A deep learning-based bimetallic additive manufacturing interface defect prediction method

Abstract

This invention relates to the field of deep learning technology and discloses a deep learning-based method for predicting interface defects in bimetallic additive manufacturing. The method constructs a temporal and spatially heterogeneous database of the entire bimetallic additive manufacturing process and establishes a hierarchical index structure. It synchronously calibrates multi-source in-situ sensor data acquired in real time and combines it with molecular dynamics simulation data of element diffusion at the bimetallic interface to construct a multi-dimensional temporal feature sequence. Based on the real-time acquisition time, it retrieves matching historical temporal and spatially correlated datasets through the hierarchical index structure, outputting feature benchmark thresholds and defect temporal evolution patterns. The multi-dimensional temporal feature sequence and defect temporal evolution patterns are input into a converter network with an attention mechanism. This invention achieves full-time dynamic prediction of bimetallic interface defects, extracts early features of defect initiation, and eliminates prediction lag caused by asynchronous multi-source data.

Description

Technical Field

[0001] This invention relates to the field of deep learning technology and discloses a method for predicting interface defects in bimetallic additive manufacturing based on deep learning. Background Technology

[0002] Predicting interface defects in bimetallic additive manufacturing typically relies on state assessment at a single time point or post-processing identification. Conventional techniques identify interface defects by performing non-destructive testing on the formed part after manufacturing, or by using real-time data collected by a single sensor during manufacturing and classifying the data features at the current moment using traditional machine learning models to determine the presence of defects. In data storage, existing technologies usually store process timing data and molten pool spatial image data independently in conventional relational databases, lacking a temporal-spatial correlation mapping between the data. In model prediction, existing technologies directly input unaligned single-sensor data into conventional neural networks for feature extraction, without incorporating simulations of the physical mechanisms of element diffusion at the bimetallic interface; the model outputs static defect classification results based solely on the data slice at the current moment.

[0003] Based on the aforementioned existing technologies, a core technical problem lies in the fact that, due to the lack of a heterogeneous database mapping temporal and spatial relationships, and the absence of physical constraints and multi-source data alignment mechanisms, existing prediction schemes can only output static current state judgments, failing to predict the dynamic evolution trend of bimetallic interface defects from initiation to expansion throughout the entire timeline. Multi-source sensor data in the manufacturing process exhibit asynchronous deviations along the time axis, and the diffusion behavior of interface elements is a dynamic physical process. Existing technologies sever the temporal correlation of data with the laws of physical evolution, causing the model to be unable to extract early temporal characteristics of defect initiation. This results in lagging prediction results, failing to provide upfront data support for real-time closed-loop control of the manufacturing process. Summary of the Invention

[0004] The purpose of this invention is to provide a solution to the problems described in the background section.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A deep learning-based method for predicting interface defects in bimetallic additive manufacturing includes: Construct a time-space heterogeneous database for the entire process of bimetallic additive manufacturing, establish a hierarchical index structure, store process time data, molten pool spatial image data, interface microstructure characterization data and defect calibration data, and establish a mapping relationship between time data and spatial data; Synchronous calibration is performed on the real-time acquired multi-source in-situ sensing data, and multi-dimensional temporal feature sequences are constructed by combining the molecular dynamics simulation data of element diffusion at bimetallic interfaces. Based on real-time acquisition time, a hierarchical index structure is used for real-time retrieval to extract matching historical time-space correlation datasets and output feature benchmark thresholds and defect time-series evolution patterns. The multi-dimensional temporal feature sequence and the temporal evolution law of defects are input into the Transformer network with attention mechanism. The network focuses on the key temporal feature intervals of defect initiation by combining feature benchmark thresholds, and outputs the prediction results of the initiation time, location, type and expansion trend of defects in bimetallic additive manufacturing interface.

[0006] Preferably, establishing a hierarchical index structure includes: We introduce a B+ tree structure from the field of resource scheduling to construct a first-order index of the time dimension for process timing data, and introduce an R-tree structure from the field of spatial information retrieval to construct a second-order index of the spatial dimension for molten pool spatial image data. The time-dimensional first-order index and the spatial second-order index are merged and connected through the adjacency table of the graph database to generate a time-series-spatial bidirectional joint index. The system utilizes a time-series-spatial bidirectional joint index to respond to real-time retrieval requests and employs a breadth-first traversal strategy of the graph to locate the physical storage nodes of historical time-series-spatial associated datasets.

[0007] Preferably, synchronous calibration of multi-source in-situ sensor data acquired in real time includes: Extract infrared time-series data of the molten pool, plasma spectral time-series data, and vibration signals; A dynamic time warping method from the field of communication signal processing is introduced, using the timestamp of the vibration signal as the reference anchor point, to perform nonlinear time axis warping alignment on the molten pool infrared time series data and the plasma spectral time series data. The token bucket rate limiting strategy in microservice resource scheduling is combined to control the data throughput during the alignment process, and the adaptive Kalman filter method is called to filter out high-frequency Gaussian noise in the aligned data.

[0008] Preferably, the construction of a multi-dimensional temporal feature sequence by combining molecular dynamics simulation data of element diffusion at bimetallic interfaces includes: Extract interface element concentration distribution data and binding energy evolution data from molecular dynamics simulation data; A physical information neural network architecture is introduced to transform the interface element concentration distribution data and binding energy evolution data into physical constraint terms of partial differential equations; The physical constraint terms are spliced ​​with the synchronously calibrated multi-source in-situ sensing data into a tensor. The spliced ​​tensor is then projected onto a unified high-dimensional latent space through a fully connected mapping layer to generate a multi-dimensional temporal feature sequence.

