A multi-sensor fusion-based additive manufacturing process online monitoring and quality evaluation method and system

Abstract

This invention provides a method and system for online monitoring and quality assessment of additive manufacturing processes based on multi-sensor fusion, relating to the technical field of additive manufacturing. The method includes: constructing a multimodal spatiotemporal dataset; extracting features from each physical modality data to obtain an independent feature representation set; inputting the independent feature representation set into a preset dynamic cross-modal attention network to calculate the mutual information correlation matrix and perform feature weighted reconstruction to obtain a deep fusion feature vector; inputting the deep fusion feature vector into a preset quality mapping model to output quality assessment index values ​​representing defect probability and defect classification; constructing a three-dimensional quality digital twin model; and outputting online monitoring and early warning commands and global quality assessment results based on the three-dimensional quality digital twin model. This invention completely eliminates the problem of spatiotemporal registration misalignment of multimodal data, achieving high-precision deep feature fusion and voxel-level panoramic three-dimensional quality space mapping and tracing.

Description

Technical Field

[0001] This invention relates to the field of additive manufacturing technology, and in particular to a method and system for online monitoring and quality assessment of additive manufacturing processes based on multi-sensor fusion. Background Technology

[0002] Additive manufacturing technology has been widely applied in industries such as aerospace and high-end equipment manufacturing due to its significant advantages in forming complex structures and improving material utilization. However, because the additive manufacturing process involves complex, transient heat transfer and multi-physics coupled dynamics, the formed parts are prone to defects such as porosity and cracks. Currently, the industry generally relies on offline non-destructive testing methods after manufacturing, such as using computed tomography (CT) scans for post-production inspection. As described in the research literature on online monitoring methods for additive manufacturing defects published in the *Journal of Mechanical Engineering*, this offline inspection method is not only costly and time-consuming, but more critically, it cannot intervene in a timely manner at the initial stage of defect formation, leading to material waste and high processing costs.

[0003] To overcome the significant limitations of post-process offline inspection, current industry technology evolution is accelerating towards online process monitoring and real-time quality assessment. Core monitoring methods are shifting from reliance on single visual sensors to the collaborative fusion of multi-modal sensors. By simultaneously acquiring multi-source heterogeneous data such as acoustic emission signals, high-speed visual images, and infrared thermal imaging, the aim is to comprehensively capture transient characteristics of the manufacturing process from different physical dimensions. A research team, in a paper published in the core Chinese journal *Electric Welding Machine* on the progress of monitoring and control in laser additive manufacturing processes, also clearly pointed out that the deep fusion of multi-physics sensor data is an inevitable evolutionary direction for improving defect identification capabilities and prediction accuracy under complex manufacturing conditions.

[0004] Despite the immense application potential of multi-sensor online monitoring, existing technologies still face significant technical barriers in practical engineering implementation. Firstly, heterogeneous data registration accuracy is low. Due to the vast differences in sampling frequencies and spatial resolutions among different sensors, existing methods often employ coarse-grained time alignment, lacking precise hardware-level synchronization of physical coordinates and system timestamps, resulting in severe spatiotemporal misalignment in the fused data. Secondly, the depth of feature fusion is severely insufficient. Existing strategies mostly remain at the level of shallow data stitching or end-of-pipe decision voting, failing to delve into the deep nonlinear correlations and mutual information between different physical modes such as thermal flow fields and acoustic responses. Finally, there is a lack of a three-dimensional quality mapping mechanism. Existing evaluation results typically only determine whether there are anomalies on the current two-dimensional processing plane, failing to accurately map transient monitoring results back to the solid space of the formed part to form a panoramic internal quality state model that grows synchronously with the processing process. This makes the evaluation results severely lack intuitive three-dimensional spatial traceability and guidance. Summary of the Invention

[0005] To overcome the shortcomings of existing technologies, the purpose of this invention is to provide a method and system for online monitoring and quality assessment of additive manufacturing processes based on multi-sensor fusion. This invention solves the problems of spatiotemporal registration misalignment of multi-source heterogeneous sensor data, shallow cross-modal feature fusion, and lack of a three-dimensional voxel-level quality space mapping traceability mechanism in existing technologies.

[0006] To achieve the above objectives, the present invention provides the following solution: a method for online monitoring and quality assessment of additive manufacturing processes based on multi-sensor fusion, comprising: Simultaneously acquire multi-source heterogeneous sensor data streams from the processing area, as well as the real-time trajectory coordinates and system timestamp of the current processing position; Based on the real-time trajectory coordinates and the system timestamp, the multi-source heterogeneous sensor data stream is time-series synchronized and spatially mapped to construct a multimodal spatiotemporal dataset. Feature extraction is performed on each physical modality data in the multimodal spatiotemporal dataset to obtain a corresponding set of independent feature representations; The independent feature representation set is input into a preset dynamic cross-modal attention network to calculate the mutual information correlation matrix between nodes within the independent feature representation set, and based on the mutual information correlation matrix, the independent feature representation set is reconstructed and adaptively weighted to obtain a deep fusion feature vector. The deep fusion feature vector is input into a preset quality mapping model, and the quality evaluation index value corresponding to the current processing position is output. The quality evaluation index value represents the defect probability and defect classification of the current processing position. Extract the discrete spatial coordinates of the current processing position in the three-dimensional solid coordinate system, assign the quality assessment index value to the discrete voxel unit corresponding to the discrete spatial coordinate, and progressively update the discrete voxel unit carrying the quality assessment index value to construct a three-dimensional quality digital twin model. Based on the aforementioned three-dimensional quality digital twin model, online monitoring and early warning commands and global quality assessment results for the additive manufacturing process are output.

[0007] An online monitoring and quality assessment system for additive manufacturing processes based on multi-sensor fusion includes: The data synchronization acquisition module is used to synchronously acquire multi-source heterogeneous sensor data streams in the processing area, as well as the real-time trajectory coordinates and system timestamp of the current processing position; The multimodal spatiotemporal mapping module is used to perform temporal synchronization and spatial mapping on the multi-source heterogeneous sensor data stream based on the real-time trajectory coordinates and the system timestamp, so as to construct a multimodal spatiotemporal dataset. An independent feature extraction module is used to extract features from each physical modality data in the multimodal spatiotemporal dataset to obtain a corresponding set of independent feature representations. The cross-modal attention fusion module is used to input the independent feature representation set into a preset dynamic cross-modal attention network to calculate the mutual information correlation matrix between nodes within the independent feature representation set, and to perform feature reconstruction and adaptive weight allocation on the independent feature representation set based on the mutual information correlation matrix to obtain a deep fusion feature vector. The quality assessment mapping module is used to input the deep fusion feature vector into a preset quality mapping model and output the quality assessment index value corresponding to the current processing position, wherein the quality assessment index value represents the defect probability and defect classification of the current processing position; The three-dimensional twin model construction module is used to extract the discrete spatial coordinates of the current processing position in the three-dimensional solid coordinate system, assign the quality evaluation index value to the discrete voxel unit corresponding to the discrete spatial coordinate, and progressively update the discrete voxel unit carrying the quality evaluation index value to construct a three-dimensional quality digital twin model. The early warning and assessment output module is used to output online monitoring and early warning commands and global quality assessment results for the additive manufacturing process based on the three-dimensional quality digital twin model.

