System for in-SITU monitoring and control using sensor fusion and machine learning

Abstract

The described technology pertains to a system for monitoring and controlling industrial or manufacturing processes, including additive manufacturing processes such as laser powder bed fusion. A sensor subsystem captures process data from at least one sensor, and, in some embodiments, contextual process data is also used. A defect monitoring subsystem comprises a sensor fusion module that fuses the captured data into a unified representation using a neural-network-based data pipeline with variational autoencoder models, and a decision-making module that generates a defect classification output and a defect probability score, and employs a defect detection dictionary to generate a defect detection indication. A process control subsystem may dynamically adjust one or more process parameters based on outputs of the defect monitoring subsystem, and a sensor simulation subsystem may validate proposed control parameter changes by generating synthetic sensor signals.

Description

[0001] SYSTEM FOR IN-SITU MONITORING AND CONTROL USING SENSOR FUSION AND MACHINE LEARNING

[0002] FIELD OF INVENTION

[0003] The present disclosure relates to a process of collecting, transporting and processing data quickly for applications that require sensory data for intelligent control. In a preferred embodiment the process is illustrated in relation to additive manufacturing technologies, specifically to in-situ monitoring and adaptive control systems for laser powder bed fusion processes.

[0004] BACKGROUND

[0005] Additive manufacturing, particularly laser powder bed fusion (LPBF), has emerged as a transformative technology in the production of complex and high-precision parts. Despite the potential, the industry faces significant challenges in ensuring consistent quality and reliability of manufactured parts. Current systems often rely on open-loop processes, where parameters are predefined and lack real-time adaptability. This approach can lead to defects such as porosity, overheating, and lack of fusion, which are only detected post-production, resulting in costly rework or complete reprints. The absence of dynamic, in-situ monitoring and control mechanisms limits the ability to make real-time adjustments, thereby affecting the overall efficiency and precision of the manufacturing process.

[0006] Existing solutions in the field have attempted to integrate sensor data for monitoring purposes, yet they often fall short in providing comprehensive, real-time feedback that can be used to adjust process parameters dynamically. These systems typically focus on single sensor inputs and produce limited outputs, which do not adequately address the complexity of the LPBF process. Moreover, the lack of standardized testing methods and the reliance on manual parameter adjustments further hinder the ability to achieve high-quality, reproducible parts. There is a pressing need for advanced systems that can seamlessly integrate multisensory data, utilize sophisticated machine learning algorithms, and provideadaptive control to enhance the precision and reliability of additive manufacturing processes.

[0007] SUMMARY

[0008] The present disclosure provides a process and system for applications where sensor fusion is required to achieve some form of intelligent detection of detrimental events in manufacturing. In one application, the process and system can be used in additive manufacture including laser bed powder fusion processes. In addition, the system can be used for a wide variety of applications including wire additive manufacturing, direct energy deposition and real time monitoring of metal removal machines (e.g. Milling Machine or Lathe) to detect and rectify chatter or detect thermal events when machining titanium or magnesium. The applications in conventional additive or metal removal manufacturing processes are endless.

[0009] Application of the process and system of the present invention can not only be focused on manufacturing. The system can be used as a basis for intelligent detection of anomalies in virtually any system requiring high speed data collection and intelligent supervisory control. For example, the system can be used to integrate an image sensor, or tactile sensors etc. onto a machine or robotic platform that may require some intelligence to accomplish a task in real time. The system is designed to be fast and intelligent. The present system can add realtime intelligence to augment existing systems that require direct sensor feedback into a controller.

[0010] In one embodiment the present disclosure provides a system for monitoring an industrial process, comprising: a sensor subsystem configured to capture process data from at least one sensor selected from the group consisting of pyrometer sensors, acoustic emissions sensors, infrared cameras, thermal imaging sensors, CT sensors, thermocouples and laser displacement sensors; a defect monitoring subsystem operatively coupled to the sensor subsystem, the defect monitoring subsystem comprising (i) a sensor fusion module configured to fuse the captured process data, optionally together with contextual process data, into a unifiedrepresentation using a neural-network-based data pipeline comprising one or more variational autoencoder models, and (ii) a decision-making module configured to generate at least a defect classification output and a defect probability score from the unified representation, wherein the defect monitoring subsystem employs a defect detection dictionary that uses a non-cryptographic hashing method together with a Hamming-based similarity measure to compare the unified representation with stored defect entries and generate a defect detection indication; and, in some embodiments, a process control subsystem configured to adjust one or more process parameters based on outputs of the defect monitoring subsystem.

[0011] In a preferred embodiment, the disclosure provides in-situ monitoring and adaptive control systems for additive manufacturing including laser powder bed fusion processes consisting of the following preferred approach.

[0012] The first relates to machine learning within the additive manufacturing field. An in-situ monitoring and control system built on a synchronized multi-sensor acquisition and data pipeline that (i) ingests multiple in-situ sensor streams and machine / process telemetry, (ii) time-aligns and synchronizes the streams, (iii) buffers / streams the synchronized data to near real-time inference and control modules, and (iv) archives the synchronized raw data and derived outputs in a traceable build record. The build record may link raw signals, process parameters, model outputs, and downstream inspection / labeling results, enabling repeatable training, validation, and continuous improvement. This monitoring and control approach is implemented by:

[0013] • Sensor fusion utilizing a neural network data pipeline

[0014] o Decision nodes providing a defect classification output and a defect probability score.