[0009] Preferably, the key temporal feature intervals for the emergence of focusing defects in Transformer networks with attention mechanisms include: A relative positional encoding strategy from the field of natural language processing is introduced to replace absolute positional encoding for positional embedding of multi-dimensional temporal feature sequences. The multi-head self-attention mechanism incorporates the deformable convolution concept from the field of computer vision, dynamically adjusting the sampling offset of each attention head according to the gradient changes of the physical constraint term; The temporal dependency weights within the feature sequence are calculated using a multi-head self-attention mechanism with adjusted sampling offsets, and the interval mapped by the weight peaks is determined as the key temporal feature interval for defect initiation.

[0010] Preferably, the key temporal feature intervals for focusing on defect initiation through network-based feature benchmark thresholding also include: We introduce a proximal policy optimization mechanism from reinforcement learning, and use the feature benchmark threshold of the real-time retrieval output as the prior boundary condition of the state space. A reward function is constructed using the cross-entropy of the predicted probability distribution output by the Transformer network in key temporal feature intervals and the actual defect labels; By using prior boundary conditions to impose a penalty term on the reward function, the loss function of the Transformer network is guided to converge in a direction that conforms to the temporal evolution of defects during backpropagation.

[0011] Preferably, after outputting the predicted results of the initiation time, location, type, and propagation trend of interface defects in bimetallic additive manufacturing, the method further includes: We introduce the matrix factorization collaborative filtering method from the recommender system and construct a user-item latent semantic matrix by combining the output prediction results with the historical time-series-spatial association dataset. The latent semantic matrix is ​​decomposed using stochastic gradient descent to extract the latent feature vector of defect evolution in the current manufacturing state. The feature benchmark threshold in the temporal-spatial bidirectional joint index is dynamically corrected using the latent feature vector of defect evolution, and the corrected data is updated to the physical storage node.

[0012] Preferably, after generating the multi-dimensional temporal feature sequence and before inputting it into the Transformer network with attention mechanism, the following steps are also included: By introducing the t-distributed random neighborhood embedding method from manifold learning and combining it with principal component analysis, nonlinear dimensionality reduction is performed on multi-dimensional temporal feature sequences in high-dimensional latent space. Principal component analysis is used to remove linear redundant dimensions from multi-dimensional time series feature sequences, while retaining principal component features that include physical constraint gradient changes. The principal component features are mapped to a low-dimensional manifold space using the t-distributed random neighborhood embedding method, and the low-dimensional manifold features are extracted as the final input of the Transformer network with attention mechanism.

[0013] Preferably, mapping principal component features to a low-dimensional manifold space using the t-distributed random neighborhood embedding method includes: Introduce dynamic elastic scaling and scheduling strategies from the cloud computing field to monitor the data inflow rate of multi-dimensional time-series characteristic sequences in real time. Based on the difference between the data inflow rate and the preset computing resource threshold, the number of parallel computing threads and memory usage of the t-distribution random neighborhood embedding method during the mapping process are dynamically allocated. An asynchronous update mechanism is used in parallel computing threads to iterate the position coordinates of feature points in the low-dimensional manifold space, thereby shortening the computation time of the dimensionality reduction process.

[0014] Preferably, after outputting the predicted expansion trend of interface defects in bimetallic additive manufacturing, the method further includes: By introducing the model predictive control framework from cybernetics, the prediction results of emergence time, location, type and expansion trend are transformed into a state variable matrix; Pre-constraints for the additive manufacturing process are constructed based on the state variable matrix, and a control adjustment sequence for additive manufacturing equipment is generated by solving a quadratic programming problem. The control adjustment sequence is encapsulated into closed-loop control instructions, and the execution status of the instructions is fed back to the time-space heterogeneous database, triggering incremental update operations of the hierarchical index structure.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This scheme constructs a time-series and spatially heterogeneous database with a hierarchical index structure, mapping and associating process time-series data, molten pool spatial image data, interface microstructure characterization data, and defect calibration data for storage. During real-time prediction, the hierarchical index structure is used to retrieve matching historical time-series and spatially associated datasets in real time, extracting feature benchmark thresholds and defect time-series evolution patterns as prior constraints. Simultaneously, multi-source in-situ sensing data is synchronously calibrated to eliminate time axis bias, and molecular dynamics simulation data is combined to generate a multi-dimensional time-series feature sequence containing physical constraints. By inputting the feature sequence and defect time-series evolution patterns into a converter network with an attention mechanism, the network, combined with prior constraints, focuses on key time-series feature intervals for defect initiation. This achieves full-time-series dynamic prediction of bimetallic additive manufacturing interface defects, from initiation time, location, and type to expansion trends. This overcomes the limitations of existing technologies that can only perform static, post-hoc identification, extracts early time-series features of defect initiation, and eliminates prediction lag caused by asynchronous multi-source data and lack of physical constraints.