[0008] The present invention discloses the following technical effects: This invention provides a method and system for online monitoring and quality assessment of additive manufacturing processes based on multi-sensor fusion. Firstly, by employing a spatial mapping mechanism between real-time trajectory coordinates and system timestamps, this invention completely eliminates the spatiotemporal registration misalignment problem between multi-source heterogeneous sensor data, achieving high-precision spatiotemporal alignment of underlying data. Secondly, by introducing a dynamic cross-modal attention network to calculate the mutual information correlation matrix and reconstruct features, it overcomes the limitations of shallow cross-modal feature fusion in existing systems, deeply mining the nonlinear complementary information between various physical fields and significantly improving the defect identification accuracy under complex working conditions. Finally, by accurately assigning quality assessment index values ​​to discrete voxel units to construct a three-dimensional quality digital twin model, a voxel-level three-dimensional quality spatial mapping and traceability mechanism is successfully established, completely solving the problem of lagging two-dimensional assessment and providing intuitive three-dimensional spatial guidance and closed-loop control for the additive manufacturing process. Attached Figure Description

[0009] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0010] Figure 1 The flowchart illustrates an online monitoring and quality assessment method for additive manufacturing processes based on multi-sensor fusion, as provided in this embodiment of the invention. Detailed Implementation

[0011] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0012] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0013] like Figure 1 As shown, this invention provides a method for online monitoring and quality assessment of additive manufacturing processes based on multi-sensor fusion, including: Step 100: Synchronously acquire the multi-source heterogeneous sensor data stream of the processing area, as well as the real-time trajectory coordinates and system timestamp of the current processing position; Step 200: Based on the real-time trajectory coordinates and the system timestamp, perform time-series synchronization and spatial mapping on the multi-source heterogeneous sensor data stream to construct a multimodal spatiotemporal dataset; Step 300: Extract features from each physical modality data in the multimodal spatiotemporal dataset to obtain the corresponding independent feature representation set; Step 400: Input the independent feature representation set into a preset dynamic cross-modal attention network to calculate the mutual information correlation matrix between nodes within the independent feature representation set, and perform feature reconstruction and adaptive weight allocation on the independent feature representation set based on the mutual information correlation matrix to obtain a deep fusion feature vector; Step 500: Input the deep fusion feature vector into a preset quality mapping model and output the quality evaluation index value corresponding to the current processing position, wherein the quality evaluation index value represents the defect probability and defect classification of the current processing position; Step 600: Extract the discrete spatial coordinates of the current processing position in the three-dimensional solid coordinate system, assign the quality assessment index value to the discrete voxel unit corresponding to the discrete spatial coordinate, and progressively update the discrete voxel unit carrying the quality assessment index value to construct a three-dimensional quality digital twin model. Step 700: Based on the three-dimensional quality digital twin model, output online monitoring and early warning commands for the additive manufacturing process and global quality assessment results.

[0014] Furthermore, the specific implementation process of step 100 is as follows: This embodiment first constructs a global hardware triggering network at the physical entity level, independent of the additive manufacturing equipment's control host. This embodiment uses a field-programmable gate array (FPGA) as the main control generator, sending nanosecond-precision crystal clock synchronization trigger signals to various sensor acquisition nodes distributed around the processing area via shielded twisted-pair cables. This physical hard-wired triggering mechanism completely bypasses the software scheduling delays and network communication jitter of general-purpose operating systems, ensuring that all heterogeneous sensors simultaneously initiate exposure or sampling actions at the microsecond-level instant of receiving the rising edge of the same physical pulse. This fundamentally eliminates the timing misalignment caused by the start-up time difference of multimodal data, achieving absolute coordination of the underlying acquisition actions.

[0015] Under the hard triggering of the aforementioned unified physical pulse, this embodiment coordinates the use of heterogeneous physical sensing entity nodes such as acoustic emission sensors, high-speed industrial cameras, and infrared thermal imagers. After capturing instantaneous physical quantities, each sensing node utilizes its built-in edge computing microprocessor to pre-allocate high-speed caches and reorganize the feature formats of massive amounts of raw acoustic waveforms and high-frame-rate thermal image sequences. This embodiment aggregates and packages data streams from different physical modes according to a unified bit width and data frame structure, forming a continuous data stream containing multimodal underlying physical attributes. This avoids queuing and congestion caused by inconsistent transmission rates and interface bandwidths of various sensors at the data transmission channel level, ensuring high-throughput, real-time, lossless convergence of multi-source heterogeneous sensor data streams.

[0016] To achieve precise spatiotemporal anchoring of multimodal sensing data with the physical processing space, this embodiment directly captures feedback pulse data from the axis encoders of each servo motor at the underlying layer of the motion controller at extremely high frequency via an industrial real-time communication bus. This embodiment then performs forward kinematic matrix calculation based on the physical geometric parameters of the actuators, converting the instantaneous angular or linear displacement pulses of each axis into three-dimensional absolute coordinates in a reference coordinate system. Subsequently, this embodiment extracts the precise absolute time at the moment the global clock synchronization pulse is generated as the system timestamp, and forcibly burns this system timestamp into the custom protocol bits of the data frame header of the multi-source heterogeneous sensing data stream and into the data packet containing the three-dimensional absolute coordinates, thus completing the high-concurrency and hard binding of physical coordinates, physical time, and multi-field physical signals on the underlying data link.

[0017] Specifically, the preset reference sampling frequency ranges from 10,000 Hz to 1,000 Hz, and the preset reference sampling frequency is greater than or equal to twice the highest effective frequency component of the highest frequency transient physical signal in the multi-source heterogeneous sensing data stream.

[0018] This embodiment establishes a stringent absolute physical sampling boundary to address the extremely short time window characteristic of the evolution of microscopic defects in the additive manufacturing processing area. This embodiment strictly locks the value range of the preset reference sampling frequency between 10,000 Hz and 1,000 Hz to perfectly adapt to the transient evolution rate of the cross-scale physical field data stream. In actual forming physical processing, the evolution cycle of physical phenomena such as metal splashing and intense thermal fluid turbulence on the molten pool surface is extremely short. Therefore, the lower limit of acquisition is set at 10,000 Hz to ensure that the high-speed visual imaging sensor hardware has sufficiently dense continuous optical observation resolution. Simultaneously, the acoustic emission high-frequency elastic stress waves accompanying the cracking of microscopic metal grains and the intense release of phase transformation stress often have a center frequency band as high as hundreds of thousands of Hz. Therefore, this embodiment pushes the upper limit of frequency sampling to 1,000 Hz by deploying an ultra-high frequency analog-to-digital converter physical chip, thereby thoroughly capturing the very early acoustic precursor physical characteristics of all dense defects on the physical signal acquisition base.

[0019] While setting the absolute physical acquisition base bandwidth as described above, this embodiment further introduces a relative physical constraint mechanism to prevent high-frequency signal aliasing. This embodiment extracts the frequency domain energy distribution characteristics of each heterogeneous mode signal in the multi-source heterogeneous sensor data stream through built-in low-level digital signal processing logic hardware, and accurately locks and tracks the highest effective frequency component of the highest frequency transient physical signal. To completely and losslessly preserve the true time-domain waveform characteristics of this extreme high-frequency physical event during the sampling and quantization process of converting analog physical quantities to discrete digital quantities, this embodiment mandates at the hardware clock distribution logic layer that the preset reference sampling frequency must be greater than or equal to twice that of the highest effective frequency component. This hardware-level physical frequency redundancy matching mechanism effectively eliminates the underlying physical risks of spectral folding and waveform distortion of high-frequency transient signals during discretization and dimensionality reduction acquisition.

