A method of predicting a deposition height of an electric arc additive manufactured component

Abstract

The present application belongs to the technical field of electric arc additive manufacturing, and specifically relates to a method for predicting the deposition height of an electric arc additive manufacturing component, which comprises the following steps: obtaining the process characteristics and time sequence characteristics of the component during the electric arc additive manufacturing process to construct a multi-factor characteristic input sequence; taking a time convolution network (TCN) as a basic model, fusing a process characteristic linear modulation (FiLM) module and a causal attention-based time sequence characteristic fusion enhancement module (CAFB) to construct a CAFi-TCN prediction model; and inputting the multi-factor characteristic input sequence into the CAFi-TCN prediction model to output the component deposition height prediction results corresponding to each deposition position of the component. The present method takes the prediction of the component deposition height as the target, constructs a multi-factor characteristic input sequence that fuses process parameters and historical deposition state information, introduces a process characteristic linear modulation mechanism and a causal attention mechanism, and constructs a prediction model based on a time convolution network, so that the prediction accuracy of the component deposition height can be significantly improved.

Description

Technical Field

[0001] This invention mainly relates to the technical field of electric arc additive manufacturing, specifically a method for predicting the deposition height of components manufactured by electric arc additive manufacturing. Background Technology

[0002] Component deposition height is an important geometric indicator characterizing the forming quality of arc additive manufacturing, and it can directly reflect the changes in the overall dimensions of the component during multi-layer deposition. If the deposition height differs significantly from the target value, it may affect subsequent processing and actual service performance.

[0003] Temporal Convolutional Networks (TCNs) can be used to predict the deposition height of components in arc additive manufacturing. In modeling the deposition process of arc additive manufacturing, the deposition height is not only closely related to historical deposition information but also influenced by process characteristics such as interlayer cooling time, the current number of deposition layers, and deposition direction. These process characteristics typically remain constant within a single deposition layer and can be considered as local static conditions; however, historical deposition states exhibit significant temporal variations. If a traditional TCN model is used to predict deposition height, all input features are processed uniformly during convolution calculations, which fails to reflect the differences in deposition height variation patterns under different process conditions. In arc additive manufacturing, the deposition state is historically dependent; as the welding torch moves, heat gradually accumulates in the deposited area and is transferred to subsequent areas. Therefore, the deposition state at different historical moments will have varying degrees of influence on the current deposition state. This characteristic places higher demands on the deposition height prediction model, requiring the model to distinguish the importance of state information from different historical deposition locations to the current prediction result. However, traditional TCNs typically use fixed weights for feature extraction when processing historical information, making it difficult to effectively reflect the contribution of different time steps to the current prediction result.

[0004] Therefore, due to the accumulation of heat input and the coupling effect of multiple process parameters in the electric arc additive manufacturing process, the deposition height of the component will show obvious nonlinear changes with the deposition position. It is difficult to accurately predict the changes in deposition height by relying solely on monitoring data within a certain period of the deposition process. Summary of the Invention

[0005] To address the shortcomings of current technologies, this invention, based on existing technologies and practical applications, provides a method for predicting the deposition height of components manufactured by electric arc additive manufacturing. This method aims to predict the deposition height of components by constructing a multi-factor feature input sequence that integrates process parameters and historical deposition state information. It introduces a linear modulation mechanism for process features and a causal attention mechanism, and builds a prediction model based on a temporal convolutional network, which can significantly improve the accuracy of component deposition height prediction.

[0006] The technical solution of the present invention is as follows:

[0007] A method for predicting the deposition height of components manufactured by electric arc additive manufacturing includes:

[0008] S1. Obtain the process characteristics and temporal characteristics of the component in the arc additive manufacturing process to construct a multi-factor feature input sequence;

[0009] S2. Based on the Temporal Convolutional Network (TCN) model, a CAFi-TCN prediction model is constructed by integrating the process feature linear modulation (FiLM) module and the causal attention-based temporal feature fusion enhancement module (CAFB).

[0010] S3. Input the multi-factor feature input sequence into the CAFi-TCN prediction model and output the predicted deposition height of the component corresponding to each deposition location.

[0011] Furthermore, the method for constructing the multi-factor feature input sequence in step S1 is as follows:

[0012] S11. Collect the current deposition layer number N, interlayer cooling time CT, and deposition direction Dir feature data as process features, and the deposition location Loc, previous layer height PreLH, and previous layer deposition height PreDH feature data as time series features.