[0016] 2. In the data retrieval stage, a first-order index for the time dimension is constructed using a plus tree, and a second-order index for the spatial dimension is constructed using a region tree. A bidirectional joint index is generated by fusing adjacency lists from a graph database, and a breadth-first traversal strategy is combined to improve the real-time retrieval response speed for heterogeneous data. In the data synchronization stage, a dynamic time warping method is used to nonlinearly warp and align the time axis of multi-source sensor data, and an adaptive Kalman filter is used to filter out high-frequency noise, improving the data purity and time alignment accuracy of multi-dimensional time-series feature sequences. In the feature dimensionality reduction stage, principal component analysis and Student's distribution random neighborhood embedding are used to perform nonlinear dimensionality reduction of high-dimensional features, and a dynamic elastic scaling scheduling strategy is used to allocate parallel computing threads and memory usage, shortening the computation time of the dimensionality reduction process. In the process control stage, the prediction results are transformed into a state variable matrix to solve a quadratic programming problem to generate a sequence of control adjustment quantities, and the feature benchmark thresholds in the hierarchical index structure are dynamically corrected using the defect evolution latent feature vector, realizing closed-loop control feedback of the prediction results to the manufacturing process and incremental updates to the database. Attached Figure Description

[0017] Figure 1 This is the overall flowchart of the method of the present invention; Figure 2 This is a flowchart illustrating the construction process of the hierarchical index structure of this invention; Figure 3 This is a flowchart of the multi-source sensor data synchronization calibration process of the present invention; Figure 4 This is a flowchart of the multi-dimensional temporal feature sequence construction process of the present invention; Figure 5 This is a flowchart illustrating the key region focusing process of the Transformer network in this invention. Figure 6 This is a flowchart of the closed-loop control and database update process for the prediction results of this invention. Detailed Implementation

[0018] This invention relates to the fields of additive manufacturing quality inspection and deep learning technology, and specifically to a deep learning-based method for predicting interface defects in bimetallic additive manufacturing. The bimetallic additive manufacturing mentioned in this specific embodiment includes, but is not limited to, additive manufacturing processes based on layer-by-layer cladding, such as directional energy deposition and selective laser melting. The bimetallic system includes, but is not limited to, dissimilar metal material systems such as stainless steel-nickel-based alloys, titanium alloys-aluminum alloys, and copper alloys-steel. The main execution component of the method is an in-situ monitoring and closed-loop control system for additive manufacturing. This system includes a multi-source in-situ sensing unit, a data storage and retrieval unit, a molecular dynamics simulation unit, a deep learning computing unit, and an additive manufacturing equipment control unit. These units are coupled together via an industrial Ethernet bus to achieve data interaction and command transmission. The technical solution of this invention will be described in complete detail below with reference to several embodiments.

[0019] Example 1: Refer to Appendix Figure 1 A time-space heterogeneous database for the entire bimetallic additive manufacturing process was constructed, establishing a hierarchical index structure to store process time-series data, molten pool spatial image data, interface microstructure characterization data, and defect calibration data. A mapping relationship between time-series and spatial data was also established. Specifically, the process time-series data includes a sequence of process parameters continuously acquired over manufacturing time, such as laser power, scanning speed, powder feeding rate, protective gas flow rate, and substrate temperature, with a sampling frequency of 10Hz–10kHz. The molten pool spatial image data includes molten pool infrared thermographic images acquired by a high-speed infrared camera and molten pool morphology images acquired by a high-speed visible light camera, with an image resolution of at least 640×480 and a sampling frame rate of at least 1000fps. The interface microstructure characterization data includes interface element distribution, grain size, and dislocation density microstructure characterization data acquired after manufacturing using scanning electron microscopy, energy dispersive spectroscopy, and electron backscatter diffraction. The defect calibration data includes manufacturing process node data corresponding to the interface defect type, defect location, defect size, and initiation time determined by ultrasonic testing, industrial CT nondestructive testing, and metallographic analysis. The mapping and association between the temporal and spatial data uses a unique manufacturing process ID as the primary key. Each molten pool spatial image data is bound to a unique timestamp, which corresponds one-to-one with the process timeline data at the same time. Simultaneously, the manufacturing layer number and scan path position corresponding to this timestamp are matched and bound with the spatial positions of subsequent interface microstructure characterization data and defect calibration data, establishing a four-element mapping relationship between process timeline, molten pool space, microstructure characterization, and defect calibration. This relationship is stored in the temporal-spatial heterogeneous database. (See Appendix) Figure 2 .

[0020] Synchronous calibration is performed on real-time acquired multi-source in-situ sensing data, combined with bimetallic interface element diffusion molecular dynamics simulation data, to construct a multi-dimensional temporal feature sequence. The multi-source in-situ sensing data comprises real-time molten pool infrared time-series data, plasma spectral time-series data, and vibration signals acquired by multiple in-situ sensors during bimetallic additive manufacturing. Synchronous calibration unifies the time reference of each sensing data point, eliminating nonlinear time axis shifts caused by differences in sampling frequencies between different sensors. The bimetallic interface element diffusion molecular dynamics simulation data is data on the interface atomic diffusion process pre-simulated using a parallel molecular dynamics simulator for the current bimetallic system and process parameter range, including interface element concentration distribution, interface binding energy evolution, and atomic diffusion flux data at different times. The multi-dimensional temporal feature sequence is a unified-dimensional temporal feature sequence generated by fusing the synchronously calibrated multi-source sensing data with the physical constraint data obtained from molecular dynamics simulation. (See attached figure.) Figure 3 .

[0021] Based on real-time acquisition time, a hierarchical index structure is used for real-time retrieval to extract matching historical time-space correlation datasets, outputting feature benchmark thresholds and defect time-series evolution patterns. The real-time acquisition time corresponds to the timestamp of the current manufacturing process. This timestamp, the current process parameter range, and the bimetallic material system are used as search keywords. The hierarchical index structure is used to perform a search in a time-space heterogeneous database, matching historical time-space correlation datasets with a similarity greater than a preset threshold to the current manufacturing condition. The feature benchmark threshold is the threshold range of multi-dimensional features extracted from the matching dataset under normal operating conditions. The defect time-series evolution pattern is the full time-series evolution path from initiation to expansion of different types of defects extracted from the matching dataset, including the inflection point of feature changes in defect initiation, the trend of feature amplitude changes, and the time-series feature evolution pattern of defect expansion.