[0020] This embodiment constructs a high-fidelity multi-source signal underlying physical mapping barrier through the dual physical entity constraint mechanism of the absolute high-frequency range and relative dynamic multiplication rate. This embodiment utilizes the high-precision clock multiplication and division circuit unit of the phase-locked loop inside the main control field-programmable gate array to directly convert the strictly preset reference sampling frequency into a hard-wired trigger clock cycle that drives the high-speed acquisition hardware chip of each heterogeneous sensing node in the front end. This wide-bandwidth and strictly anti-aliasing physical entity collaborative sampling architecture completely blocks the omission or missed detection of extremely transient micro-defect features caused by hardware undersampling from the source of physical information. This provides a continuous underlying physical observation slice with extremely high signal-to-noise ratio for the subsequent construction of multimodal spatiotemporal datasets and deep fusion of cross-modal features, ultimately ensuring the industrial-grade output reliability of quality assessment indicators under complex operating conditions.

[0021] Furthermore, the specific implementation process of step 200 is as follows: This embodiment first addresses the timing acquisition deviation caused by the inherent sampling frequency differences of various physical modal devices within a multi-source heterogeneous sensing data stream by executing underlying physical time axis reconstruction and alignment logic. This embodiment parses the underlying communication messages, extracts the local hardware timestamps attached to the headers of each physical modal data frame, and establishes the precise system timestamp captured by globally triggered hardware as a globally unified absolute time reference. To compensate for the physical signal blind spots caused by asynchronous discrete sampling within extremely short transient time slices, this embodiment employs a time interpolation resampling algorithm based on high-order polynomial fitting to materialize the logic, performing temporal smoothing and continuous reconstruction of the transient elastic waveforms or instantaneous fused pool pixel grayscale change rates of each physical modal data. This process forcibly resamples and maps the asynchronous discrete physical sampling nodes of different underlying sensing devices to the same microsecond-level absolute physical time baseline, thereby generating a timing-aligned sensing data set in the memory physical address that completely eliminates the transmission delay differences and inherent sampling rate gaps between heterogeneous devices, effectively ensuring the absolute time synchronization prerequisite before cross-field physical feature fusion calculation.

[0022] Based on establishing a rigorous underlying physical time anchor, this embodiment calls upon pre-programmed calibration parameters in non-volatile memory to read a spatial pose transformation matrix encompassing the relative geometric relationships between the field of view of each sensing physical entity and the plane of the additive manufacturing machine tool forming process entity. This embodiment utilizes hardware geometric computation acceleration operators within the edge computing node to perform precise inverse physical space projection calculations of coordinate system rotation and rigid body translation based on this spatial pose transformation matrix. This involves processing various discrete modal data stored in the time-aligned sensing data set, such as the two-dimensional pixel coordinate stream of a high-speed vision camera and the one-dimensional vibration energy signal sequence of acoustic physical entities, through coordinate system rotation and rigid body translation. This cross-dimensional underlying geometric projection operation forcibly transforms and normalizes heterogeneous sensing physical quantities from different observation perspectives and heterogeneous acquisition dimensions into a three-dimensional entity coordinate system preset based on the machine tool's physical origin. This derives a spatial mapping feature stream endowed with prior information on absolute three-dimensional physical spatial position, greatly eliminating the geometric interference and misleading effects of lens perspective distortion and physical installation position offset on subsequent three-dimensional spatial tracing of microscopic defects.

[0023] This embodiment further uses the real-time trajectory coordinates generated by the physical feedback analysis of the motion actuator as the spatial anchor point index of the solid processing and forming area, and performs tensor encapsulation and splicing weaving logic of high-dimensional data features in the physical memory block. This embodiment performs spatial gridding binding operation on the spatial mapping feature stream output by the above operation according to the physical spatial anchor point index, and uses the microprocessor's low-level continuous memory addressing mechanism to complete the multi-dimensional tensor splicing of cross-modal heterogeneous physical quantities in the feature dimension. Subsequently, this embodiment dynamically extracts and isolates a three-dimensional spatial physical data window in the solid digital memory space with the current physical processing laser cladding point as the absolute center and the boundary geometric dimensions strictly defined. Finally, it outputs a multi-modal spatiotemporal dataset that has multi-dimensional transient physical properties and achieves perfect closed-loop binding in the absolute time reference axis and the three-dimensional solid space network. This provides a low-level solid perception data base with extremely high fidelity and holographic traceability for the next stage of capturing the strong coupling precursors of transient multiphysics fields generated by internal dense defects.

[0024] Furthermore, the specific implementation process of step 300 is as follows: This embodiment first parses the underlying data structure attributes encapsulated within the pre-built multimodal spatiotemporal dataset in the underlying data flow architecture. Utilizing the direct memory access controller built into the edge computing motherboard, this embodiment explores the data bit width and permutation tensor dimension of each physical modality data stream, and then adaptively divides each physical modality data stream into a one-dimensional temporal signal stream and a multi-dimensional spatial image stream. To achieve underlying hardware-level parallel processing for heterogeneous physical field data, this embodiment instantiates and constructs a parallel backbone network containing one-dimensional temporal feature extraction branches and multi-dimensional spatial feature extraction branches in the high-speed video memory of the graphics processing unit or tensor processing unit. This data splitting and parallel physical computing topology design based on underlying data structure attributes fundamentally avoids video memory read / write collisions and serial queuing delays caused by heterogeneous data in the same central processing unit computing thread, greatly improving the throughput rate of massive physical quantities and the transient computing response bandwidth in online additive manufacturing monitoring scenarios.

[0025] After establishing the underlying parallel computing link, this embodiment inputs the split one-dimensional temporal signal stream into the pre-deployed one-dimensional temporal feature extraction branch to extract temporal dynamic features characterizing transient changes in the manufacturing process, and synchronously inputs the multi-dimensional spatial image stream into the multi-dimensional spatial feature extraction branch to extract spatial morphological features characterizing the evolution of the processing area's morphology. Specifically, this embodiment utilizes the physical sliding of a one-dimensional solid convolution kernel to capture the extremely transient envelope abrupt changes and decaying oscillations of high-frequency acoustic emission stress waves, while simultaneously utilizing two-dimensional multi-scale convolution residual blocks to capture the morphological distortion features of the molten pool surface thermal radiation temperature gradient distribution and the geometric contour of metal spatter. This decoupled parallel feature extraction mechanism, oriented towards specific manufacturing physical field characteristics, not only completely preserves the original single-field physical evolution laws captured by each independent sensing device, but also eliminates a large amount of environmental background physical noise unrelated to the evolution of microscopic defects in the shallow solid computing units of the network, achieving high-fidelity abstraction and purification of the underlying dense defect precursor physical features.

[0026] To address the technical challenge of a significant physical gap in the underlying tensor dimension and numerical distribution scale of the physical features extracted by the parallel branches, this embodiment further utilizes a pre-defined nonlinear mapping layer configured at the end of each extraction branch to perform channel dimension alignment and deep semantic space projection on the temporal dynamic features and the spatial morphological features, respectively. This embodiment uses a fully connected network node matrix within the physical computing nodes and an activation gating function to forcibly map heterogeneous features of varying lengths and channel depths to an implicit high-dimensional feature basis with a unified fixed physical dimension, thereby obtaining individual single-modal high-dimensional feature vectors that eliminate differences in the underlying physical domain. Subsequently, this embodiment calls memory concatenation instructions in the cache block to perform pointer-level concatenation and combination of all individual single-modal high-dimensional feature vectors according to a fixed modal channel displacement sequence, ultimately statically solidifying them at the physical memory level to generate the independent feature representation set for downstream network use. This multidimensional spatial alignment and underlying memory splicing entity operation completely overcomes the physical barriers of the underlying analog signals of heterogeneous sensors, and builds a standardized high-dimensional digital array base with regular dimensions and unified underlying semantics for subsequent deep mutual information association mining computing networks.

