Hull fillet welding and butt welding oriented geometric dimension control and parameter optimization method

Abstract

The invention belongs to the technical field of ship structure manufacturing and welding processes, and discloses a geometric dimension control and parameter optimization method for ship body angle welding and butt welding, which comprises the following steps: S1, constructing a classified welding process database and carrying out data preprocessing; s2, constructing and training a neural network forward prediction model special for a welding seam; s3, the process parameters are intelligently and optimally adjusted through the neural network forward prediction model special for the welding seam, and an optimal process parameter combination is output; and S4, performing experimental verification on the optimal process parameter combination output in the step S3, and performing closed-loop evolution on the optimized fillet weld time sequence enhanced neural network and butt weld time sequence enhanced neural network. According to the method, the design variable, which directly determines the material consumption, of the geometric dimension of the welding seam is used as an accurate control target, and direct and active control over the welding cost is achieved on the premise that the welding quality is guaranteed through intelligent reverse optimal adjustment of the technological parameters.

Description

Technical Field

[0001] This invention relates to the field of ship structure manufacturing and welding technology, and in particular to a method for geometric dimension control and parameter optimization for fillet welds and butt welds on ship hulls. Background Technology

[0002] Welding is a key process in ship hull structure forming, with fillet welds and butt welds being the two most widely used and load-bearing joint types. The weld leg size of fillet welds (such as the connection between ribs, rib plates, and shell plates) directly determines the static load and fatigue strength of the joint; the reinforcement of butt welds (such as the splicing of decks and bottom outer plates) significantly affects their stress concentration and fatigue performance. During the ship design phase, the weld geometry determined based on mechanical calculations and specifications not only ensures structural safety but also directly relates to material consumption and costs during manufacturing. However, how to accurately and stably achieve these design dimensions in the manufacturing process and efficiently and scientifically correlate them with welding process parameters has always been a key challenge in the field of intelligent ship manufacturing.

[0003] Currently, research and optimization of welding processes mainly proceed in two directions: First, numerical simulation based on physical mechanisms, such as the finite element method and computational fluid dynamics, is used to deeply analyze the thermo-mechanical-metallurgical behavior during welding. While these methods can reveal the underlying mechanisms, the models are complex, computationally expensive, and heavily reliant on precise physical property parameters, making them unsuitable for rapid response and online optimization process design scenarios. Second, data-driven methods, especially utilizing machine learning algorithms (such as neural networks and support vector machines) to establish "forward prediction models" that predict weld quality or morphology from welding process parameters (current, voltage, speed, etc.). This type of research has made significant progress and can effectively answer the question "Given a set of parameters, what kind of weld will be obtained?", becoming an effective process analysis and auxiliary tool.

[0004] However, there is a significant limitation and gap in the existing technology system: the objective function of most optimization studies focuses on improving the mechanical properties of the weld joint itself, reducing defects, or improving welding efficiency, but there are few systematic efforts to minimize costs by taking the "weld geometry"—a variable that directly determines the amount of material used—as the core control object. Specifically, this is manifested in the following ways: (1) Most existing data-driven models are "forward" predictions and cannot directly solve the core reverse engineering problem of "how to set process parameters to achieve a certain target size"; (2) Even if a few studies attempt to combine forward models with optimization algorithms, their optimization objectives usually point to tensile strength, penetration depth, or welding deformation, etc., and have not established an optimization model with the direct objective of "precisely controlling the weld leg size or reinforcement height to achieve welding material saving". This means that the current intelligent methods have failed to effectively solve the phenomenon of "over-welding" caused by improper process parameters in shipbuilding—that is, the actual weld size far exceeds the minimum design requirements, resulting in a large waste of welding materials and energy.

[0005] Furthermore, in the upstream design stage of the welding process, although it is possible to collaboratively optimize the "target material-saving size" that can theoretically ensure safety and save materials based on the principles of mechanics and economics, when this size information is transmitted to the manufacturing stage, due to the lack of accurate and reliable reverse mapping methods for process parameters, it is often forced to adopt conservative empirical parameters with margins for welding. As a result, "material-saving design" cannot be implemented into "material-saving manufacturing", and the economic optimization goal is lost in the final product. Summary of the Invention

[0006] To address the problem of actual weld dimensions far exceeding the minimum design requirements in existing technologies, resulting in significant waste of welding materials and energy, this invention proposes a geometric dimension control and parameter optimization method for fillet welds and butt welds in ship hulls. Specifically targeting the most common fillet welds and butt welds in ship component manufacturing, it provides a method and system that can reverse and accurately map target geometric dimensions (weld leg size, reinforcement height) based on economic trade-offs to optimal welding process parameters (such as current, voltage, speed, etc.). The aim is to control the fillet weld leg size and butt weld reinforcement height to achieve material savings and cost reduction while ensuring joint strength.

[0007] This invention is achieved through the following technical solution, including the following steps:

[0008] S1. Construct a classified welding process database and perform data preprocessing;

[0009] S2. Construct and train a forward prediction model for a weld-specific neural network to obtain optimized time-series enhanced neural networks for fillet welds and butt welds;

[0010] S3. The optimized time-enhanced neural network for fillet welds and butt welds is used to intelligently adjust the process parameters and output the optimal combination of process parameters.