[0022] Multi-dimensional temporal feature sequences and defect temporal evolution patterns are input into a Transformer network with an attention mechanism. The network, combined with feature benchmark thresholds, focuses on key temporal feature intervals for defect initiation, outputting predictions of the initiation time, location, type, and expansion trend of defects in bimetallic additive manufacturing interfaces. The Transformer network with an attention mechanism includes an embedding layer, an encoder layer, a decoder layer, and an output head. The encoder layer is composed of multiple multi-head self-attention modules and feedforward neural network modules stacked together. The decoder layer is used to fuse prior information about the defect temporal evolution patterns. After embedding the input multi-dimensional temporal feature sequence into its position, the network calculates the temporal dependency of the feature sequence through a multi-head self-attention mechanism. Combined with a feature baseline threshold, it assigns higher attention weights to feature intervals exceeding the baseline threshold, thereby focusing on the key temporal feature intervals for defect initiation. Finally, the network outputs the predicted results of defect initiation time, location, type, and expansion trend through the output head. The initiation time corresponds to the timestamp and layer number of the manufacturing process, the location corresponds to the planar coordinates of the scanning path and the interface depth direction, the type is the classification probability of porosity, crack, lack of fusion, and inclusion, and the expansion trend is the predicted sequence of defect size and expansion direction for multiple future time steps.

[0023] As a preferred embodiment, the establishment of the hierarchical index structure specifically includes: A B+ tree structure is introduced to construct a first-order index for the time dimension of the process timing data. The key value of the B+ tree is the timestamp of the process timing data, and the leaf nodes of the B+ tree store the physical storage address of the process timing data corresponding to the timestamp. The leaf nodes are connected by a doubly linked list, supporting range retrieval within a time interval. The order m of the B+ tree is determined according to the sampling frequency of the process timing data. In this embodiment, m is set to 200 to ensure that the number of I / O operations for a single node retrieval does not exceed 2.

[0024] An R-tree structure is introduced to construct a second-order spatial index for molten pool spatial image data. The minimum bounding rectangle of the R-tree corresponds to the spatial coordinate range of the molten pool spatial image, including the manufacturing layer number, the XY plane coordinates of the scan path, and the Z coordinate in the interface depth direction. The leaf nodes of the R-tree store the physical storage address of the corresponding molten pool spatial image data. The spatial range can be quickly retrieved by judging the overlap of the minimum bounding rectangles.

[0025] The first-order index of the time dimension and the second-order index of the spatial dimension are merged and joined using the adjacency list of the graph database to generate a time-series-spatial bidirectional joint index. Nodes in the graph database represent leaf nodes of the B+ tree and R-tree, respectively. Node attributes include the physical storage address, timestamp, and spatial coordinate range of the corresponding data. Edges in the graph database represent the mapping relationship between timestamps and spatial coordinates, with the edge weight representing the matching degree between the corresponding time-series data and spatial data. The adjacency list stores the relationships between nodes and edges, enabling bidirectional retrieval mapping from the time dimension to the spatial dimension and from the spatial dimension to the time dimension.

[0026] This system utilizes a time-series-spatial bidirectional joint index to respond to real-time retrieval requests, and employs a breadth-first search strategy in the graph to locate the physical storage nodes of historical time-series-spatial associated datasets. Upon receiving a real-time retrieval request, the system first parses the time interval and spatial range within the request. It then uses a first-order B+ tree index to locate the set of leaf nodes matching the time interval. Starting from this set, it performs a breadth-first traversal using an adjacency list, with a traversal depth not exceeding three levels, to quickly locate the R-tree leaf nodes matching the spatial range. Finally, it obtains the physical storage address of the corresponding historical time-series-spatial associated dataset, completing the retrieval.

[0027] The synchronous calibration of the real-time acquired multi-source in-situ sensing data specifically includes: The molten pool infrared time-series data, plasma spectral time-series data, and vibration signals are extracted. Specifically, the molten pool infrared time-series data consists of the peak temperature, average temperature, and temperature gradient time-series data collected by an infrared thermometer at a sampling frequency of 2 kHz; the plasma spectral time-series data consists of the characteristic spectral line intensity, spectral line half-width at half-maximum, electron temperature, and electron density time-series data collected by a spectrometer at a sampling frequency of 5 kHz; and the vibration signals are the vibration acceleration time-series data in the X, Y, and Z directions collected by an accelerometer at a sampling frequency of 10 kHz. The vibration signals are synchronized with the movement of the scanning galvanometer of the additive manufacturing equipment, with timestamp accuracy at the microsecond level, serving as a reference anchor point for time axis alignment.

[0028] A dynamic time warping method is introduced, using the timestamp of the vibration signal as the reference anchor point, to nonlinearly warp and align the molten pool infrared time series data and the plasma spectral time series data. The distance calculation and path solving process of dynamic time warping is as follows: Given a reference vibration signal sequence The sequence length is N; the molten pool infrared timing sequence to be aligned is The sequence length is M, and N≠M.

[0029] Construct the distance matrix , where matrix elements , representing the Euclidean distance between the i-th element of the reference sequence and the j-th element of the sequence to be aligned.

[0030] Construct the cumulative distance matrix The recursive formula for the cumulative distance is:

[0031] The boundary conditions are as follows: , , .