[0027] Furthermore, the specific implementation process of step 400 is as follows: This embodiment first inputs the set of independent feature representations into a pre-defined dynamic cross-modal attention network deployed in high-performance tensor-accelerated physical hardware. To decouple the intrinsic correlation representations of heterogeneous physical signals in the underlying feature space, this embodiment utilizes pre-initialized query mapping matrices, key mapping matrices, and value mapping matrices within the pre-defined dynamic cross-modal attention network to perform high-concurrency matrix linear multiplication operations with the set of independent feature representations in high-speed video memory. This hardware-level large-scale matrix multiplication operation, without disrupting the original distribution of underlying physical features, forcibly projects single-modal high-dimensional features from heterogeneous physical sensors such as acoustic, optical, and thermal sensors into a unified metric latent space. This accurately extracts and dynamically generates query feature vectors, key feature vectors, and value feature vectors corresponding to each physical modality, providing a fundamental digital tensor foundation with absolute alignment in dimensionality and semantic scale for subsequent cross-modal interactive comparisons between heterogeneous physical field features.

[0028] After establishing a unified feature tensor space, this embodiment calculates the underlying tensor inner product metric between the query feature vector and the key feature vector corresponding to any two different physical modes, targeting the nonlinear coupling evolution mechanism of strong transient multiphysics fields during processing. This embodiment calls the physical tensor core to perform the underlying dot product similarity calculation operation and uses a hardware-accelerated exponential normalization function to perform a nonlinear probability distribution mapping on the dot product similarity, thereby constructing the mutual information correlation matrix at the physical feature layer that quantifies the information dependency strength between heterogeneous sensor sensing channels. Subsequently, this embodiment performs a high-dimensional tensor matrix multiplication operation between the mutual information correlation matrix and the corresponding value feature vector. This cross-modal aggregation calculation process of physical entities enables a single physical mode to adaptively absorb implicit complementary information from other heterogeneous physical fields in the underlying feature space, thereby accurately filtering out local independent physical background noise unrelated to defect evolution and inferring and outputting a highly purified enhanced feature representation in the memory block.

[0029] To prevent the submergence of critical weak precursor signals of specific physical fields during the deep aggregation of cross-modal features, this embodiment further imports the enhanced feature representation into an adaptive gated network unit preset within the backbone network layer. This embodiment utilizes the fully connected computation nodes and nonlinear activation hardware operators within the adaptive gated network unit to dynamically evaluate the confidence level of the physical contribution of each modal enhanced feature to the current microscopic defect state mapping, thereby accurately generating the adaptive fusion weights. This embodiment then uses the underlying multiplication and accumulation physical operation unit to perform tensor-level dynamic weighting processing on the enhanced feature representation according to the adaptive fusion weights. In the high-speed cache of the graphics processing unit, the dynamically weighted features and the original input set of independent feature representations are summed at the underlying physical memory address and concatenated with the channel dimension. This materialized residual cross-modal fusion architecture completely overcomes the risks of hardware computation degradation and gradient decay in deep physical feature reconstruction networks, perfectly balancing cross-modal global complementary perception and single-field local feature fidelity, and finally statically solidifies and outputs the deep fusion feature vector for high-precision defect space mapping.

[0030] Specifically, the expression for calculating the mutual information correlation matrix is ​​as follows: Among them, among them, The number of physical modes; For the first The high-dimensional feature vector of a single mode corresponding to each physical mode; The dimension of the single-modal high-dimensional feature vector; In order to make each Independent feature representation matrix formed by modal stacking; To be A learnable mapping matrix that linearly maps to the query subspace; To be A learnable mapping matrix that linearly maps to the key space; To query the feature matrix; The key feature matrix; To query the key subspace dimension; This is the cross-modal dot product similarity matrix; For row-wise exponential normalization functions; Mutual information correlation matrix; The expression for the enhanced feature representation is: ; in, The matrix represents the independent features; To be A learnable mapping matrix that linearly maps to a value subspace; The characteristic matrix is ​​a value; Mutual information correlation matrix; To enhance the feature representation matrix; The formula for calculating the adaptive fusion weights is: ; in, To enhance the feature representation matrix; The matrix represents the independent features; For concatenation operators along the feature dimension; The learnable matrix for adaptive gated mapping; A column vector of all 1s is used to broadcast the bias to each modal node; For the learnable bias vector of the adaptive gating mapping; It is a Sigmoid gated function; For adaptive fusion weight matrix; For element-wise multiplication operators; The fused feature matrix; MeanPool ( ) is a pooling operator that averages the modal node dimensions; Concat( ) is the vector concatenation operator; This is a deep fusion feature vector.

[0031] More specifically, to meet the deep coupling requirements of the underlying physical tensor features, this embodiment implements a strict hierarchical physical topology definition for the preset dynamic cross-modal attention network within the high-speed memory block of the tensor acceleration physical hardware. This embodiment sets the initial physical channel dimension of the input independent feature representation matrix to a fixed tensor depth strongly correlated with the sampling accuracy of the front-end heterogeneous physical sensor, and instantiates and constructs multi-layered cascaded linear fully connected network nodes within the computing unit as the physical carrier for the learnable mapping matrix. For the query mapping matrix, key mapping matrix, and value mapping matrix declared in the underlying formula, this embodiment adopts a normal distribution-based random initialization strategy for physical memory to assign initial level parameters to the learnable network weights, thereby breaking the symmetry of the underlying computing nodes. This network architecture design, accurate to the hardware dimension, and parameter memory initialization mechanism lay a complete foundation for the underlying computation graph topology entity for subsequent high-concurrency tensor multiplication calculations of massive cross-modal physical quantities, enabling seamless mapping of data from any heterogeneous physical sensing channel to the specified implicit physical metric space for underlying feature alignment.

[0032] In the specific physical application scenario of additive manufacturing transient forming, this embodiment profoundly reveals the intrinsic mapping and physical resonance logic between multi-source heterogeneous physical field input data and microscopic quality output evaluation through the aforementioned cross-modal physical operations. This embodiment abstractly fuses acoustic vibration signals, optical thermal radiation signals, and molten pool visual morphology signals into the independent feature representation set, which serves as the physical input data stream for the underlying algorithm's computational nodes. When dense physical defects such as incomplete fusion or microscopic pores occur locally in the molten pool, they inevitably trigger a sharp change in the local thermal temperature gradient and a sudden jump in the transient acoustic emission high-frequency elastic stress wave on the underlying processed entity surface. This embodiment utilizes a cross-modal dot product similarity matrix to quantify and capture this physical abrupt resonance correlation phenomenon of different physical modes within the same microsecond-level absolute time slice at the underlying hardware layer. Furthermore, it uses an enhanced feature representation matrix to achieve cross-complementarity of physical field information and encapsulates the fusion operation result into the deep fusion feature vector as the deterministic output of the physical operator, thereby achieving precise computational penetration from discrete surface physical representation to the deep intrinsic dense defect evolution mechanism.