[0013] S12. Discretize the deposition path within the effective sampling interval according to the set sampling step size to obtain the feature vectors corresponding to each deposition location.

[0014] S13. Organize the feature vectors of each deposition location into a multi-factor feature input sequence according to the deposition path order.

[0015] Furthermore, the effective sampling range is 5-95 mm from the deposition start point, the sampling step size is set to 0.5 mm, and the deposition direction characteristics are numerically encoded: odd-numbered layers are encoded as 1, and even-numbered layers are encoded as -1.

[0016] Furthermore, in the CAFi-TCN prediction model, the temporal convolutional network TCN is composed of multiple conditional TCN modules stacked together. Causal convolution and dilated convolution are used to extract temporal features, and residual connections are used to alleviate gradient vanishing in deep networks.

[0017] The process feature linear modulation module FiLM takes the process feature as a conditional input and outputs a scaling vector and a bias vector to perform scaling and offset linear modulation on the timing feature output by the TCN module.

[0018] The temporal feature fusion enhancement module CAFB uses causal multi-head attention to differentiate and weight historical features at different times, and outputs enhanced temporal features.

[0019] Furthermore, in the CAFi-TCN prediction model, at the input end, temporal features are input into the TCN backbone network, and process features are input into the process feature linear modulation module FiLM. The intermediate features of each conditional TCN module are linearly modulated. After the output of the conditional TCN module, it is further input into the temporal feature fusion enhancement module CAFB. Under the constraint of causal mask, the temporal features are weighted and fused. Finally, the fused features are input into the output layer to obtain the predicted values ​​of the component deposition height corresponding to each deposition position in the current deposition layer.

[0020] Furthermore, the execution steps of the process characteristic linear modulation module FiLM include:

[0021] The interlayer cooling time CT, the current number of deposition layers N, and the deposition direction Dir of the process features are used as conditional inputs and applied to the intermediate feature layer of the conditional TCN module. In the conditional TCN module, the process features are first converted into modulation parameters through linear mapping:

[0022]

[0023] In the formula, and These represent the scaling vector and the bias vector, respectively. Indicates a linear layer. Indicates process characteristics;

[0024] Then the output features of the intermediate feature layer are processed. H i Perform linear modulation:

[0025]

[0026] In the formula, This indicates element-wise multiplication.

[0027] Furthermore, the execution steps of the temporal feature fusion enhancement module CAFB include:

[0028] The modulated temporal features are mapped to obtain query vector, key vector, and value vector, respectively.

[0029] A causal mask is applied to block information about future moments, allowing only historical and current time steps to participate in attention calculations; multi-head causal attention operations are performed, dynamically assigning weights to features at different historical time steps; the multi-head attention outputs are concatenated, processed by a feedforward network, and then added to the input features as residuals to output the temporal feature fusion enhancement result.

[0030] The beneficial effects of this invention are:

[0031] This invention aims to predict the deposition height of structural members. It constructs a multi-factor feature input sequence that integrates process parameters and historical depositional state information. A linear modulation mechanism for process features and a causal attention mechanism are introduced. A CAFi-TCN prediction model is established based on a temporal convolutional network. This model incorporates a FiLM mechanism on top of the TCN, enabling process conditions to participate in the temporal feature extraction process. Combined with a CAFB module, historical time-step information is differentially weighted, thereby enhancing the model's ability to represent key historical features. Experiments demonstrate that the proposed method not only possesses high prediction accuracy and stability but also provides theoretical basis and methodological support for subsequent depositional process control and forming quality optimization. Attached Figure Description

[0032] Figure 1 This is a flowchart of the method proposed in this invention.

[0033] Figure 2 This is a schematic diagram of the linear modulation module structure, which is a process feature of the present invention.

[0034] Figure 3 This is a schematic diagram of the temporal feature fusion enhancement module structure of the present invention.

[0035] Figure 4 This is a schematic diagram of the prediction model structure of the present invention. In the figure, part (a) is the overall structure of the model, part (b) is the structure of the conditional TCN module, and part (c) is the causal attention structure.

[0036] Figure 5 This is a schematic diagram of the comparative experimental results of deposition height prediction in an embodiment of the present invention. Part (a) is a comparison diagram after 2 minutes of interlayer cooling, part (b) is a comparison diagram after 3 minutes of interlayer cooling, part (c) is a comparison diagram after 4 minutes of interlayer cooling, and part (d) is a comparison diagram of prediction accuracy.