[0032] By backtracking the cumulative distance matrix from arrive The optimal path is obtained to determine the time mapping relationship between the sequence to be aligned and the reference sequence. Based on this mapping relationship, the time axis of the sequence to be aligned is nonlinearly distorted to achieve time axis alignment with the reference vibration signal. The same alignment operation is performed on the plasma spectral time series data to complete the time axis synchronization of multi-source sensing data.

[0033] A token bucket rate limiting strategy is used to control the data throughput during the alignment process, and an adaptive Kalman filter is invoked to filter out high-frequency Gaussian noise in the aligned data. In the token bucket rate limiting strategy, the token bucket capacity is C, the token generation rate is r, and one token is consumed for processing each data frame. When there are no tokens in the token bucket, the data frame is buffered in a waiting queue to avoid computational blocking caused by data bursts. In this embodiment, C is set to 1000, and r is set to match the sensor's maximum sampling frequency. The state equation and observation equation of the adaptive Kalman filter are as follows: Equations of state:

[0034] Observation equation:

[0035] in, Let be the system state vector at time k, corresponding to the aligned true value of the sensor data; The state transition matrix is ​​taken as the identity matrix in this embodiment; The observation matrix is ​​taken as the identity matrix; For process noise, For the purpose of observation, all noises follow a Gaussian distribution.

[0036] Adaptive Kalman filtering estimates the process noise covariance in real time. Covariance of observation noise Update the filter gain This achieves adaptive filtering of high-frequency Gaussian noise, and the filter gain update formula is:

[0037] in, This is the prior estimate of the error covariance matrix.

[0038] As a preferred embodiment, the construction of a multi-dimensional temporal feature sequence by combining bimetallic interface element diffusion molecular dynamics simulation data specifically includes: Interface element concentration distribution data and binding energy evolution data were extracted from molecular dynamics simulation data. The molecular dynamics simulation data was obtained by simulating the interface atomic diffusion process of the current bimetallic system under the temperature and stress fields corresponding to the current process parameters using a large-scale atomic / molecular parallel simulator. The simulation time step was 1 fs, covering the entire cycle of single-channel cladding. The interface element concentration distribution data represents the atomic concentration distribution of alloying elements at different times and spatial locations in the interface region. The binding energy evolution data represents the time-series evolution sequence of atomic binding energy and interface energy at different times in the interface region, referencing... Figure 4 .

[0039] A physical information neural network architecture is introduced to transform the interface element concentration distribution data and binding energy evolution data into physical constraints of partial differential equations. The element diffusion process at the bimetallic interface satisfies Fick's second law, and its partial differential equation form is:

[0040] in, Let be the element concentration at time t and spatial location x. The diffusion coefficient is related to the element concentration C and temperature T.

[0041] The physical information neural network uses the above partial differential equations as physical constraints to construct a constraint loss function. , The sum of squared residuals of the partial differential equation is expressed as:

[0042] Where N is the number of sampling points, , This represents the element concentration and temperature value at the corresponding sampling point.

[0043] The physical constraint terms are concatenated with the synchronously calibrated multi-source in-situ sensing data using tensors. A fully connected mapping layer then projects the concatenated tensors onto a unified high-dimensional latent space, generating a multi-dimensional temporal feature sequence. The synchronously calibrated multi-source in-situ sensing data is then transformed into a sequence with dimension [missing information]. The sensing feature tensor, where T is the time series length. The sensing feature dimension is used to transform the physical constraint term into a dimension of . The physical characteristic tensor The physical feature dimension is used as the basis for concatenating the two tensors along this dimension, resulting in a tensor with dimension . The fused feature tensor is then input into a fully connected mapping layer, which comprises two linear layers and a GELU activation function, with an output dimension of... The high-dimensional latent space feature tensor, i.e., the multi-dimensional temporal feature sequence, in which... This represents the model dimension of the Transformer network; in this embodiment, it is set to 512.

[0044] The key temporal feature intervals for the emergence of focusing defects in the Transformer network with attention mechanism specifically include: A relative positional encoding strategy is introduced to replace absolute positional encoding for positional embedding of multi-dimensional temporal feature sequences. For a feature sequence of length T, the relative positional offset between any two positions i and j is... The relative position encoding matrix is ,in For the dimension of attention head, This represents the maximum range of relative positional offset. The attention score for relative positional encoding is calculated using the following formula:

[0045] Where Q is the query matrix, K is the key matrix, and V is the value matrix, all obtained from the input feature sequence through linear projection. (Refer to...) Figure 5 .

[0046] This paper incorporates deformable convolution into a multi-head self-attention mechanism, dynamically adjusting the sampling offset of each attention head based on the gradient changes of the physical constraint term. The multi-head self-attention mechanism includes h parallel attention heads, each corresponding to a set of sampling points, initially set at uniform sampling positions of the feature sequence. The offset of each sampling position is calculated based on the gradient changes of the physical constraint term. , N is the number of sampling points for each attention head, and the offset is calculated using the following formula:

[0047] in, The offset mapping weight matrix, Physical constraint loss at sampling location The gradient value at that point.

[0048] Based on the calculated offset Adjust the sampling position of each attention head so that the sampling area of ​​the attention head focuses on the region where the physical constraint gradient changes drastically.

[0049] A multi-head self-attention mechanism with adjusted sampling offset is used to calculate the temporal dependency weights within the feature sequence. The interval mapped by the weight peak is determined as the key temporal feature interval for defect initiation. The multi-head self-attention mechanism with adjusted sampling position outputs the attention weight for each temporal position. The attention weight ranges from 0 to 1, with a higher weight indicating a greater contribution of the feature at that position to defect prediction. Peak detection is performed on the attention weight sequence, and continuous temporal intervals with weights exceeding a preset weight threshold are determined as key temporal feature intervals for defect initiation. The preset weight threshold is three times the average attention weight under normal operating conditions.