[0011] S4. Experimentally verify the optimal combination of process parameters output in step S3, and perform closed-loop evolution on the optimized fillet weld time-enhanced neural network and butt weld time-enhanced neural network.

[0012] As a further preferred option, the specific steps of step S2 are as follows:

[0013] S2.1. Using a sliding window, extract the time-domain features within each window of the welding current I and arc voltage U to construct the fillet weld input feature vector. Or input feature vector of butt weld ;

[0014] S2.2 Construct timing-enhanced neural networks for fillet welds and butt welds respectively. Both types of neural networks include:

[0015] Input layer: The number of nodes corresponds to the dimension of their respective input feature vectors;

[0016] Temporal feature extraction layer: employs a gated recurrent unit layer;

[0017] Attention layer: Connects to a multi-head self-attention mechanism module to weight and focus the temporal state output by the GRU;

[0018] Fully connected hidden layer: Employs a two-layer fully connected structure using the Swish activation function;

[0019] Output layer: Single-node linear output, predicting solder pad size K or margin height H.

[0020] Regularization: DropConnect is dropped after the GRU layer using random connections, and Gaussian weighted noise is injected between fully connected layers;

[0021] S2.3. Train and optimize the time-series enhancement neural network for fillet welds and the time-series enhancement neural network for butt welds constructed in step S2.2 to obtain the optimized time-series enhancement neural network for fillet welds and the time-series enhancement neural network for butt welds.

[0022] As a further preferred option, the specific steps of step S3 are as follows:

[0023] S3.1 Construct an objective function based on minimizing the weighted error between the predicted size and the target size, as shown in the following formula:

[0024]

[0025] Where λ is the penalty coefficient; Penalty(X) is the soft penalty term for parameters approaching the constraint boundary; Y target Indicates the target size; f model (X) represents the predicted size;

[0026] S3.2. The objective function constructed in step S3.1 is solved using the hybrid improved particle swarm optimization-simulated annealing algorithm, and the optimal combination of process parameters is output.

[0027] This invention also provides an optimization system based on the geometric dimension control and parameter optimization method for fillet and butt welds of ship hulls described in this invention, including...

[0028] The database management module is used to build and maintain the dedicated welding process database;

[0029] The timing feature extraction module, integrated into the database management module, is responsible for real-time feature calculation and storage of timing signals during the welding process;

[0030] The intelligent optimization and solution module is used to construct the process parameter optimization model and call the optimization algorithm to solve it.

[0031] Hybrid optimization solver: Integrated into the intelligent optimization solver module, it has a built-in hybrid improved particle swarm optimization-simulated annealing algorithm and its adaptive parameter adjustment mechanism;

[0032] The validation and evolution module is used to manage the experimental validation process and drive the closed-loop update of the database and model;

[0033] Online learning controller: Integrated into the validation and evolution module, it is responsible for monitoring production errors, triggering online fine-tuning, and managing incremental learning processes.

[0034] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0035] 1. This invention is the first to use the "weld geometry"—a design variable that directly determines material consumption—as a precise control target. By intelligently reverse-engineering process parameters, it achieves direct and proactive control over welding costs (mainly reflected in welding material consumption) while ensuring welding quality.

[0036] 2. This invention effectively prevents "over-welding" by precisely achieving optimized target weld dimensions. It transforms cost control from vague, experience-based management to precise, numerical quantification, directly reducing shipbuilding costs.

[0037] 3. This invention constructs a complete closed-loop workflow that includes data acquisition, model training, parameter tuning, experimental verification, and knowledge feedback. It has strong self-learning and adaptive evolution capabilities, can continuously optimize cost control accuracy, and continuously enhances the model prediction and parameter tuning capabilities over time through continuous iteration. It also has robustness in dealing with the influence of variables such as material batches and equipment status.

[0038] 4. The technical solution of this invention is highly targeted and has significant engineering practical value. The solution focuses on the most critical and material-intensive fillet welds and butt welds in shipbuilding, and is tied to AH36 / EH36 high-strength steel and mainstream welding methods. All input and output variables directly correspond to actual engineering practices, and the optimization results can directly generate welding process specifications with a clear implementation path. It has clear economic benefits in saving welding materials and reducing energy consumption, and is of great significance for promoting the intelligent upgrading and green transformation of the shipbuilding industry. Attached Figure Description

[0039] Figure 1 This is a schematic diagram of the overall process of the optimization method of the present invention.

[0040] Figure 2 This is a schematic diagram of the weld leg in the welding process.

[0041] Figure 3 This is a schematic diagram of the weld reinforcement in the welding process. Detailed Implementation

[0042] The advantages and features of the present invention will be illustrated and explained by the following non-limiting description of preferred embodiments, which are given by way of example only with reference to the accompanying drawings.

[0043] This invention proposes a process parameter optimization method integrating the core ideas of "classification modeling, intelligent reverse engineering, and closed-loop evolution." The overall concept is as follows: First, for different welding methods and output dimensions of fillet welds and butt welds, an independent high-precision data-driven forward prediction model is constructed as a "digital twin" for understanding the process-dimensional relationship; second, the inverse problem of parameter optimization is transformed into a mathematical optimization problem with the forward model as the evaluation function and the process parameters as the decision variables, and a metaheuristic algorithm is used for efficient search; finally, through physical verification and data feedback, a continuously iterative intelligent closed-loop system is formed. Before describing the detailed steps of this invention, the core geometric dimension concept of fillet welds in the welding process is first explained, such as... Figure 2 As shown, this represents the weld leg in a welding process. The weld leg is the minimum distance from the weld toe on one right-angled face to another in the cross-section of a fillet weld; as... Figure 3 As shown, the weld reinforcement in the welding process refers to the maximum height of the weld metal that extends beyond the weld toe line on the base material surface.