[0037] Figure 6 This is a schematic diagram of the predicted height curves under different interlayer cooling time conditions according to an embodiment of the present invention. Detailed Implementation

[0038] The present invention will be further described in conjunction with the accompanying drawings and specific embodiments. It should be understood that these embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. Furthermore, it should be understood that after reading the teachings of this invention, those skilled in the art can make various alterations or modifications to the invention, and these equivalent forms also fall within the scope defined in this application.

[0039] This embodiment provides a method for predicting the deposition height of components manufactured by electric arc additive manufacturing. The main process is as follows: Figure 1 As shown.

[0040] The arc additive manufacturing process is influenced by a variety of process factors and historical depositional conditions. To comprehensively describe the characteristics of the deposition process, this embodiment selects the deposition location (Loc), current deposition layer number (N), interlayer cooling time (CT), deposition direction (Dir), and the depositional state of the previous layer as model inputs. The depositional state of the previous layer includes the previous layer height (PreLH) and the previous layer depositional height (PreDH). In subsequent model construction, the current deposition layer number (N), interlayer cooling time (CT), and deposition direction (Dir) are used as process features, while the deposition location (Loc), the previous layer height (PreLH), and the previous layer depositional height (PreDH) are used as temporal features.

[0041] In this embodiment, a multi-factor feature input sequence is first constructed based on process characteristics and time-series characteristics, as follows.

[0042] During the deposition process, each depositional location corresponds to a set of multi-factor features. For ease of processing, each input feature is constructed as a feature vector at a single depositional location, expressed as follows:

[0043]

[0044] In the formula, This represents the input feature vector corresponding to the i-th deposition location. Since the deposition locations have a clear order during the deposition of a single-pass multi-layer component, the feature vectors of all locations are organized into a sequence according to the deposition path order:

[0045]

[0046] In the formula, This represents the input feature sequence corresponding to the Nth sedimentary layer, and T represents the number of sampling points in a single sedimentary path, i.e., the sequence length.

[0047] Based on the above-mentioned multi-factor feature construction method, this embodiment models the component deposition height prediction problem as a time series mapping problem. The model takes the multi-factor feature sequence of the Nth layer as input. As input, the output is the sequence of deposition heights corresponding to each deposition location within this layer:

[0048]

[0049] In the formula, This indicates the deposition height of the component where the Nth layer is deposited; Let represent the deposition height at the i-th deposition location. Deposition height is primarily influenced by technological characteristics and historical depositional conditions; therefore, this mapping relationship can be further expressed as:

[0050]

[0051] In the formula, This represents the constructed sedimentation height prediction model.

[0052] In this embodiment, the data used by the model was acquired through an electric arc additive manufacturing experimental system. The interlayer cooling time was set to 2 min, 3 min, and 4 min, respectively. Under these conditions, data were collected for different deposition processes to construct a multi-factor feature dataset. During data processing, considering the potential errors in the deposition height measurement process, and the influence of arc initiation and extinction on the initial and final deposition stages, which could interfere with model training, this embodiment selected data within a range of 5-95 mm from the deposition initiation point as the effective sample area. Within this sampling interval, the deposition path was discretized with a sampling step size of 0.5 mm to obtain the multi-factor feature data corresponding to each deposition layer location. Therefore, the number of sampling points T in a single deposition path is 181.

[0053] Because the input variables differ in units and ranges, normalization is necessary to avoid unbalanced impacts on model training caused by different features. For continuous variables (such as deposition location, number of layers, interlayer cooling time, and height), the min-max normalization method is used, and its calculation formula is as follows:

[0054]

[0055] In the formula, For the normalized data, This is the original data. The minimum value in the original training set. This represents the maximum value in the original training data. Regarding the deposition direction variable, since the additive manufacturing process uses a cyclic deposition method, the deposition directions of adjacent layers are opposite. Therefore, the direction of odd-numbered layers is defined as 1, and the direction of even-numbered layers is defined as -1.

[0056] Based on the multi-factor feature input sequence constructed in the above embodiments, deposition height prediction is essentially a time-series modeling problem constrained by process conditions. Considering the obvious local correlation, long-term dependence, and process condition coupling characteristics of the deposition process, this embodiment selects the TCN temporal convolutional network as the basic model, and further improves it from two aspects: process condition modulation and key historical information modeling.