[0050] The method of focusing on key temporal feature intervals for defect initiation through network-based feature benchmark thresholding also includes: A proximal policy optimization mechanism from reinforcement learning is introduced, using the feature benchmark threshold from real-time retrieval output as the prior boundary condition of the state space. The agent performing the proximal policy optimization corresponds to the encoder parameters of the Transformer network, and the state space... The action space is the vector of differences between the multi-dimensional temporal feature sequence at the current moment and the feature benchmark threshold. The attention weights of the encoder are adjusted by prior boundary conditions, which are the upper and lower limits of the feature baseline threshold. When the value of the feature sequence exceeds the prior boundary conditions, a penalty mechanism is triggered.

[0051] A reward function is constructed using the cross-entropy of the predicted probability distribution output by the Transformer network in key temporal feature intervals and the true defect labels. Reward Function The expression is:

[0052] in, Let cross-entropy be the loss function. Predict the probability distribution of defect types output by the network. This represents the probability distribution of actual defect labels; Let be the intersection-union ratio function. The key temporal feature intervals for network prediction The time interval in which real defects arise; , As the weighting coefficient, in this embodiment , .

[0053] By applying a priori boundary conditions to the reward function with a penalty term, the loss function of the Transformer network is guided to converge in a direction consistent with the temporal evolution of defects during backpropagation. (Penalty term) The expression is:

[0054] in, The penalty coefficient is... It is a state-space vector. This represents the prior boundary corresponding to the feature benchmark threshold.

[0055] Adding a penalty term to the total loss function of the Transformer network; total loss function ,in For cross-entropy loss, The loss is determined by physical constraints. During backpropagation, the policy update magnitude is limited by the pruning objective function optimized from the near-end policy. The pruning objective function is:

[0056] in, For network parameters, This represents the probability ratio between the old and new strategies. For the dominant function, In this embodiment, the cutting factor is used. .

[0057] As a preferred embodiment, after generating the multi-dimensional temporal feature sequence and before inputting it into the Transformer network with attention mechanism, a dimensionality reduction process for the multi-dimensional temporal feature sequence is also included. The specific process is as follows: A t-distributed random neighborhood embedding method is introduced in combination with principal component analysis to perform nonlinear dimensionality reduction on multi-dimensional temporal feature sequences in high-dimensional latent space.

[0058] Principal component analysis is used to remove linear redundant dimensions from multi-dimensional time-series feature sequences, retaining principal component features that include physical constraint gradient changes. The specific steps are as follows: For dimension Multidimensional temporal feature sequences Perform zero-mean transformation to obtain the zero-mean matrix. ; calculate covariance matrix ; For covariance matrix Perform eigenvalue decomposition to obtain eigenvalues. and the corresponding feature vector; Calculate the cumulative variance contribution rate, select the eigenvectors corresponding to the top k eigenvalues ​​with a cumulative variance contribution rate of not less than 95%, and construct the projection matrix. ; Zero-mean matrix Projected onto projection matrix The dimension is obtained as Principal component features are used to eliminate linear redundant dimensions while retaining principal component features that include physical constraint gradient changes.

[0059] The principal component features are mapped to a low-dimensional manifold space using the t-distributed random neighborhood embedding method, and the low-dimensional manifold features are extracted as the final input of the Transformer network with attention mechanism.

[0060] The method of mapping principal component features to a low-dimensional manifold space using the t-distributed random neighborhood embedding method specifically includes: A dynamic elastic scaling scheduling strategy from the cloud computing field is introduced to monitor the data inflow rate of multi-dimensional time-series characteristic sequences in real time. The data inflow rate is the number of frames of the input time-series characteristic sequence per unit time, which is collected in real time through the traffic monitoring module of the industrial Ethernet bus with a sampling period of 100ms.

[0061] Based on the difference between the data inflow rate and a preset computing resource threshold, the number of parallel computing threads and memory usage in the t-distribution random neighborhood embedding method are dynamically allocated during the mapping process. The preset computing resource threshold includes a CPU core count threshold and a memory usage threshold. When the data inflow rate exceeds the upper limit of the threshold, the number of parallel computing threads is increased proportionally, with the maximum number of threads not exceeding the number of physical CPU cores, and a larger memory block is allocated. When the data inflow rate is below the lower limit of the threshold, the number of parallel computing threads is reduced proportionally, releasing idle memory resources, thus achieving dynamic and elastic scaling of computing resources.

[0062] An asynchronous update mechanism is used in the parallel computing thread to iterate the position coordinates of feature points in the low-dimensional manifold space, shortening the computation time of the dimensionality reduction process. During the iterative process of t-distributed random neighborhood embedding, high-dimensional feature points are divided into multiple batches, each batch corresponding to a parallel computing thread. Each thread asynchronously calculates the low-dimensional coordinate gradient of the feature points in its corresponding batch. The gradient calculation formula is:

[0063] in, The KL divergence loss is used to calculate the conditional probability distributions in high-dimensional and low-dimensional spaces. Let be the joint probability of feature points i and j in high-dimensional space. Let be the joint probability of feature points i and j in the low-dimensional space. , Let i and j be the coordinates of feature points i and j in the low-dimensional space.

[0064] Each thread updates the feature point coordinates of its corresponding batch asynchronously, without waiting for the calculation results of other threads. After the iteration is complete, the output dimension is... The low-dimensional manifold features are used, and in this embodiment, d is set to 64, which serves as the final input to the Transformer network.