[0044] like Figure 1 As shown, this invention provides a method for geometric dimension control and parameter optimization for fillet welds and butt welds on ship hulls, with the specific steps as follows:

[0045] S1. Construct a classified welding process database and perform data preprocessing;

[0046] This step forms the basis for all subsequent data-driven models. Its core lies in classifying and collecting data for different weld types and implementing strict quality control.

[0047] S1.1 Divide the welding scenarios into fillet weld scenarios and butt weld scenarios according to the actual working conditions, and collect the corresponding data;

[0048] The CO2 gas shielded welding (Gas Metal Arc Welding, abbreviated as GMAW-CO2) process for marine AH36 / EH36 grade high-strength steel, with ER50-6 solid welding wire (commonly with a diameter of 1.2mm) as the fillet weld scenario, was used to collect data on key conditions such as base metal thickness and welding position.

[0049] The submerged arc welding (SAW) process for marine AH36 / EH36 high-strength steel, using H10Mn2 welding wire (commonly 4.0mm in diameter) and SJ101 sintered flux, was classified as a butt weld scenario, and parameters for long weld splicing conditions in straight positions were collected.

[0050] The data sources for this invention can be multi-source data, including: a) laboratory controlled process test data; b) historical process data and result data automatically recorded by welding robots or digital welding machines; c) weld sample data collected on the production site under strict quality monitoring and judged to be excellent.

[0051] S1.2, Construct input and output variables;

[0052] Input variables (i.e., process parameters):

[0053] For fillet weld scenarios, at least the following should be included: welding current I, in amperes (A); arc voltage U, in volts (V); welding speed V, in centimeters per minute (cm / min); and shielding gas flow rate G, in liters per minute (L / min).

[0054] For butt weld scenarios, at least the welding current I, in amperes (A), the arc voltage U, in volts (V), and the welding speed V, in centimeters per minute (cm / min) should be included.

[0055] Output variables (i.e., process parameters):

[0056] Fillet weld scenario: Weld leg size K, in millimeters (mm), is defined as the length of the straight side of the isosceles right triangle in the weld cross section.

[0057] In the case of butt welds: the weld height H, in millimeters (mm), is defined as the maximum height of the weld metal above the surface of the base material.

[0058] All geometric dimensions must be obtained using high-precision measuring equipment to ensure that the data is objective and accurate.

[0059] S1.3, Data preprocessing;

[0060] S1.3.1 Outlier Cleaning: Apply the 3σ criterion (Raida criterion) individually to each process parameter and dimensional variable. Calculate the mean μ and standard deviation σ of the variable, and remove all data points outside the range of (μ-3σ, μ+3σ) to eliminate interference from recording errors or extreme abnormal operating conditions.

[0061] S1.3.2 Data Standardization: All numerical variables are processed using the Z-score standardization method. The formula is: , where x is the original value and Z represents the standardized value. This step aims to eliminate the influence of parameters with different dimensions such as current, voltage, and speed on model training, accelerate neural network convergence, and improve numerical stability.

[0062] S1.3.3 Dataset Partitioning: The preprocessed complete dataset is randomly divided into a training set, a validation set, and a test set in a 7:2:1 ratio. The training set is used for model parameter learning, the validation set is used for hyperparameter tuning and training process monitoring, and the test set is used for objective performance evaluation of the final model.

[0063] S1.3.4, Timing Feature Engineering: For timing signals such as welding current and arc voltage, while storing the original values, calculate their sliding statistical features (window of 1 second, step size of 0.2 seconds), including mean, standard deviation, peak-to-peak value, and waveform factor, as high-dimensional input features for subsequent modeling.

[0064] S1.3.5 Multi-rule anomaly cleaning: Based on the 3σ criterion, a rule base based on welding process knowledge is introduced for joint cleaning:

[0065] (1) Current-voltage matching rule: If (where k is an empirical constant for a specific welding wire-gas combination, and δ is the allowable deviation), then it is marked as abnormal.

[0066] (2) Speed ​​change rule: If the welding speed changes by more than 20% between adjacent sampling points, it is considered abnormal.

[0067] (3) Physical limit rule: Data that exceeds the theoretical limit value of the base material thickness for weld leg size or reinforcement height shall be rejected.

[0068] S2. Construct and train a forward prediction model for a weld-specific neural network to obtain optimized time-series enhanced neural networks for fillet welds and butt welds;

[0069] This step aims to build a specialized predictive model that can accurately characterize the complex nonlinear and time-dependent relationships between welding process parameters and weld dimensions.

