatically adjusts constraints. For example, in the transition region between the weld head and tail, the stability constraint coefficient is reduced accordingly; the current stability constraint coefficient may be lowered to 0.03 to ensure smoother parameter adjustments. In regions with abrupt changes in material thickness, the coupling threshold may be relaxed to 2.0, allowing for greater parameter adjustment differences to accommodate changes in material properties. For high-strength titanium alloy welding, the system automatically adjusts the base step size based on material characteristics; for example, the current base step size may be reduced to 3 amperes for finer control.

[0135] The update frequency of the optimal adjustment strategy is adaptively adjusted based on the dynamic characteristics of the welding process. During the parameter stabilization phase, the update frequency is once per second; during the rapid parameter change phase, the update frequency can be increased to five times per second to ensure the system can respond promptly to changes in operating conditions. The system also has a safety protection mechanism to limit the cumulative effect of parameter adjustments. For example, within 60 seconds, the cumulative adjustment of current must not exceed 15% of the current value, voltage must not exceed 12%, and speed must not exceed 20%, preventing excessive adjustments from causing welding instability.

[0136] The weighted modulation and multi-constraint optimization method proposed in this invention effectively solves the stability and coordination problems in the welding parameter adjustment process. It introduces a dual guarantee mechanism of stability constraints and coupling constraints to ensure a smooth and reliable adjustment process; and employs an iterative optimization strategy to handle complex interactions between parameters, avoiding adjustment conflicts. This method is highly adaptable and can automatically adjust constraints according to different welding conditions, significantly improving the welding quality and consistency of titanium-steel composite plates, and providing effective technical support for the intelligent welding of high-performance composite materials.

[0137] In a second aspect, a computer-readable storage medium is provided, having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

Claims

1. A neural network-based method for predicting the welding strength of a titanium-steel clad plate and controlling parameters, characterized by, The method comprises the steps of: acquiring welding process parameters, weld surface images and weld ultrasonic detection data in a titanium-steel clad plate welding process, and preprocessing to obtain a standardized data set; a dual-branch heterogeneous neural network prediction mechanism is constructed, wherein a first branch adopts a lightweight RNN structure with an attention mechanism to quickly extract features from the standardized data set and output a welding strength trend; a second branch adopts an improved LSTM structure to accurately model the standardized data set and output welding strength prediction data; the welding strength trend and the welding strength prediction data are adaptively fused, specifically including: calculating a time sequence consistency score based on the time sequence gradient features of the welding strength trend and the fluctuation features of the welding strength prediction data, combining the time sequence consistency score with real-time state data of the welding process parameters to generate a dynamic fusion weight; adaptively modulating the dynamic fusion weight based on historical prediction accuracy to generate an optimized fusion weight coefficient; weighting and combining the optimized fusion weight coefficient with the welding strength trend and the welding strength prediction data to obtain an initial fusion result; calculating a residual compensation value based on the deviation of the initial fusion result and historical prediction data, and adaptively correcting the residual compensation value and the initial fusion result to obtain a comprehensive prediction result; calculating the prediction error between the comprehensive prediction result and the actual welding strength data obtained in real time; based on reinforcement learning, a collaborative optimization control scheme is constructed, the welding process parameters, weld surface images and prediction error are taken as state inputs, the parameter coupling degree between welding current, welding voltage and welding speed is calculated, the control weight of welding current, welding voltage and welding speed is generated according to the parameter coupling degree, and the optimal adjustment strategy data is generated; according to the optimal adjustment strategy data, welding current, welding voltage and welding speed are adjusted in real time.

2. The method of claim 1, wherein, The step of the first branch adopting a lightweight RNN structure with an attention mechanism to quickly extract features from the standardized data set and output a welding strength trend comprises the steps of: a multi-scale time sequence feature extraction module is constructed to extract features from the standardized data set using different size sliding windows, calculate time sequence attention weights based on the current time hidden layer state and the previous time state, and adaptively weight and fuse the features extracted under different scales to obtain multi-scale fusion features; a bidirectional gated recurrent unit is used to process the multi-scale fusion features, the bidirectional gated recurrent unit calculates an adaptive gating threshold value according to the mean and standard deviation of the input data, dynamically modulates the output of the update gate and the reset gate, and obtains an optimized gating state; a forward propagation path and a backward propagation path are constructed, the forward propagation path generates a forward candidate state based on the modulation result of the optimized reset gate state and the previous time hidden layer state and the multi-scale fusion features, and the backward propagation path generates a backward candidate state based on the modulation result of the optimized reset gate state and the next time hidden layer state and the multi-scale fusion features; The forward candidate state and the reverse candidate state are connected in residual connection with the hidden layer state at the previous moment, and the result of the residual connection is selectively updated by the optimized update gate state to obtain the hidden layer state at the current moment; the hidden layer state at the current moment is extracted and fused to generate the welding strength trend data.