[0065] After outputting the predicted results of the initiation time, location, type, and expansion trend of interface defects in bimetallic additive manufacturing, the process also includes dynamic correction and updating of the temporal-spatial bidirectional joint index. The specific process is as follows: This paper introduces a matrix factorization collaborative filtering method from recommender systems, constructing a user-item latent semantic matrix by combining the output prediction results with a historical time-series-spatial association dataset. The feature sequence corresponding to the current manufacturing condition is used as the user, and the defect evolution features corresponding to the historical time-series-spatial association dataset are used as the item, thus constructing a user-item rating matrix. Where M is the number of users, N is the number of items, and the matrix elements are... The similarity score between the current working condition and the historical working condition is scored, with a score range of 0 to 5. The higher the similarity, the higher the score.

[0066] The latent semantic matrix is ​​decomposed using stochastic gradient descent to extract the latent feature vectors of defect evolution under the current manufacturing state. The scoring matrix is ​​then used. Decomposed into user latent vector matrix With the item latent vector matrix , where F is the dimension of the latent features, and in this embodiment, F is taken as 32. The objective function of the decomposition is to minimize the squared error:

[0067] in, The regularization coefficient is used to avoid overfitting. Iterative updates are performed using stochastic gradient descent. and The iterative formula is:

[0068]

[0069] in, The learning rate is used. After iterative convergence, the user latent vector corresponding to the current manufacturing condition is extracted. , which serves as the latent feature vector of defect evolution in the current manufacturing state.

[0070] The feature benchmark threshold in the temporal-spatial bidirectional joint index is dynamically corrected using the defect evolution latent feature vector, and the corrected data is updated to the physical storage node. Based on the defect evolution latent feature vector, a correction coefficient for the feature benchmark threshold under the current operating condition is calculated. This correction coefficient is the cosine similarity between the latent feature vector and the latent feature vector under normal operating conditions. The upper and lower limits of the feature benchmark threshold are adjusted according to the correction coefficient. The corrected feature benchmark threshold, the current temporal feature sequence, the melt pool spatial image data, and the prediction results are encapsulated into a new dataset, which is then updated to the physical storage node of the heterogeneous database through the temporal-spatial bidirectional joint index, triggering an incremental update of the index.

[0071] As a preferred embodiment, after outputting the predicted expansion trend of interface defects in bimetallic additive manufacturing, the process also includes closed-loop control of the additive manufacturing process, specifically as follows: A model predictive control framework from cybernetics is introduced, transforming the prediction results of emergence time, location, type, and expansion trend into a state variable matrix. The prediction time domain of this model predictive control is... Control time domain is ,and The state variable matrix , where n is the dimension of the state variables, including the defect prediction initiation time, location coordinates, defect type probability, propagation rate, propagation direction, and current laser power, scanning speed, powder feeding rate and other process parameters.

[0072] Pre-constraints for the additive manufacturing process are constructed based on a state variable matrix. A control regulation sequence for the additive manufacturing equipment is generated by solving a quadratic programming problem. The pre-constraints include upper and lower limits for process parameters, rate of change constraints for control regulation, and an upper limit constraint for defect propagation rate. The control regulation sequence... ,in The process parameter adjustment amounts at time k include laser power adjustment, scanning speed adjustment, and powder feeding rate adjustment.

[0073] The objective function of the quadratic programming problem is:

[0074] in, Let k be the predicted state value at time k+i. Let Q be the target state matrix without defects, Q be the state weight matrix, and R be the control quantity weight matrix.

[0075] Under the pre-constraint conditions, the above quadratic programming problem is solved to obtain the optimal control adjustment sequence. The first element of the sequence is taken as the actual control adjustment at the current moment.

[0076] The control adjustment sequence is encapsulated into closed-loop control commands, and the execution status of these commands is fed back to a time-space heterogeneous database, triggering incremental updates to the hierarchical index structure. The control adjustments are converted into standard communication protocol commands recognizable by the additive manufacturing equipment and sent to the equipment's controller via an industrial Ethernet bus to perform real-time adjustments to process parameters. Simultaneously, the execution status of the commands, adjusted process timing data, and molten pool spatial image data are collected and fed back to the time-space heterogeneous database. The new dataset is stored in the physical storage node, triggering incremental updates to the hierarchical index structure, thus completing the prediction-control-feedback closed-loop process. (Refer to...) Figure 6 .

[0077] All the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for predicting interface defects in bimetallic additive manufacturing based on deep learning, characterized in that, include: Construct a time-space heterogeneous database for the entire process of bimetallic additive manufacturing, establish a hierarchical index structure, store process time data, molten pool spatial image data, interface microstructure characterization data and defect calibration data, and establish a mapping relationship between time data and spatial data; Synchronous calibration is performed on the real-time acquired multi-source in-situ sensing data, and multi-dimensional temporal feature sequences are constructed by combining the molecular dynamics simulation data of element diffusion at bimetallic interfaces. Based on real-time acquisition time, a hierarchical index structure is used for real-time retrieval to extract matching historical time-space correlation datasets and output feature benchmark thresholds and defect time-series evolution patterns. The multi-dimensional temporal feature sequence and the temporal evolution law of defects are input into the Transformer network with attention mechanism. The network focuses on the key temporal feature intervals of defect initiation by combining feature benchmark thresholds, and outputs the prediction results of the initiation time, location, type and expansion trend of defects in bimetallic additive manufacturing interface.