[0070] S2.1. A sliding window method is used to extract the time-domain features of the welding current I and arc voltage U within each window, including the mean μ, standard deviation σ, peak-to-peak value P−P, and waveform factor. Construct the input feature vector of the fillet weld. Or input feature vector of butt weld , where X rms Represents the root mean square value; Represents the absolute average value of a time-domain signal. This represents the average value of the welding current I within the window; This represents the standard deviation of the welding current I within the window; This represents the peak-to-peak value of the welding current I within the window; This represents the waveform factor of the welding current I within the window; This represents the average value of the arc voltage U within the window; This represents the standard deviation of the arc voltage U within the window; This represents the peak-to-peak value of the arc voltage U within the window; This represents the waveform factor of the arc voltage U within the window.

[0071] The sliding window has a window length of 1 second and a step size of 0.2 seconds.

[0072] S2.2 Construct timing-enhanced neural networks for fillet welds and butt welds respectively. Both types of neural networks include:

[0073] Input layer: The number of nodes corresponds to the dimension of their respective input feature vectors.

[0074] Temporal feature extraction layer: A gated recurrent unit (GRU) layer (32 hidden units) is used to model the short-term and long-term dependencies of parameter fluctuations during the welding process.

[0075] Attention layer: Connects to a multi-head self-attention mechanism module (4 heads) to weight and focus the timing state output by the GRU, and identify key process stages that have a significant impact on weld size.

[0076] Fully connected hidden layer: Employs a two-layer fully connected structure, using the Swish activation function ( , This provides a smoother and non-monotonic nonlinear fitting capability.

[0077] Output layer: Single-node linear output, predicting solder pad size K or margin height H.

[0078] Regularization: Random connection drop (DropConnect) (drop rate 0.2) is used after the GRU layer, and Gaussian weighted noise (standard deviation 0.01) is injected between fully connected layers.

[0079] S2.3. Train and optimize the time-series enhancement neural network for fillet welds and the time-series enhancement neural network for butt welds constructed in step S2.2 to obtain the optimized time-series enhancement neural network for fillet welds and the time-series enhancement neural network for butt welds.

[0080] S2.3.1 Construct the loss function, as shown in the following formula:

[0081] ;

[0082] In the formula, y represents the Huber loss threshold parameter, set to 0.1; y represents the actual weld size, i.e., the actual value of the weld leg size K or the excess height H obtained by measurement in the training data. This represents the average Huber loss function value, a regularized loss function used to train neural networks that is insensitive to outliers. This represents the weld size predicted by the model, i.e., the output value f of the neural network forward prediction model. model (X).

[0083] The temporal augmentation neural network constructed in this invention employs Huber loss, combining the advantages of mean squared error and mean absolute error to enhance robustness to potential outliers in the training data.

[0084] S2.3.2. Use the Nadam optimizer (Nesterov accelerated gradient method Adam), set the initial learning rate to 0.001, and use cosine annealing learning rate scheduling, with a minimum learning rate of 1e-5.

[0085] S2.3.3: Monitor the validation set loss during the training process. If the loss does not decrease for 10 consecutive epochs, trigger early stopping and automatically save the model with the minimum validation loss.

[0086] S2.3.4. The time-series enhancement neural network for fillet welds and butt welds constructed in step S2.2 must satisfy the following on the independent test set: Mean Absolute Error (MAE) < 0.1 mm, and Coefficient of Determination (R²) < 0.1 mm. 2After reaching >0.98, the optimized timing enhancement neural network for fillet welds and butt welds was obtained.

[0087] S3. The optimized timing-enhanced neural network for fillet welds and butt welds is used to intelligently adjust the process parameters and output the optimal combination of process parameters.

[0088] This step is the core of reverse design, transforming the question of "how to weld" into a mathematical optimization problem. It should be noted that the "target weld core geometry" input to the method described in this invention refers to the economically optimal dimension determined by the structural design and cost analysis process, while satisfying all strength and performance specifications. The core value of this method lies in using the "material-saving target" from the design as input and intelligently reverse-engineering to output executable process parameters. The idea behind this step is to formalize the problem; when the user inputs a target material-saving dimension Y determined based on the co-optimization of strength and economy... target The system automatically identifies the weld type and calls the corresponding positive prediction model f. model (X). The decision variable is the welding process parameter combination X. For fillet welds, For butt welds, The specific steps are as follows:

[0089] S3.1 Construct an objective function based on minimizing the weighted error between the predicted size and the target size, as shown in the following formula:

[0090]

[0091] Where λ is the penalty coefficient; Penalty(X) is the soft penalty term for parameters approaching the constraint boundary; Y target Indicates the target size; f model (X) represents the predicted size of the optimized fillet weld time-enhanced neural network and the butt weld time-enhanced neural network outputs.

[0092] Constraints: Each process parameter x i It must be within the scope of its physical feasibility, equipment availability, and process specifications: x i，min ≤x i ≤x i，max x i，min and x i，max These represent the physical feasibility, equipment availability, and lower and upper limits specified in the process specifications for process parameters of type i, respectively.

[0093] S3.2. The objective function constructed in step S3.1 is solved using a hybrid improved particle swarm optimization-simulated annealing algorithm, and the optimal combination of process parameters is output. The steps are as follows:

[0094] a. Population initialization: 30 initial particles are generated within the parameter feasible region using Latin hypercube sampling to ensure uniform spatial distribution.

[0095] b. Adaptive update:

[0096] Inertial weight ω decays nonlinearly:

[0097] Learning factors are dynamically adjusted: early stage focuses on exploration (larger C1), later stage focuses on development (larger C2).