[0033] To address the massive amount of learnable hardware parameters contained within the aforementioned cross-modal physical feature reconstruction network, this embodiment utilizes a pre-collected dataset of additive manufacturing standard sample defects, annotated by experts from the physical processing platform, to execute a rigorous underlying physical gradient backpropagation entity training step within a high-performance computing cluster. This embodiment employs an adaptive moment estimation optimization hardware operator, using a preset small floating-point number as the initial base learning rate in the physical memory block, and employing a cosine annealing learning rate decay strategy to precisely control the physical convergence step size of the underlying weight updates. In each iteration of the physical computation involving forward tensor propagation and backpropagation error differentiation, this embodiment calculates the tensor error gradient between the predicted output of the physical node and the actual physical defect label based on the multidimensional cross-entropy loss function. This physical gradient is then used to backpropagate the underlying multiplication and accumulation physical operation unit, performing entity weight overwriting and parameter approximation based on physical memory addresses for core operation nodes such as the query mapping matrix, key mapping matrix, and the learnable matrix of the adaptive gating mapping. This materialized closed-loop gradient update architecture ensures that the model can accurately converge to the globally optimal physical feature representation manifold domain under massive real-world manufacturing conditions, significantly improving the model's robustness in low-level inference in unknown and complex manufacturing environments.

[0034] Furthermore, the specific implementation process of step 500 is as follows: This embodiment first imports the deeply fused feature vector statically output by the front-end module directly into the quality mapping model pre-deployed in the underlying tensor processing entity unit. To mitigate the dimensionality curse and underlying memory overflow risks that are easily caused by high-dimensional cross-modal physical features in direct classification and regression mapping, this embodiment utilizes the nonlinear feature decoding network within the pre-defined quality mapping model to perform high-dimensional feature space dimensionality reduction and quality semantic information abstraction on the input underlying tensor. The layer-by-layer pooling and fully connected dimensionality reduction operations of this physical entity, through the nonlinear mapping mechanism of hardware-accelerated operators, accurately remove redundant physical field background noise and non-critical morphological fluctuations within the deeply fused feature vector. This purifies and statically generates a highly condensed global quality state vector specifically for inferring the physical state of microscopic defects within the physical memory block, significantly reducing the underlying computational load and physical memory bandwidth consumption of subsequent parallel inference branches.

[0035] Based on the multi-threaded concurrent processing architecture of the underlying hardware, this embodiment performs a lossless copy and synchronous splitting of the extracted global quality state vector in physical memory, and then inputs it in parallel into the defect probability prediction branch and the defect classification branch within the preset quality mapping model. In the concurrently executed inference computation physical link, this embodiment uses the logistic regression hardware operator within the defect probability prediction branch to perform nonlinear activation mapping operations on the global quality state vector to accurately output a defect occurrence probability value that quantitatively characterizes the possibility of micro-physical defects occurring at the current physical processing cladding point. At the same time, this embodiment uses the tensor multiplication core within the defect classification branch to accurately project the global quality state vector onto a preset defect type distribution space. By performing a high-throughput entity dot product comparison with various known typical micro-defect feature matrices pre-programmed in the underlying memory, a multi-dimensional confidence score is calculated. Then, through hardware-level extreme value addressing logic, the category corresponding to the highest confidence score is hard-determined as the defect classification label under the current physical slice coordinates. This underlying concurrent dual-branch hardware inference architecture completely breaks through the latency bottleneck of traditional serial evaluation algorithms, meeting the stringent time constraints of microsecond-level online quality evaluation in additive manufacturing.

[0036] After completing the aforementioned dual-branch underlying physical tensor concurrent inference, this embodiment performs strict memory alignment and physical assembly operations on the output discrete and continuous heterogeneous inference results at the register level of the edge computing microprocessor. This embodiment calls underlying bit operation instructions to perform data structure concatenation and underlying protocol-level joint encoding of the defect occurrence probability value, representing continuous quantization risk, and the defect classification label, representing discrete physical attributes. This materialized data encapsulation mechanism forcibly bridges the data type gap between probabilistic floating-point numbers and categorical integer sequences in physical memory space, thereby fusing and generating the quality assessment index value that possesses both continuous probability attributes and discrete category physical attributes. This highly integrated multidimensional assessment data stream is directly pushed into the downstream 3D twin modeling cache queue via an industry-standard bus, not only realizing a digital holographic physical mapping of the microscopic quality state of the current processing position in the processing area, but also completely eliminating the underlying data structure barriers in the process of transferring multimodal monitoring data to the 3D physical space mapping layer.

[0037] More specifically, to meet the review requirements for the full disclosure of artificial intelligence algorithms, this embodiment defines an extremely rigorous hierarchical structure and solidifies physical connections for the preset quality mapping model in the underlying edge computing hardware. This embodiment instantiates the nonlinear feature decoding network as a deep residual topology containing multiple cascaded physical convolutional layers and batch normalization hardware operators, and directly connects its tail to the defect probability prediction branch and the defect classification branch, which contain multilayer perceptron computing nodes. This physical architecture design, which directly opens a common feature dimensionality reduction base pool in the underlying physical memory and uses a hardware bus to distribute global quality state vectors in parallel to two independent fully connected physical inference branches, thoroughly clarifies the tensor flow logic and memory interaction network from the input high-dimensional feature flow to the dual-ended output nodes. This enables the model to possess extremely high underlying computing power determinism and physical structure reproducibility when dealing with the strong transient defect mapping in additive manufacturing.

[0038] This embodiment addresses the practical engineering application scenarios where additive manufacturing processes are prone to sudden microscopic pores or macroscopic cracks, revealing the physical mapping and intrinsic correlation between the input and output data of the underlying artificial intelligence algorithm model. This embodiment uses the deeply fused feature vector, which is a physical fusion of upstream multi-sensor hardware with an extremely high signal-to-noise ratio, as the physical tensor input. This input data is essentially a digitally reduced-dimensional holographic physical representation of local abnormal thermal gradients and high-frequency acoustic bursting signals. When this highly refined physical feature is input to the quality mapping model, the underlying physical operator is activated through nonlinear mapping and outputs the probability value of the defect occurrence, representing the risk of microscopic defects, as well as the defect classification label for the specific physical type. This data setting mechanism, which strongly binds the underlying heterogeneous physical perception features with the internal physical quality state of the formed entity across domains, perfectly matches the underlying logic of defect evolution in complex multi-physics environments, providing a solid high-dimensional data mapping physical anchor for subsequent digital twin modeling of the entity.

[0039] To address the massive nonlinear physical mapping weights and hardware gating bias parameters contained in the underlying quality mapping model, this embodiment utilizes a multimodal entity processing defect dataset pre-labeled with physical ground truth values ​​using an industrial computed tomography (CT) scanner. A multi-task joint physical gradient training step is executed on a graphics processing unit (GPU) cluster. This embodiment performs a weighted summation in physical memory between the physical calculation logic of the binary cross-entropy loss function for the probability prediction dimension and the entity hardware operator of the multi-classification focus loss function for the classification dimension, constructing a global multi-task joint physical loss quantity to guide the updating of underlying hardware parameters. During entity backpropagation training, this embodiment sets a fixed floating-point physical baseline learning rate and uses an adaptive momentum optimization operator to calculate the physical error gradient for the decoding network and parallel dual-branch hierarchical nodes. Then, through underlying memory overwrite instructions, iterative approximation updates are performed on the learnable weight matrix within each physical inference branch until the global physical error loss converges below a preset tolerance threshold. This endows the model with the underlying generalization inference physical capability for high-fidelity defect diagnosis under harsh real-world processing conditions.