3. The method of claim 1, wherein, The second branch adopts an improved LSTM structure to accurately model the standardized data set, and the steps of outputting the welding strength prediction data include: The standardized data set is decomposed into features, the importance weights of the decomposed features are calculated based on a multi-layer perceptron, the importance weights are combined with the corresponding decomposed features to obtain weighted fusion features, a dynamic head number adaptive multi-head attention module is constructed, the number of attention heads is adaptively determined based on the current state of the weighted fusion features, the weighted fusion features are mapped into a query matrix, a key matrix and a value matrix, a position perception matrix is generated based on sine-cosine position encoding, the product result of the position perception matrix and the query matrix and the key matrix is combined to obtain attention weights, a channel attention enhancement module is constructed, the channel features are obtained by globally average pooling the attention weights, the weight coefficients of the channel features are calculated based on a multi-layer perceptron, and the weight coefficients are multiplied with the attention weights channel by channel to obtain enhanced features; An improved long short-term memory unit is constructed, the enhanced features are taken as inputs, the forget gate of the long short-term memory unit is adaptively adjusted based on the memory unit state at the previous moment, the input gate is adaptively adjusted based on the hyperbolic tangent transformation of the memory unit state at the previous moment, and the hidden layer state at the current moment is obtained; a dynamic coefficient is calculated based on the feature similarity of adjacent layer hidden layer states, the hidden layer states of adjacent layers are connected in residual connection using the dynamic coefficient, and the inter-layer fusion features are obtained by densely connecting the multi-layer hidden layer states; and the welding strength prediction result is generated based on the inter-layer fusion features.

4. The method of claim 1, wherein, A collaborative optimization control scheme is constructed based on reinforcement learning, the welding process parameters, the weld surface image and the prediction error are taken as state inputs, the parameter coupling degree between the welding current, the welding voltage and the welding speed is calculated, the control weights of the welding current, the welding voltage and the welding speed are generated according to the parameter coupling degree, and the steps of generating the optimal adjustment strategy data include: The welding process parameters, the weld surface image and the prediction error are combined to obtain fusion state features; the Pearson correlation coefficients between the welding current, the welding voltage and the welding speed are calculated based on the fusion state features to generate a parameter coupling degree matrix; The initial weights of each welding process parameter are calculated based on the coupling strength between the parameters in the parameter coupling degree matrix, the initial weights are combined with the weight adjustment gradient of the reward function calculated based on the welding quality index and the process economy index to obtain the optimized dynamic control weights; constructing a deep Q network structure, inputting the fusion state feature and the optimized dynamic control weight into the deep Q network structure to generate a state-action value estimation, and selecting an adjusting action based on the state-action value estimation using an ε-greedy strategy; calculating a time difference error based on an immediate reward and a maximum value estimation of a next state, and updating parameters of the deep Q network structure using the time difference error; weighting and modulating the optimized dynamic control weight and the adjusting action selected based on the deep Q network structure to generate welding current, voltage and speed adjustment amounts, and obtaining an optimal adjustment strategy.

5. The method of claim 4, wherein, The steps of calculating an initial weight of each welding process parameter based on the coupling strength between parameters in the parameter coupling degree matrix, combining the initial weight with a weight adjustment gradient of a reward function calculated based on a welding quality index and a process economy index, and obtaining an optimized dynamic control weight include: calculating a basic coupling degree between the welding current, voltage and speed using Pearson correlation coefficients; performing wavelet transform on the welding current, voltage and speed, calculating time-frequency coupling characteristics in different frequency bands based on cross wavelet transform coefficients, and calculating the causal relationship strength between the welding current, voltage and speed based on a Granger causality test method; and weighting and fusing the basic coupling degree, the time-frequency coupling characteristics and the causal relationship strength to generate a fusion coupling degree matrix; calculating the initial weight of the welding current, voltage and speed based on the fusion coupling degree matrix; calculating a mechanical property score based on the tensile strength, yield strength and elongation, and calculating a process stability score based on arc characteristic fluctuations and molten pool dynamic behavior; weighting and combining the mechanical property score and the process stability score to obtain a welding quality evaluation value; calculating a process economy evaluation value based on energy consumption and material utilization indexes; and weighting and calculating the welding quality evaluation value and the process economy evaluation value to obtain a comprehensive reward value; calculating a gradient of the comprehensive reward value on the initial weight, combining the initial weight, and introducing a momentum term of the difference between adjacent two weights to generate a dynamic weight; calculating a compensation adjustment amount based on historical adjustment effects; and combining the compensation adjustment amount and the dynamic weight to obtain an optimized dynamic control weight.

6. The method of claim 4, wherein, The steps of weighting and modulating the optimized dynamic control weight and the adjusting action selected based on the deep Q network structure to generate welding current, voltage and speed adjustment amounts, and obtaining an optimal adjustment strategy include: standardizing and multiplying the optimized dynamic control weight and the adjusting action output by the deep Q network to obtain current, voltage and speed basic adjustment amounts; calculating a compensation value based on the deviation between a target parameter value and a current parameter value, and combining the compensation value and the corresponding basic adjustment amount to obtain current, voltage and speed total adjustment amounts; The stability constraint value of the total current regulation amount, the total voltage regulation amount and the total speed regulation amount is calculated, which is the ratio of the absolute value of the total regulation amount and the product of the current parameter value and the stability constraint coefficient; the first adjustment is made on the total current regulation amount, the total voltage regulation amount and the total speed regulation amount according to the preset proportion according to the stability constraint value; the coupling constraint value between the total current regulation amount, the total voltage regulation amount and the total speed regulation amount after the first adjustment is calculated; the iterative optimization is made on the total current regulation amount, the total voltage regulation amount and the total speed regulation amount after the first adjustment according to the coupling constraint value until the coupling constraint value meets the preset coupling threshold value; and the optimized total current regulation amount, the total voltage regulation amount and the total speed regulation amount are output as the optimal regulation strategy.

7. A computer-readable storage medium having stored thereon computer program instructions, wherein, The computer program instructions are executed by the processor to implement the method in any one of claims 1 to 6.

Images