[0098]

[0099] In the formula, T max The total number of iterations is represented by t; the current iteration number is represented by t.

[0100] c. Constraint handling: If a particle goes out of bounds, the boundary reflection method is used to make it oscillate within the feasible region, rather than simply truncation.

[0101] d. Hybrid Local Search: Every 5 iterations, simulated annealing local search is performed on the current globally optimal particle, and the probability of accepting inferior solutions decays exponentially.

[0102] e. Convergence criterion: When the global optimal fitness value changes by less than 10 over 10 consecutive generations. -3 It will terminate when the maximum number of iterations, 150, is reached.

[0103] f. Output: The globally optimal particle position is the recommended optimal combination of process parameters X. opt .

[0104] S4. Experimentally verify the optimal combination of process parameters output in step S3, and perform closed-loop evolution on the optimized fillet weld time-enhanced neural network and butt weld time-enhanced neural network.

[0105] This step ensures the reliability of the optimization results and gives the system the ability to learn continuously.

[0106] S4.1 Physical Experiment Verification:

[0107] Specimen preparation: Strictly follow the optimal parameter combination X output in step S3. opt Standard welding specimens were prepared under the same conditions as those for data acquisition, including base material, welding materials, and equipment.

[0108] Precision measurement: Non-contact scanning of the cooled weld seam is performed using equipment such as a 3D optical scanner to acquire high-density point cloud data. The 3D morphology of the weld seam is reconstructed using specialized software, and the actual weld leg size K is accurately extracted. actual Or actual remaining height H actual .

[0109] Accuracy assessment and quality inspection: Calculate the absolute and relative errors between the actual and target dimensions. Simultaneously, macroscopic metallographic inspection and hardness testing can be performed to comprehensively evaluate the weld formation quality.

[0110] S4.2 If the verification result meets the preset accuracy requirements, such as the dimensional error being within ±0.1mm, then the successful [process parameters, measured dimensions] data pair will be added as a new high-quality sample to the corresponding fillet weld or butt weld database in S1.

[0111] S4.3 When the update trigger mechanism is met, the mechanism is triggered, and step S2 is re-executed using the latest database to perform full retraining or incremental learning on the corresponding neural network positive prediction model.

[0112] The system of this invention also includes a model update triggering mechanism, for example, when the cumulative number of newly added valid samples in a certain type of weld database reaches 15% of the original training set size, or when it is triggered periodically.

[0113] Once triggered, the system automatically starts the model update process: using the expanded database, step S2 is re-executed to perform full retraining or incremental learning on the corresponding neural network positive prediction model.

[0114] Incremental learning and knowledge consolidation: When a model update is triggered, an elastic weight consolidation algorithm is used for incremental training. This algorithm calculates the importance matrix of old task parameters and imposes constraints on important parameters during fine-tuning, effectively mitigating the problem of catastrophic forgetting and enabling the model to maintain its predictive ability for historical data while absorbing new knowledge.

[0115] Online adaptive fine-tuning: After system deployment, prediction errors are continuously monitored. If the average error of the actual size of three consecutive production batches exceeds a threshold (e.g., 0.15mm), the system automatically enters online learning mode. It uses recent production data to form a small batch dataset and rapidly fine-tunes the model's output layer and last hidden layer with a low learning rate (1 / 10 of the initial learning rate) to achieve adaptive optimization in the production environment.

[0116] This invention also provides an optimization system based on a geometric dimension control and parameter optimization method for fillet welds and butt welds on ship hulls, including...

[0117] The database management module is used to build and maintain the dedicated welding process database;

[0118] The timing feature extraction module, integrated into the database management module, is responsible for real-time feature calculation and storage of timing signals during the welding process;

[0119] The intelligent optimization and solution module is used to construct the process parameter optimization model and call the optimization algorithm to solve it.

[0120] Hybrid optimization solver: Integrated into the intelligent optimization solver module, it has a built-in hybrid improved particle swarm optimization-simulated annealing algorithm and its adaptive parameter adjustment mechanism.

[0121] The validation and evolution module is used to manage the experimental validation process and drive the closed-loop updates of the database and model.

[0122] Online learning controller: Integrated into the validation and evolution module, it is responsible for monitoring production errors, triggering online fine-tuning, and managing incremental learning processes.

[0123] This invention establishes a complete intelligent closed loop of perception, decision-making, execution, and learning through a cycle of "target input → intelligent optimization → experimental verification → data feedback → model update." Each cycle enriches the system's knowledge base, improves the accuracy of the prediction model, and enhances its optimization capabilities, thereby achieving long-term, autonomous performance evolution and better adapting to subtle changes in the production environment.

[0124] Example 1

[0125] The optimization method of this invention will be explained below for specific welding scenarios and in conjunction with specific working parameters:

[0126] Optimization of fillet weld size control parameters for hull ribs

[0127] Application Scenario: During the construction of hull side sections, a shipyard needs to weld EH36 high-strength steel ribs (12mm thick) using fillet welds. Based on the finite element stress analysis of this node and the company's welding material cost model, after strength-economic co-optimization, the optimal weld leg size was determined to be K=6.0mm. The welding method is CO2 gas shielded welding.

[0128] S1. Data Construction and Preprocessing

[0129] 1. Data Acquisition: Retrieve historical welding data records for this type of joint from the factory's Manufacturing Execution System (MES) and laboratory database, totaling 300 sets. The welding wire used is ER50-6 (φ1.2mm), and the shielding gas is pure CO2.