[0040] Furthermore, the specific implementation process of step 600 is as follows: This embodiment first allocates a contiguous physical memory addressing block in the high-speed dynamic random access memory of the underlying industrial control computer to carry the digital holographic mapping base of the part to be formed. This embodiment reads the pre-imported 3D computer-aided design model by calling the underlying graphics parsing hardware operator, and uses physical boundary extraction logic based on extreme coordinate traversal to obtain the global 3D bounding box of the 3D computer-aided design model in the absolute 3D solid coordinate system of the machine tool. To transform the contiguous physical forming space into a discrete digital space that the computer can execute low-level read / write operations, this embodiment performs orthogonal meshing discretization of the global 3D bounding box at the underlying physical memory level according to a preset 3D spatial resolution that strictly matches the accuracy of industrial non-destructive testing. This physical discretization operation instantiates and constructs a 3D voxelized basic framework containing a massive number of initial hollow voxels in the underlying video memory block. Each hollow voxel is forcibly assigned a unique physical memory address pointer by the underlying operating system, thus providing a solidified digital mesh skeleton with absolute physical size constraints for subsequent high-precision spatial anchoring of micro-defects.

[0041] After constructing the aforementioned underlying voxelized digital skeleton, this embodiment performs a hard real-time dimensionality reduction mapping from physical coordinates to memory addresses, addressing the strong transient movement characteristics of the laser molten pool in additive manufacturing. This embodiment extracts the continuous physical coordinate values ​​of the current processing position in the three-dimensional solid coordinate system from the underlying motion control bus at high frequency. These continuous coordinates are essentially absolute spatial physical positions containing floating-point precision. To eliminate the memory overhead and precision drift caused by underlying floating-point addressing, this embodiment utilizes a spatial discretization operator deployed within a programmable gate array to perform a fast physical space mapping based on floor division and shift operations on these continuous physical coordinate values. This mapping is then forcibly projected onto the three-dimensional integer index space of the three-dimensional voxelized framework, resulting in a three-dimensional mesh index representing the discrete spatial coordinates. This embodiment then uses the three-dimensional mesh index as the absolute physical base address offset for underlying memory addressing, performs ultra-fast addressing in the huge voxel memory block, accurately locates the corresponding target initial hollow voxel with extremely low latency, and hard-confirms the physical memory node as the discrete voxel unit that carries the microscopic quality state of the current processing position, thus completely breaking down the spatial physical mapping barrier between the physical forming trajectory and the underlying digital memory mesh.

[0042] After successfully anchoring the physical mapping memory address, this embodiment executes the hard write of the underlying quality status data and the dynamic physical growth logic of the digital entity model. This embodiment uses low-level bit manipulation instructions to parse the physical attribute data within the quality evaluation index value inferred from the front end, encodes it into a multi-dimensional quality feature vector with a fixed byte length, and uses direct memory access technology to brute-force overwrite this multi-dimensional quality feature vector into the attribute memory block pre-allocated by the discrete voxel unit. Simultaneously, a low-level hardware interrupt is triggered, flipping the lifecycle status flag of the discrete voxel unit from an empty, inactive state to an entity-assigned state. With the layer-by-layer physical stacking of the additive manufacturing process, this embodiment, under the physical scheduling of the central processing unit, executes the above-mentioned spatial mapping and low-level attribute assignment steps in a high-frequency loop. Triggered by the physical pulse at the end of each processing layer, all discrete voxel units in the entity-assigned state are laterally topologically connected at the physical memory topology level to generate single-layer quality slice data containing the panoramic microscopic defect distribution of the current layer. Finally, in this embodiment, high-speed video memory is driven between adjacent physical processing layers to perform vertical memory stacking and stitching of all the single-layer quality slice data. Through layer-by-layer entity accumulation and dynamic memory expansion in three-dimensional digital space, the three-dimensional quality digital twin model is completely materialized and presented, which is completely consistent with the macroscopic spatial morphology of the actual physical processing part and has an internal holographic mapping of the quality distribution state of each micrometer-level grid.

[0043] Specifically, the expression for the three-dimensional quality digital twin model is: ; in, For time indexing; For a moment The current machining position in the continuous spatial coordinates of the three-dimensional solid coordinate system; This is a spatial discretization mapping function used to convert continuous coordinates... Mapped to voxel grid integer indices For a moment The corresponding voxel grid integer index; voxels The cumulative number of times the value has been assigned and updated; For a moment The joint encoding vector of the quality assessment index value corresponding to the current position; voxels Aggregated quality feature vector stored internally; Integer index of the voxel grid; It is a set of three-dimensional quality digital twin models used to represent the index of all activated voxels, their quality vectors, and update counts.

[0044] Furthermore, the specific implementation process of step 700 is as follows: This embodiment first establishes a high-frequency hardware polling and hard real-time interception mechanism for extremely transient microscopic quality mutations during the continuous growth of the underlying digital twin physical mesh. This embodiment uses a high-speed physical memory bus to extract in real time the most recently physically addressed and updated discrete voxel units in the 3D quality digital twin model, and then directly reads the quality assessment index values ​​encapsulated within the physical mesh node using underlying registers. To intervene with underlying computing power before irreversible macroscopic waste is generated during physical entity processing, this embodiment imports the defect probability representing the physical density risk within the extracted quality assessment index values ​​into a high-speed comparator of a field-programmable gate array (FPGA), and performs a nanosecond-level real-time sliding comparison calculation with a preset dynamic tolerance threshold stored in non-volatile memory. This high-frequency polling and hard-core comparison mechanism at the physical layer completely eliminates the software scheduling latency redundancy of the upper-layer operating system, ensuring absolute real-time monitoring and underlying interception barriers for the microscopic physical defect state of each micrometer-level forming space.

[0045] When the aforementioned high-speed hardware comparator determines at the physical level that the defect probability has hardened beyond the dynamic tolerance threshold, this embodiment immediately triggers a nanosecond-level underlying manufacturing physical anomaly hardware interrupt response mechanism, and rapidly encapsulates and generates the online monitoring and early warning command with strong real-time performance at the edge computing core layer. This embodiment then utilizes a real-time industrial communication physical link with deterministically low transmission latency to inject the online monitoring and early warning command into the underlying motion control board of the additive manufacturing equipment's control host at the millisecond level. This physical early warning data packet forcibly oversteps its authority to take over or interfere with the machine tool's original digital control machining code, thereby directly driving the underlying servo motor driver and high-energy laser generator hardware of the additive manufacturing equipment to perform closed-loop adaptive feedback actions at the physical level, such as instantaneously reducing laser output power or directly implementing an emergency stop at the current physical machining slice layer. This hard-wired closed-loop physical intervention architecture, which penetrates directly from the underlying memory mutation of the digital twin to the underlying hardware execution mechanism of the physical forming machine tool, greatly shortens the physical response time of abnormal defects. This completely blocks the macroscopic spread and deterioration accumulation of dense defects such as micropores or cracks between the three-dimensional solid processing layers at the source of physical forming.

[0046] After the physical additive manufacturing process of all physical slices is completed, or when the underlying data acquisition card receives a global physical evaluation trigger signal issued externally, this embodiment initiates a closed-loop logic for physical spatial panoramic tracing and macroscopic quality assessment of the entire digital twin base in the physical computing node. This embodiment utilizes the high-concurrency video memory read thread pool built into the graphics processor to completely and deeply traverse all discrete voxel units in the physical continuous memory address segment of the three-dimensional quality digital twin model that are in a state of being assigned values, and batch extracts the full set of quality assessment index values ​​bound to their underlying addresses. In a preset absolute three-dimensional solid coordinate system, this embodiment calls the underlying tensor parallel aggregation operator to spatially statistically summarize the global distribution density of different defect categories and the macroscopic cumulative defect volume ratio physical feature quantity. Then, combined with the macroscopic service force mechanical performance evaluation standard of the solid part pre-statically burned into non-volatile memory, the underlying multi-dimensional feature tensor comprehensive judgment logic is executed. Finally, this embodiment outputs the global quality assessment result, which includes a precise 3D solid visualization perspective view of internal micro-defects and an overall macroscopic solid forming pass rate judgment, within the high-speed rendering hardware physical pipeline of the graphics processor. This solidified panoramic digital 3D perspective not only achieves a complete and non-destructive perspective on the internal physical quality of the formed workpiece, but also provides high-fidelity closed-loop data guidance with absolute 3D spatial anchor points for the targeted iterative optimization of subsequent precision manufacturing process physical control parameters.