[0057] Temporal Convolutional Network (TCN) is a time-series modeling method based on a one-dimensional convolutional structure, effectively extracting temporal dependencies from sequential data. TCN networks are typically composed of multiple stacked TCN modules, employing causal convolution and dilated convolution to model temporal features, and incorporating residual connections to alleviate the gradient vanishing problem during deep network training. If a traditional TCN model is used to predict deposition height, all input features are processed uniformly during convolution calculations. This approach struggles to reflect the differences in deposition height variation under different process conditions. Therefore, this embodiment proposes a Conditional Temporal Convolutional Network (CTN) structure based on Feature-wise Linear Modulation (FiLM), as follows... Figure 2 As shown, this structure takes the current number of deposition layers N, interlayer cooling time CT, and deposition direction Dir as input conditions and applies them to the intermediate feature layer of the TCN block to achieve adaptive modulation of the process timing characteristics.

[0058] Let the input historical sedimentary state temporal characteristics be:

[0059]

[0060] The technological characteristics are:

[0061]

[0062] In the formula, B represents the batch size, and T is the sequence length. Represents the temporal feature dimension. This represents the dimension of the process features. In the i-th conditional TCN module, the process features are first converted into modulation parameters through a linear mapping:

[0063]

[0064] In the formula, and These represent the scaling vector and the bias vector, respectively. This represents a linear layer. The output feature H of this layer is then... i Perform linear modulation:

[0065]

[0066] In the formula, This indicates element-wise multiplication.

[0067] Traditional TCNs typically use fixed weights for feature extraction when processing historical information, which makes it difficult to effectively reflect the contribution of different time steps to the current prediction result. Therefore, this embodiment introduces a causal attention-based temporal feature fusion enhancement module (CAFB) based on conditional TCN to further model long-term dependencies and highlight features at key time steps. Its structure is as follows: Figure 3 As shown. Let the feature mapping of the conditional TCN output be:

[0068]

[0069] In the formula, C represents the number of feature channels. First, the query, key, and value are obtained through linear projection:

[0070]

[0071] A causal mask is used in the attention calculation process to ensure that only historical information up to and including time t is used in the prediction at time step t, thereby avoiding the leakage of future information. The attention calculation formula is as follows:

[0072]

[0073] In the formula, This represents the causal mask matrix. Through a multi-head attention mechanism, the model can differentially weight historical features at different times, thereby enhancing its ability to represent information at key time steps. Finally, the attention output and conditional TCN features are fused through residual connections to form an enhanced temporal feature representation.

[0074] like Figure 4 As shown, this embodiment constructs a CAFi-TCN prediction model that integrates linear modulation of process features with causal attention mechanism enhancement. Based on the TCN framework, this model improves upon traditional TCN in two aspects: process condition modeling and long-term dependency characterization, thereby enhancing the model's ability to learn the variation patterns of component deposition height. Figure 4 Part (a) is the overall structure of the model, part (b) is the structure of the conditional TCN module, and part (c) is the causal attention structure.

[0075] At the input end, the deposition location (Loc), the previous layer height (PreLH), and the previous layer deposition height (PreDH) are input as temporal features into the TCN backbone network to characterize the historical depositional state along the depositional path. The current depositional layer number (N), interlayer cooling time (CT), and depositional direction (Dir) are input as process condition features into the FiLM module to describe the technological background information corresponding to the current depositional layer. The main body of the model consists of seven stacked conditional TCN blocks, each containing causal convolutions, dilated convolutions, and residual connection structures. The process condition features in each TCN block generate scaling and bias vectors through the FiLM mechanism, linearly modulating the intermediate features. After the conditional TCN output, it is further input into the causal attention module to complete the weighted fusion of temporal features under causal mask constraints. Finally, the fused features are input into the output layer to obtain the predicted depositional height of the components corresponding to each depositional location within the current depositional layer.

[0076] To verify the performance advantages of the CAFi-TCN model proposed in this embodiment in the task of predicting the deposition height of additive components, this embodiment selects two representative benchmark models in the field of time series prediction for comparative experiments: Long Short-Term Memory (LSTM) network and standard temporal convolutional network. As a classic variant of recurrent neural network (RNN), LSTM effectively solves the gradient explosion and gradient vanishing problems that traditional RNNs are prone to when processing long sequence data by introducing unique gating mechanisms (input gate, forget gate, and output gate), and is widely used in various time series prediction tasks.