[0130] 2. Data cleaning: The inspection of the data revealed that 2 sets of current records were 0 (mistakenly recorded because the equipment was not started) and 3 sets of weld leg dimensions were obviously abnormal (>12mm, measurement error). They were removed according to the 3σ criterion.

[0131] 3. Variables and Range: 295 sets of valid data. The process parameters are approximately: welding current I (180-280A), arc voltage U (24-32V), welding speed V (25-45 cm / min), and gas flow rate G (18-22 L / min). The output is the measured weld leg size K.

[0132] 4. Standardization and Splitting: Z-score standardization was performed on I, U, V, G, and K respectively. Then, the 295 sets of data were randomly shuffled and divided into three sets in a 7:2:1 ratio: 206 sets for training, 59 sets for validation, and 30 sets for testing.

[0133] S2. Construct and train a neural network-specific positive prediction model for weld seams;

[0134] 1. Model Construction: A fully connected neural network was constructed using the PyTorch framework, with a 4-16-8-1 structure. This means the input layer has 4 nodes, the first hidden layer has 16 ReLU neurons, the second hidden layer has 8 ReLU neurons, and the output layer has 1 linear neuron. During training, Dropout was applied to the outputs of both hidden layers with a dropout rate of 0.1.

[0135] 2. Training configuration: Loss function is MSE, optimizer is Adam (lr=0.001, betas=(0.9, 0.999)). Batch size is set to 32, maximum training epochs are 800. Early stopping strategy is enabled, patience value is set to 30.

[0136] 3. Model Performance: Training stops early after approximately 350 epochs. Performance is evaluated on the test set:

[0137] The mean absolute error (MAE) for predicting solder leg size is 0.08 mm.

[0138] The coefficient of determination (R²) between predicted and actual values 2 The value is 0.97.

[0139] The performance far exceeded the preset standards (MAE<0.15mm, R²>0.95), the model verification was successful, and it was deployed to the optimization system.

[0140] S3. The optimized time-enhanced neural network for fillet welds and butt welds is used to intelligently adjust the process parameters and output the optimal combination of process parameters.

[0141] Input: User-defined target solder pad size K determined through the above collaborative optimization. target =6.0mm.

[0142] S3.1 Optimize model construction:

[0143] Decision variables .

[0144] Objective function:

[0145] Constraints: 180≤I≤280, 24≤U≤32, 25≤V≤45, 18≤G≤22.

[0146] S3.2, PSO solution:

[0147] Algorithm parameters: number of particles = 30, maximum iterations = 120, inertia weight ω decreases linearly from 0.9 to 0.4, learning factors c1 = c2 = 1.5.

[0148] Solution process: Initialize the particle swarm and perform iterative calculations. The fitness of each particle is obtained by calculating the predicted K value using the deployed fillet weld forward model. The algorithm converges after 68 generations, with the global optimal fitness value decreasing to 0.004.

[0149] Optimized output: The system recommends the following optimal combination of process parameters: welding current I=238A, arc voltage U=28.2V, welding speed V=38cm / min, and shielding gas flow rate G=20L / min.

[0150] S4. Experimentally verify the optimal combination of process parameters output in step S3, and perform closed-loop evolution on the optimized fillet weld time-enhanced neural network and butt weld time-enhanced neural network.

[0151] S4.1 Physical Experiment Verification

[0152] 1. Physical welding: CO2 gas shielded fillet welding was performed on the same EH36 steel plate (12mm) using recommended parameters, and three sets of parallel specimens were prepared.

[0153] 2. Precision Measurement: A 3D scanner was used to scan the weld. The cross-sectional point cloud was analyzed using software. The weld leg dimensions were measured as 5.93 mm, 5.97 mm, and 5.95 mm on the three sets of specimens, respectively. The average value was taken as K. actual =5.95mm.

[0154] 3. Results Analysis: The absolute error between the actual average size and the target size is 0.05 mm, which is far superior to the control level of ±0.5 mm typically achieved by traditional methods. The weld appearance is uniform, and the macroscopic metallographic display shows good fusion.

[0155] Data backflow and system evolution:

[0156] S4.2. Add these three sets of successful experimental data (parameters and measured dimensions) as three new samples to the fillet weld database. The database sample size increases from 295 sets to 298 sets.

[0157] S4.3 The system's accumulated trigger threshold for reflow data is set to "Add 50 sets". After this operation, the add counter is 3. As subsequent production verification data continues to accumulate and reaches 50 sets, the system will automatically trigger the retraining of the fillet weld forward model, using all 348 sets of data to generate a more powerful new model, thus evolving the optimization capabilities.

[0158] Example 2

[0159] Optimization of control parameters for butt weld reinforcement of marine AH36 steel plates

[0160] Application Scenario: A shipyard needs to butt-weld AH36 high-strength steel plates (20mm thick) during the construction of the ship's bottom sections. Based on fatigue strength assessment and manufacturing cost analysis, the optimal weld reinforcement height is determined to be H=2.0mm.

[0161] S1. Data Construction and Preprocessing

[0162] 1. Data Acquisition: Collect submerged arc welding process data and inspection records for the production line over the past two years, totaling 280 sets. The welding wire is H10Mn2 (φ4.0mm), and the flux is SJ101.