[0047] This embodiment also provides an online monitoring and quality assessment system for additive manufacturing processes based on multi-sensor fusion, including: The data synchronization acquisition module is used to synchronously acquire multi-source heterogeneous sensor data streams in the processing area, as well as the real-time trajectory coordinates and system timestamp of the current processing position; The multimodal spatiotemporal mapping module is used to perform temporal synchronization and spatial mapping on the multi-source heterogeneous sensor data stream based on the real-time trajectory coordinates and the system timestamp, so as to construct a multimodal spatiotemporal dataset. An independent feature extraction module is used to extract features from each physical modality data in the multimodal spatiotemporal dataset to obtain a corresponding set of independent feature representations. The cross-modal attention fusion module is used to input the independent feature representation set into a preset dynamic cross-modal attention network to calculate the mutual information correlation matrix between nodes within the independent feature representation set, and to perform feature reconstruction and adaptive weight allocation on the independent feature representation set based on the mutual information correlation matrix to obtain a deep fusion feature vector. The quality assessment mapping module is used to input the deep fusion feature vector into a preset quality mapping model and output the quality assessment index value corresponding to the current processing position, wherein the quality assessment index value represents the defect probability and defect classification of the current processing position; The three-dimensional twin model construction module is used to extract the discrete spatial coordinates of the current processing position in the three-dimensional solid coordinate system, assign the quality evaluation index value to the discrete voxel unit corresponding to the discrete spatial coordinate, and progressively update the discrete voxel unit carrying the quality evaluation index value to construct a three-dimensional quality digital twin model. The early warning and assessment output module is used to output online monitoring and early warning commands and global quality assessment results for the additive manufacturing process based on the three-dimensional quality digital twin model.

[0048] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0049] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A method for online monitoring and quality assessment of additive manufacturing processes based on multi-sensor fusion, characterized in that, include: Simultaneously acquire multi-source heterogeneous sensor data streams from the processing area, as well as the real-time trajectory coordinates and system timestamp of the current processing position; Based on the real-time trajectory coordinates and the system timestamp, the multi-source heterogeneous sensor data stream is time-series synchronized and spatially mapped to construct a multimodal spatiotemporal dataset. Feature extraction is performed on each physical modality data in the multimodal spatiotemporal dataset to obtain a corresponding set of independent feature representations; The independent feature representation set is input into a preset dynamic cross-modal attention network to calculate the mutual information correlation matrix between nodes within the independent feature representation set, and based on the mutual information correlation matrix, the independent feature representation set is reconstructed and adaptively weighted to obtain a deep fusion feature vector. The deep fusion feature vector is input into a preset quality mapping model, and the quality evaluation index value corresponding to the current processing position is output. The quality evaluation index value represents the defect probability and defect classification of the current processing position. Extract the discrete spatial coordinates of the current processing position in the three-dimensional solid coordinate system, assign the quality assessment index value to the discrete voxel unit corresponding to the discrete spatial coordinate, and progressively update the discrete voxel unit carrying the quality assessment index value to construct a three-dimensional quality digital twin model. Based on the aforementioned three-dimensional quality digital twin model, online monitoring and early warning commands and global quality assessment results for the additive manufacturing process are output.

2. The method for online monitoring and quality assessment of additive manufacturing processes based on multi-sensor fusion according to claim 1, characterized in that, The synchronous acquisition of multi-source heterogeneous sensor data streams of the processing area, as well as the real-time trajectory coordinates and system timestamp of the current processing position, includes: A global hardware triggering network independent of the additive manufacturing equipment control host is established, and a global clock synchronization triggering signal is sent to each sensor acquisition node through the global hardware triggering network according to a preset reference sampling frequency; Using the various sensing nodes deployed around the processing area, at least two physical mode data of the processing area are collected under the triggering of the same pulse edge of the global clock synchronization trigger signal, and the at least two physical mode data are aggregated and packaged into the multi-source heterogeneous sensing data stream; The feedback pulse data of the shaft encoder inside the motion controller of the additive manufacturing equipment is read at high frequency through the industrial real-time communication bus, and the feedback pulse data of the shaft encoder is analyzed through forward kinematics calculation to obtain the three-dimensional absolute coordinates of the current processing position, and the three-dimensional absolute coordinates are used as the real-time trajectory coordinates. The precise absolute time of generating the global clock synchronization trigger signal is extracted as the system timestamp, and the system timestamp is embedded into the data frame header protocol of the multi-source heterogeneous sensor data stream and the data packet of the real-time trajectory coordinates, respectively, to complete the strict timing binding of multi-source heterogeneous data at the physical acquisition layer.

3. The method for online monitoring and quality assessment of additive manufacturing processes based on multi-sensor fusion according to claim 1, characterized in that, The preset reference sampling frequency ranges from 10,000 Hz to 1,000 Hz, and the preset reference sampling frequency is greater than or equal to twice the highest effective frequency component of the highest frequency transient physical signal in the multi-source heterogeneous sensing data stream.

4. The method for online monitoring and quality assessment of additive manufacturing processes based on multi-sensor fusion according to claim 1, characterized in that, Based on the real-time trajectory coordinates and the system timestamp, the multi-source heterogeneous sensor data streams are time-series synchronized and spatially mapped to construct a multimodal spatiotemporal dataset, including: Local time stamps of each physical modality data in the multi-source heterogeneous sensing data stream are extracted, and the system timestamp is used as a global unified time reference. A time interpolation resampling algorithm is used to reconstruct and align the time axis of each physical modality data to obtain a time-aligned sensing data set. Read the pre-calibrated spatial pose transformation matrix of the multi-sensor field of view and the additive manufacturing processing plane, and accurately project the two-dimensional pixel coordinates and one-dimensional signal sequence in the temporally aligned sensing data set onto the preset three-dimensional solid coordinate system according to the spatial pose transformation matrix to obtain the spatial mapping feature flow. Using the real-time trajectory coordinates as spatial anchor point indexes, the spatial mapping feature stream is spatially gridded and bound and multidimensional tensor spliced ​​according to the spatial anchor point indexes, and a three-dimensional spatial physical data window centered on the current processing position is extracted to obtain the multimodal spatiotemporal dataset containing multidimensional physical attributes and fully aligned in time and space.

5. The method for online monitoring and quality assessment of additive manufacturing processes based on multi-sensor fusion according to claim 1, characterized in that, Feature extraction is performed on each physical modality data in the multimodal spatiotemporal dataset to obtain a corresponding independent feature representation set, including: Based on the underlying data structure attributes, the physical modal data are adaptively divided into a one-dimensional time-series signal stream and a multi-dimensional spatial image stream. Construct a parallel backbone network that includes a one-dimensional temporal feature extraction branch and a multi-dimensional spatial feature extraction branch; The one-dimensional time-series signal stream is input into the one-dimensional time-series feature extraction branch to extract time-series dynamic features; The multidimensional spatial image stream is input into the multidimensional spatial feature extraction branch to extract spatial morphological features; By using a preset nonlinear mapping layer, the temporal dynamic features and the spatial morphological features are respectively aligned in the channel dimension and projected into the deep semantic space to obtain each single-modal high-dimensional feature vector. All the single-modal high-dimensional feature vectors are then concatenated and combined to form the independent feature representation set.