2. The method according to claim 1, characterized in that, Establishing a hierarchical index structure includes: We introduce a B+ tree structure from the field of resource scheduling to construct a first-order index of the time dimension for process timing data, and introduce an R-tree structure from the field of spatial information retrieval to construct a second-order index of the spatial dimension for molten pool spatial image data. The time-dimensional first-order index and the spatial second-order index are merged and connected through the adjacency table of the graph database to generate a time-series-spatial bidirectional joint index. The system utilizes a time-series-spatial bidirectional joint index to respond to real-time retrieval requests and employs a breadth-first traversal strategy of the graph to locate the physical storage nodes of historical time-series-spatial associated datasets.

3. The method according to claim 2, characterized in that, Synchronous calibration of real-time acquired multi-source in-situ sensing data includes: Extract infrared time-series data of the molten pool, plasma spectral time-series data, and vibration signals; A dynamic time warping method from the field of communication signal processing is introduced, using the timestamp of the vibration signal as the reference anchor point, to perform nonlinear time axis warping alignment on the molten pool infrared time series data and the plasma spectral time series data. The token bucket rate limiting strategy in microservice resource scheduling is combined to control the data throughput during the alignment process, and the adaptive Kalman filter method is called to filter out high-frequency Gaussian noise in the aligned data.

4. The method according to claim 3, characterized in that, A multi-dimensional time-series feature sequence was constructed by combining molecular dynamics simulation data of element diffusion at bimetallic interfaces, including: Extract interface element concentration distribution data and binding energy evolution data from molecular dynamics simulation data; A physical information neural network architecture is introduced to transform the interface element concentration distribution data and binding energy evolution data into physical constraint terms of partial differential equations; The physical constraint terms are spliced ​​with the synchronously calibrated multi-source in-situ sensing data into a tensor. The spliced ​​tensor is then projected onto a unified high-dimensional latent space through a fully connected mapping layer to generate a multi-dimensional temporal feature sequence.

5. The method according to claim 4, characterized in that, The key temporal feature intervals for the emergence of focusing defects in Transformer networks with attention mechanisms include: A relative positional encoding strategy from the field of natural language processing is introduced to replace absolute positional encoding for positional embedding of multi-dimensional temporal feature sequences. The multi-head self-attention mechanism incorporates the deformable convolution concept from the field of computer vision, dynamically adjusting the sampling offset of each attention head according to the gradient changes of the physical constraint term. The temporal dependency weights within the feature sequence are calculated using a multi-head self-attention mechanism with adjusted sampling offsets, and the interval mapped by the weight peaks is determined as the key temporal feature interval for defect initiation.

6. The method according to claim 5, characterized in that, The key temporal feature intervals for defect initiation, which are focused on by combining network features with benchmark thresholds, also include: We introduce a proximal policy optimization mechanism from reinforcement learning, and use the feature benchmark threshold of the real-time retrieval output as the prior boundary condition of the state space. A reward function is constructed using the cross-entropy of the predicted probability distribution output by the Transformer network in key temporal feature intervals and the actual defect labels; By using prior boundary conditions to impose a penalty term on the reward function, the loss function of the Transformer network is guided to converge in a direction that conforms to the temporal evolution of defects during backpropagation.

7. The method according to claim 6, characterized in that, After outputting the predicted results of the initiation time, location, type, and propagation trend of interface defects in bimetallic additive manufacturing, the following are also included: We introduce the matrix factorization collaborative filtering method from the recommender system and construct a user-item latent semantic matrix by combining the output prediction results with the historical time-series-spatial association dataset. The latent semantic matrix is ​​decomposed using stochastic gradient descent to extract the latent feature vector of defect evolution in the current manufacturing state. The feature benchmark threshold in the temporal-spatial bidirectional joint index is dynamically corrected using the latent feature vector of defect evolution, and the corrected data is updated to the physical storage node.

8. The method according to claim 7, characterized in that, After generating the multi-dimensional temporal feature sequence, and before inputting it into the Transformer network with attention mechanism, the following steps are also included: By introducing the t-distributed random neighborhood embedding method from manifold learning and combining it with principal component analysis, nonlinear dimensionality reduction is performed on multi-dimensional temporal feature sequences in high-dimensional latent space. Principal component analysis is used to remove linear redundant dimensions from multi-dimensional time series feature sequences, while retaining principal component features that include physical constraint gradient changes. The principal component features are mapped to a low-dimensional manifold space using the t-distributed random neighborhood embedding method, and the low-dimensional manifold features are extracted as the final input of the Transformer network with attention mechanism.

9. The method according to claim 8, characterized in that, Mapping principal component features to a low-dimensional manifold space using the t-distributed random neighborhood embedding method includes: Introduce dynamic elastic scaling and scheduling strategies from the cloud computing field to monitor the data inflow rate of multi-dimensional time-series characteristic sequences in real time. Based on the difference between the data inflow rate and the preset computing resource threshold, the number of parallel computing threads and memory usage of the t-distribution random neighborhood embedding method during the mapping process are dynamically allocated. An asynchronous update mechanism is used in the parallel computing thread to iterate the position coordinates of feature points in the low-dimensional manifold space.

10. The method according to claim 9, characterized in that, After outputting the predicted expansion trend of interface defects in bimetallic additive manufacturing, the following are also included: By introducing the model predictive control framework from cybernetics, the prediction results of emergence time, location, type and expansion trend are transformed into a state variable matrix; Pre-constraints for the additive manufacturing process are constructed based on the state variable matrix, and a control adjustment sequence for additive manufacturing equipment is generated by solving a quadratic programming problem. The control adjustment sequence is encapsulated into closed-loop control instructions, and the execution status of the instructions is fed back to the time-space heterogeneous database, triggering incremental update operations of the hierarchical index structure.

Images