[0163] 2. Data Cleaning and Variables: 275 sets of valid data were obtained after cleaning. Process parameter range: welding current I (580-700A), arc voltage U (32-36V), welding speed V (40-60 cm / min). The output is the measured weld height H.

[0164] 3. Standardization and partitioning: Z-score standardization was performed on I, U, V, and H, and the training set (192 sets), validation set (55 sets), and test set (28 sets) were partitioned in a 7:2:1 ratio.

[0165] S2. Construct and train a neural network-specific positive prediction model for weld seams;

[0166] 1. Model Construction: Construct a fully connected neural network with a 3-14-6-1 structure. The input layer has 3 nodes, the first hidden layer has 14 ReLU neurons, the second hidden layer has 6 ReLU neurons, and the output layer has 1 linear neuron.

[0167] 2. Training and Performance: A similar training strategy as in Example 1 was adopted. The final model was evaluated on the test set: the MAE of the predicted residual height was 0.07 mm, and R... 2 It achieves an accuracy of 0.98, demonstrating excellent precision.

[0168] S3. The optimized time-enhanced neural network for fillet welds and butt welds is used to intelligently adjust the process parameters and output the optimal combination of process parameters.

[0169] Problem Input: Set the above-mentioned economically optimal objective residual height H target =2.0mm.

[0170] S3.1 Optimization Solution: Construct an optimization problem with the forward model of the butt weld as the core, and the decision variables are [I, U, V]. The PSO algorithm (with parameters finely tuned as in Example 1) is used for solution.

[0171] S3.2 Optimization Output: After the algorithm converges, the system recommends the following optimal process parameters: welding current I = 642A, arc voltage U = 34.1V, and welding speed V = 52cm / min.

[0172] S4. Experimentally verify the optimal combination of process parameters output in step S3, and perform closed-loop evolution on the optimized fillet weld time-enhanced neural network and butt weld time-enhanced neural network.

[0173] S4.1 Physical Experiment Verification

[0174] Physical welding and measurement: Submerged arc welding specimens were prepared according to recommended parameters. Three-dimensional scanning measurement was used to obtain H... actual =1.96mm.

[0175] S4.2 Results and Reflow: The error is only 0.04mm, fully meeting the high standard requirements. This data is used for input into the butt weld database.

[0176] S4.3 System Evolution: Similarly, this data will be included in the new samples, contributing to future model updates and continuous improvement of the accuracy of butt weld parameter optimization.

[0177] Overall Effects and Prospects

[0178] The above embodiments fully verify the effectiveness of the method of the present invention. It not only provides high-precision parameter tuning results for different weld types, but its built-in closed-loop evolution mechanism also ensures the long-term viability of the system. With its widespread application in actual shipyard production, the system's accumulated "process-dimension" mapping knowledge will become increasingly rich, ultimately expected to develop into a core intelligent decision support platform for welding processes, providing solid technical support for improving quality, efficiency, and cost reduction in shipbuilding. Simultaneously, this methodology can also be extended to the process optimization of other materials (such as aluminum alloys and stainless steel) or other welding methods (such as laser welding and friction stir welding), showing broad application prospects.

[0179] In addition to the above embodiments, the present invention may have other implementation methods. All technical solutions formed by equivalent substitution or equivalent transformation fall within the protection scope claimed by the present invention.

Claims

1. A method for geometric dimension control and parameter optimization for fillet and butt welds on ship hulls, characterized by: It includes the following steps: S1. Construct a classified welding process database and perform data preprocessing; S2. Construct and train a forward prediction model for a weld-specific neural network to obtain optimized time-series enhanced neural networks for fillet welds and butt welds; S3. The optimized time-enhanced neural network for fillet welds and butt welds is used to intelligently adjust the process parameters and output the optimal combination of process parameters. S4. Experimentally verify the optimal combination of process parameters output in step S3, and perform closed-loop evolution on the optimized fillet weld time-enhanced neural network and butt weld time-enhanced neural network.

2. The method for geometric dimension control and parameter optimization for fillet and butt welds of ship hulls according to claim 1, characterized in that: The specific steps of step S1 are as follows: S1.1 Divide the welding scenarios into fillet weld scenarios and butt weld scenarios according to the actual working conditions, and collect the corresponding data; S1.2, Construct input and output variables; S1.3 Data preprocessing.

3. The method for geometric dimension control and parameter optimization for fillet and butt welds of ship hulls according to claim 2, characterized in that: The specific steps of step S2 are as follows: S2.

1. Using a sliding window, extract the time-domain features within each window of the welding current I and arc voltage U to construct the fillet weld input feature vector. Or input feature vector of butt weld ; S2.2 Construct timing-enhanced neural networks for fillet welds and butt welds respectively. Both types of neural networks include: Input layer: The number of nodes corresponds to the dimension of their respective input feature vectors; Temporal feature extraction layer: employs a gated recurrent unit layer; Attention layer: Connects to a multi-head self-attention mechanism module to weight and focus the temporal state output by the GRU; Fully connected hidden layer: Employs a two-layer fully connected structure using the Swish activation function; Output layer: Single-node linear output, predicting solder pad size K or margin height H; Regularization: DropConnect is dropped after the GRU layer using random connections, and Gaussian weighted noise is injected between fully connected layers; S2.