6. The method for online monitoring and quality assessment of additive manufacturing processes based on multi-sensor fusion according to claim 1, characterized in that, The process involves inputting the set of independent feature representations into a preset dynamic cross-modal attention network to calculate the mutual information correlation matrix between nodes within the set of independent feature representations. Based on the mutual information correlation matrix, feature reconstruction and adaptive weight allocation are performed on the set of independent feature representations to obtain a deep fusion feature vector, including: Using the query mapping matrix, key mapping matrix, and value mapping matrix within the preset dynamic cross-modal attention network, linear multiplication is performed with the independent feature representation set to extract and generate query feature vectors, key feature vectors, and value feature vectors corresponding to each physical modality data. Calculate the dot product similarity between the query feature vector and the key feature vector corresponding to any two different physical modes, and use the exponential normalization function to perform probability distribution mapping on the dot product similarity to construct the mutual information association matrix; The mutual information correlation matrix is ​​multiplied with the corresponding value feature vector to obtain the enhanced feature representation; The enhanced feature representation is input into a preset adaptive gating network unit to generate adaptive fusion weights; The enhanced feature representation is dynamically weighted according to the adaptive fusion weights, and the dynamically weighted features are summed with the independent feature representation set by residual summation and dimension concatenation to obtain the deep fusion feature vector.

7. The method for online monitoring and quality assessment of additive manufacturing processes based on multi-sensor fusion according to claim 1, characterized in that, The step of inputting the deep fusion feature vector into a preset quality mapping model and outputting the quality evaluation index value corresponding to the current processing position includes: The deep fusion feature vector is input into the nonlinear feature decoding network inside the preset quality mapping model to perform dimensionality reduction of the high-dimensional feature space and abstraction of quality semantic information to obtain a global quality state vector. The global quality state vector is synchronously split and input in parallel into the defect probability prediction branch and the defect classification branch within the preset quality mapping model; The defect probability prediction branch is used to perform a nonlinear activation mapping operation on the global quality state vector to obtain the defect occurrence probability value; The global quality state vector is projected onto a preset defect type distribution space using the defect classification branch to calculate the confidence score with each known typical micro-defect feature, and the category corresponding to the highest confidence score is determined as the defect classification label. The probability value of the defect occurrence and the defect classification label are concatenated and jointly encoded using a data structure to generate the quality assessment index value, which has both continuous probability attributes and discrete category attributes.

8. The method for online monitoring and quality assessment of additive manufacturing processes based on multi-sensor fusion according to claim 1, characterized in that, The step of extracting the discrete spatial coordinates of the current processing position in a three-dimensional solid coordinate system, assigning the quality assessment index value to the discrete voxel unit corresponding to the discrete spatial coordinates, and progressively updating the discrete voxel unit carrying the quality assessment index value to construct a three-dimensional quality digital twin model includes: Read the three-dimensional computer-aided design model of the part to be formed, and obtain the global three-dimensional bounding box of the three-dimensional computer-aided design model in the three-dimensional solid coordinate system; The global three-dimensional bounding box is orthogonally meshed and discretized according to the preset three-dimensional spatial resolution to establish a three-dimensional voxelized basic framework containing several initial hollow voxels. Extract the continuous physical coordinate values ​​of the current processing position in the three-dimensional solid coordinate system, and use a spatial discretization algorithm to map the continuous physical coordinate values ​​to the three-dimensional integer index space of the three-dimensional voxelized basic framework to obtain a three-dimensional mesh index representing the discrete spatial coordinates; The target initial hollow voxel is precisely located according to the three-dimensional mesh index, and the target initial hollow voxel is determined as the discrete voxel unit corresponding to the discrete spatial coordinates; The attribute data is encoded into a multidimensional quality feature vector, and the multidimensional quality feature vector is written into the attribute memory block of the discrete voxel unit. The state of the discrete voxel unit is updated from the hollow inactive state to the entity assigned a value state. The spatial mapping and attribute assignment steps are executed repeatedly, and all discrete voxel units in the entity-assigned state are horizontally connected within the same processing layer to generate single-layer quality slice data. All the single-layer quality slice data are stacked vertically between adjacent processing layers, and the three-dimensional quality digital twin model is obtained by layer-by-layer accumulation in three-dimensional space and dynamic memory update.

9. The method for online monitoring and quality assessment of additive manufacturing processes based on multi-sensor fusion according to claim 1, characterized in that, Based on the aforementioned three-dimensional quality digital twin model, online monitoring and early warning commands and global quality assessment results for the additive manufacturing process are output, including: The quality assessment index value contained in the latest updated discrete voxel unit in the three-dimensional quality digital twin model is extracted in real time, and the defect probability in the quality assessment index value is compared with the preset dynamic tolerance threshold in real time. If the defect probability exceeds the dynamic tolerance threshold, the manufacturing anomaly response mechanism is triggered and the online monitoring and early warning instruction is generated. The online monitoring and early warning instruction is sent to the equipment control host of the additive manufacturing equipment in real time to drive the additive manufacturing equipment to perform closed-loop adaptive feedback actions such as laser power adjustment or layer-by-layer shutdown intervention. After the additive manufacturing process is completed or when a global evaluation trigger signal is received, all discrete voxel units in the solid assigned state inside the three-dimensional quality digital twin model are completely traversed. Extract the quality assessment index values ​​from all discrete voxel units in the entity-assigned state, and aggregate and statistically analyze the global distribution density and cumulative defect volume ratio of different defect categories in a preset three-dimensional entity coordinate system. Based on the global distribution density and the cumulative defect volume ratio, and combined with the preset part service performance evaluation standards, a comprehensive judgment is made to obtain the global quality assessment result, which includes a three-dimensional visualization perspective view of internal micro-defects and an overall forming pass rate judgment.

10. An online monitoring and quality assessment system for additive manufacturing processes based on multi-sensor fusion, characterized in that, include: The data synchronization acquisition module is used to synchronously acquire multi-source heterogeneous sensor data streams in the processing area, as well as the real-time trajectory coordinates and system timestamp of the current processing position; The multimodal spatiotemporal mapping module is used to perform temporal synchronization and spatial mapping on the multi-source heterogeneous sensor data stream based on the real-time trajectory coordinates and the system timestamp, so as to construct a multimodal spatiotemporal dataset. An independent feature extraction module is used to extract features from each physical modality data in the multimodal spatiotemporal dataset to obtain a corresponding set of independent feature representations. The cross-modal attention fusion module is used to input the independent feature representation set into a preset dynamic cross-modal attention network to calculate the mutual information correlation matrix between nodes within the independent feature representation set, and to perform feature reconstruction and adaptive weight allocation on the independent feature representation set based on the mutual information correlation matrix to obtain a deep fusion feature vector. The quality assessment mapping module is used to input the deep fusion feature vector into a preset quality mapping model and output the quality assessment index value corresponding to the current processing position, wherein the quality assessment index value represents the defect probability and defect classification of the current processing position; The three-dimensional twin model construction module is used to extract the discrete spatial coordinates of the current processing position in the three-dimensional solid coordinate system, assign the quality evaluation index value to the discrete voxel unit corresponding to the discrete spatial coordinate, and progressively update the discrete voxel unit carrying the quality evaluation index value to construct a three-dimensional quality digital twin model. The early warning and assessment output module is used to output online monitoring and early warning commands and global quality assessment results for the additive manufacturing process based on the three-dimensional quality digital twin model.

Images