[0077] In the comparative experiments, both LSTM and the standard TCN used the same input feature settings as the model in this embodiment, namely Loc, N, CT, Dir, PreLH, and PreDH, to ensure the fairness of the model comparison. The dataset was divided based on different interlayer cooling times and the number of deposition layers. Data from deposition layers 1-9 with interlayer cooling times of 2, 3, and 4 minutes were selected as the training set for model training and optimization; data from deposition layer 10 was selected as the test set to evaluate the predictive performance of each model.

[0078] To evaluate the prediction performance of each model on deposition height, this embodiment selects Mean Absolute Error (MAE) and Prediction Accuracy (PA) as evaluation metrics. Mean Absolute Error (MAE) is the average of the absolute values ​​of the differences between the predicted and actual values ​​of the samples; it is a conventional evaluation metric. The formula for calculating PA is as follows:

[0079]

[0080] In the formula, This represents the actual component height value at time step t. This represents the component height value predicted by the model at time step t. PA reflects the consistency between the predicted result and the actual value; the closer its value is to 1, the higher the accuracy of the model prediction.

[0081] This embodiment trains the model on an NVIDIA GeForce RTX 4090 in a WSL Ubuntu 22.04.5 environment, using the PyTorch deep learning framework. During training, the same data input format and training strategy are used for LSTM, TCN, and the CAFi-TCN proposed in this embodiment, while maintaining the model size at roughly the same order of magnitude. MSE is used as the loss function, the initial learning rate is set to 0.001, and the Adam optimizer is used to update the model parameters.

[0082] Figure 5 The prediction results of LSTM, TCN, and the model in this embodiment for the height of the 10th deposited component under different interlayer cooling time conditions are presented. Figure 5 (a) Figure 5 (b) Figure 5 In diagram (c), the black line represents the true value of the component height; the blue, green, and red lines represent the predicted values ​​and absolute values ​​of the prediction errors of the LSTM, TCN, and CAFi-TCN models in this embodiment, respectively. Figure 5 In the middle (d), the average prediction accuracy of each model is shown on the three test sets. The prediction accuracy of the test sets with different interlayer cooling times is shown in Table 1.

[0083] Table 1. Comparison of prediction accuracy of each model in the test set experiment.

[0084] Experimental results show that the CAFi-TCN model proposed in this embodiment exhibits the best predictive performance under all three interlayer cooling time conditions. Whether in the initial deposition stage or the steady-state deposition stage, the predicted curves of CAFi-TCN maintain a high degree of consistency with the actual deposition height curves, and the overall prediction error remains at a low level. Compared with LSTM and the standard TCN, the model in this embodiment can more accurately characterize the variation law of the deposition height of the component.

[0085] In actual additive manufacturing processes, when it is necessary to adjust process parameters such as interlayer cooling time, repeated experiments are usually required to observe the impact of parameter changes on the forming results. This not only increases experimental costs but also reduces process optimization efficiency. To address this issue, this embodiment utilizes a trained CAFi-TCN model to perform parameter substitution prediction: deposition data under the condition of a real interlayer cooling time CT = 3 min is selected, keeping other input features unchanged, and only the condition variable CT is set to 2 min, 3 min, and 4 min respectively, thereby obtaining the predicted height curves under different interlayer cooling time conditions, such as... Figure 6 As shown.

[0086] The experimental results show that, while keeping other real input features constant, the model's prediction results can shift with changes in CT parameters. Specifically, the prediction curve under the condition of CT = 4 min is generally higher than that under the condition of CT = 3 min, while the prediction curve under the condition of CT = 2 min is generally lower, indicating that the model can provide an intuitive reference for the trend of deposition height changes after interlayer cooling time adjustment.

[0087] In this embodiment, a multi-factor input feature sequence was constructed by combining information such as deposition location, current deposition layer number, interlayer cooling time, deposition direction, and the height and depositional height of the previous layer. Data acquisition, effective interval screening, and normalization were then completed. Subsequently, this embodiment proposes a CAFi-TCN model that integrates linear modulation of process features and a causal attention mechanism. This model introduces the FiLM mechanism on the basis of TCN, enabling process conditions to participate in the temporal feature extraction process; combined with the CAFB module, historical time step information is differentially weighted, thereby enhancing the model's ability to represent key historical features. Comparative experimental results show that the proposed model achieves good prediction results under different interlayer cooling time conditions, with an average prediction error of 0.18 mm and a prediction accuracy of 0.99, significantly better than LSTM and the standard TCN model. In summary, the CAFi-TCN model proposed in this embodiment not only has high prediction accuracy and stability, but also provides a theoretical basis and methodological support for subsequent deposition process control and forming quality optimization.