3. Train and optimize the time-series enhancement neural network for fillet welds and the time-series enhancement neural network for butt welds constructed in step S2.2 to obtain the optimized time-series enhancement neural network for fillet welds and the time-series enhancement neural network for butt welds.

4. The method for geometric dimension control and parameter optimization for fillet and butt welds of ship hulls according to claim 3, characterized in that: The specific steps of step S2.3 are as follows: S2.3.1 Construct the loss function, as shown in the following formula: ； In the formula, This represents the Huber loss threshold parameter; y represents the actual weld size. This represents the average Huber loss function value; This indicates the weld size predicted by the model; S2.3.

2. Use the Nadam optimizer, set the initial learning rate to 0.001, and use cosine annealing learning rate scheduling with a minimum learning rate of 1e-5. S2.3.3: Monitor the validation set loss during the training process. If the loss does not decrease for 10 consecutive epochs, trigger early stopping and automatically save the model with the minimum validation loss. S2.3.

4. The time-series enhancement neural network for fillet welds and butt welds constructed in step S2.2 must satisfy the following on the independent test set: Mean Absolute Error (MAE) < 0.1 mm, and Coefficient of Determination (R²) < 0.1 mm. 2 After reaching >0.98, the optimized timing enhancement neural network for fillet welds and butt welds was obtained.

5. The method for geometric dimension control and parameter optimization for fillet and butt welds of ship hulls according to claim 4, characterized in that: The specific steps of step S3 are as follows: S3.1 Construct an objective function based on minimizing the weighted error between the predicted size and the target size, as shown in the following formula: ； Where λ is the penalty coefficient; Penalty(X) is the soft penalty term for parameters approaching the constraint boundary; Y target Indicates the target size; f model (X) represents the predicted size; S3.

2. The objective function constructed in step S3.1 is solved using the hybrid improved particle swarm optimization-simulated annealing algorithm, and the optimal combination of process parameters is output.

6. The method for geometric dimension control and parameter optimization for fillet and butt welds of ship hulls according to claim 5, characterized in that: The solution process of the hybrid improved particle swarm optimization-simulated annealing algorithm in step S3.2 is as follows: a. Population initialization: 30 initial particles are generated within the parameter feasible region using Latin hypercube sampling to ensure uniform spatial distribution; b. Adaptive update: Inertial weight ω decays nonlinearly: ; Dynamic adjustment of learning factors: ； In the formula, T max The total number of iterations is represented by t; the current iteration number is represented by t. c. Constraint handling: If a particle goes out of bounds, the boundary reflection method is used to make it oscillate within the feasible region; d. Hybrid Local Search: Every 5 iterations, simulated annealing local search is performed on the current globally optimal particle, and the probability of accepting a suboptimal solution decays exponentially; e. Convergence criterion: When the global optimal fitness value changes by less than 10 over 10 consecutive generations. -3 The iteration will terminate when the maximum number of iterations is reached. f. Output: The globally optimal particle position is the recommended optimal combination of process parameters X. opt .

7. The method for geometric dimension control and parameter optimization for fillet and butt welds of ship hulls according to claim 6, characterized in that: The specific steps of step S4 are as follows: S4.1 Physical Experiment Verification: Specimen preparation: Based on the optimized parameter combination X output in step S3 opt Standard welding specimens were prepared under the same conditions as those for data acquisition, including base material, welding materials, and equipment. Precision measurement: Accurately extracting the actual solder leg size K actual Or actual remaining height H actual ; Accuracy assessment and quality inspection: Calculate the absolute and relative errors between the actual and target dimensions; S4.2 If the verification result meets the preset accuracy requirements, such as the dimensional error being within ±0.1mm, then add this data pair to the corresponding fillet weld or butt weld database in S1. S4.3 When the update trigger mechanism is met, the mechanism is triggered, and step S2 is re-executed using the latest database to perform full retraining or incremental learning on the corresponding neural network positive prediction model.

8. The method for geometric dimension control and parameter optimization for fillet and butt welds of ship hulls according to claim 7, characterized in that: The sliding window in step S2.1 has a window length of 1 second and a step size of 0.2 seconds.

9. The method for geometric dimension control and parameter optimization for fillet and butt welds of ship hulls according to claim 7, characterized in that: The data preprocessing method in step S1.3 includes: outlier cleaning, data standardization, dataset partitioning, and multi-rule outlier cleaning.

10. An optimization system based on the geometric dimension control and parameter optimization method for fillet and butt welds of ship hulls as described in any one of claims 7 to 9, characterized in that: include The database management module is used to build and maintain the dedicated welding process database; The timing feature extraction module, integrated into the database management module, is responsible for real-time feature calculation and storage of timing signals during the welding process; The intelligent optimization and solution module is used to construct the process parameter optimization model and call the optimization algorithm to solve it. Hybrid optimization solver: Integrated into the intelligent optimization solver module, it has a built-in hybrid improved particle swarm optimization-simulated annealing algorithm and its adaptive parameter adjustment mechanism; The validation and evolution module is used to manage the experimental validation process and drive the closed-loop update of the database and model; Online learning controller: Integrated into the validation and evolution module, it is responsible for monitoring production errors, triggering online fine-tuning, and managing incremental learning processes.

Images