Claims

1. A method for predicting the deposition height of components manufactured by electric arc additive manufacturing, characterized in that, include: S1. Obtain the process characteristics and temporal characteristics of the component in the arc additive manufacturing process to construct a multi-factor feature input sequence; S2. Based on the Temporal Convolutional Network (TCN) model, a CAFi-TCN prediction model is constructed by integrating the process feature linear modulation (FiLM) module and the causal attention-based temporal feature fusion enhancement module (CAFB). S3. Input the multi-factor feature input sequence into the CAFi-TCN prediction model and output the predicted deposition height of the component corresponding to each deposition location.

2. The method for predicting the deposition height of arc additive manufacturing components according to claim 1, characterized in that, The method for constructing the multi-factor feature input sequence in step S1 is as follows: S11. Collect the current deposition layer number N, interlayer cooling time CT, and deposition direction Dir feature data as process features, and the deposition location Loc, previous layer height PreLH, and previous layer deposition height PreDH feature data as time series features. S12. Discretize the deposition path within the effective sampling interval according to the set sampling step size to obtain the feature vectors corresponding to each deposition location. S13. Organize the feature vectors of each deposition location into a multi-factor feature input sequence according to the deposition path order.

3. The method for predicting the deposition height of arc additive manufacturing components according to claim 2, characterized in that, The effective sampling range is 5-95 mm from the deposition start point, with a sampling step size of 0.5 mm. The deposition direction characteristics are numerically encoded: odd-numbered layers are encoded as 1, and even-numbered layers are encoded as -1.

4. The method for predicting the deposition height of arc additive manufacturing components according to claim 1, characterized in that, In the CAFi-TCN prediction model, the temporal convolutional network TCN is composed of multiple conditional TCN modules stacked together. Causal convolution and dilated convolution are used to extract temporal features, and residual connections are used to alleviate gradient vanishing in deep networks. The process feature linear modulation module FiLM takes the process feature as a conditional input and outputs a scaling vector and a bias vector to perform scaling and offset linear modulation on the timing feature output by the TCN module. The temporal feature fusion enhancement module CAFB uses causal multi-head attention to differentiate and weight historical features at different times, and outputs enhanced temporal features.

5. The method for predicting the deposition height of arc additive manufacturing components according to claim 4, characterized in that, In the CAFi-TCN prediction model, at the input end, temporal features are input into the TCN backbone network, and process features are input into the process feature linear modulation module FiLM. The intermediate features of each conditional TCN module are linearly modulated. After the output of the conditional TCN module, it is further input into the temporal feature fusion enhancement module CAFB. Under the constraint of causal mask, the temporal features are weighted and fused. Finally, the fused features are input into the output layer to obtain the predicted values ​​of the component deposition height corresponding to each deposition position in the current deposition layer.

6. The method for predicting the deposition height of arc additive manufacturing components according to claim 4, characterized in that, The execution steps of the process characteristic linear modulation module FiLM include: The interlayer cooling time CT, the current number of deposition layers N, and the deposition direction Dir of the process features are used as conditional inputs and applied to the intermediate feature layer of the conditional TCN module. In the conditional TCN module, the process features are first converted into modulation parameters through linear mapping: ， In the formula, and These represent the scaling vector and the bias vector, respectively. Indicates a linear layer. Indicates process characteristics; Then the output features of the intermediate feature layer are processed. H i Perform linear modulation: ， In the formula, This indicates element-wise multiplication.

7. The method for predicting the deposition height of arc additive manufacturing components according to claim 4, characterized in that, The execution steps of the temporal feature fusion enhancement module CAFB include: The modulated temporal features are mapped to obtain query vector, key vector, and value vector, respectively. A causal mask is applied to block information about future moments, allowing only historical and current time steps to participate in attention calculations; multi-head causal attention operations are performed, dynamically assigning weights to features at different historical time steps; the multi-head attention outputs are concatenated, processed by a feedforward network, and then added to the input features as residuals to output the temporal feature fusion enhancement result.

Images