[0015] • A defect detection dictionary applying a non-cryptographic hashing method together with a Hamming-based similarity measure to compare encoded representations with stored defect entries and generate a defect detection indication.The second is regarding the application of the machine learning algorithm where:

[0016] • The algorithm simulates the system utilizing a custom kernel which is physics based in the way that the tensor weighting activation function triggers as the kernel provides insights for the system with which each tensor is trained;

[0017] • An identified workflow defines a training method for new materials or material families providing a response signature to produce trained and deployable neural network models quickly and effectively;

[0018] • The algorithm is also set up to provide modular units that can be independently adjusted to the need of the material, thus the outputs that it produces and the interaction between the systems that it considers can be effectively fine-tuned like what is seen in other large language models (LLM). Through each of the interactions in the algorithm, the algorithm’s individual parts allow for faster turnaround in fine tuning for the defects that are relevant, the controls that are required and the library that produces defect probability scores within the defect groupings;

[0019] • Finally, the application of the algorithm also provides multiple outputs from the models which in turn provide information on the process part quality. Finally, the system is tuned using a standardized method resulting in a quick turnaround on new models for material or material families. This standard testing method is set up to identify the significant defects (cracks, pores, and keyholes) within complex geometries (overhangs, thin features, enclosed features and supports) within each material or material-family. It also provides a repeatable way to iterate over new monitor and control models. Especially with the use of multiple independent models, the setup of a standardized testing method allows for the exchange of hyper-tuned models without requiring extensive commitment in material and material family control. Th / s hyper-tuning produces reduction in the time of training, development, and implementation of new Al models. Together with the standardized training models, this independent system allows the models to detect various defects regardless of geometries accurately and repeatedly.Accordingly, in one aspect, the present invention relates to an apparatus and system for monitoring and controlling an additive manufacturing process. The system includes a sensor subsystem configured to capture process data from the manufacturing environment using one or more sensors selected from the group consisting of pyrometer sensors, acoustic emissions sensors, infrared cameras, thermal imaging sensors, CT sensors, thermocouples and laser displacement sensors. A defect monitoring subsystem comprises a sensor fusion module that fuses the captured sensor data and contextual process data using a neural-network-based data pipeline employing variational autoencoder models, and a decision-making module that generates a defect classification output and a defect probability score. In this embodiment, a defect detection dictionary is implemented using a non-cryptographic hashing method together with a Hamming-based similarity measure to compare encoded representations with stored defect entries and generate a defect detection indication, and a process control subsystem dynamically adjusts parameters — such as laser power and scanning speed — based on outputs of the defect monitoring subsystem.

[0020] In yet another aspect, a method for detecting defects in an additive manufacturing process is provided. The method involves receiving process data from at least one sensor and fusing the data using a neural network-based sensor fusion module that incorporates variational autoencoder models to generate a unified output. A non cryptographic hashing algorithm is then applied to the unified output to form a defect detection dictionary. A Hamming-based similarity measure is then determined between the encoded representation and stored defect entries in the dictionary to generate a defect detection indication.

[0021] In another aspect focused on defect detection, a system is provided that includes a sensor subsystem for capturing process data and a sensor fusion module that integrates this data into a unified output. A decision-making module generates a defect classification output and a defect probability score, and a dedicated defect detection module applies a non-cryptographic hashing algorithm in combination with a Hamming-based similarity measure to compare encoded representationswith stored defect entries and generate a defect detection indication. An output module then provides the defect detection indication.

[0022] In further aspects, a physics-based custom kernel is integrated within a machine learning algorithm to simulate system responses during the additive manufacturing process. The custom kernel adjusts tensor weighting activation functions based on physics-based insights, thereby enhancing the accuracy of process outcome predictions. The algorithm is further configured to supply training data for specific materials or material families and includes modular units that can be independently adjusted to optimize interactions between system components for improved defect detection and process control. In a related embodiment, a modular machine learning algorithm for material-specific training is provided. This algorithm includes a sensor data acquisition module, a material-specific training module that generates a response signature, and a physics-based custom kernel that adjusts tensor weighting activation functions based on sensor and process condition data. A modular unit interface permits discrete adjustment of training parameters, and an output module generates a defect classification output and a defect detection indication using a defect detection dictionary implemented with a non-cryptographic hashing method and a Hamming-based similarity measure. In additional aspects, a method for standardized testing to achieve quick model turnaround in an additive manufacturing process is provided. This method comprises printing standardized test coupons with predetermined geometries and preset process parameter variations; capturing sensor data during coupon printing via data from at least two sensors selected from the group consisting of pyrometer sensors, acoustic emissions sensors, infrared cameras, thermal imaging sensors, and laser displacement sensors; and fusing the captured sensor data using a neural network-based data pipeline that incorporates variational autoencoder models to produce a unified output. The pre-determined geometries can include overhangs, horizontal channels (enclosed features), thin walls, supports, and gyroids. A defect classification output and a defect detection indication are generated using a defect detection dictionary implemented with non-cryptographic hashing and Hamming-based similarity measure. The unified sensor data and thedefect detection results are correlated with known process parameters of the printed coupons to calibrate a machine learning model. In a further embodiment, the system includes a test coupon printing module, a sensor subsystem, a sensor fusion module, a decision-making module, a defect detection module for direct coupling of outputs, a calibration module to correlate data with process parameters, and a model updating module to iteratively refine the machine learning model for hyper-tuned, in-situ monitoring and control.

[0023] These and other features will be more clearly understood from the following detailed description taken together with the accompanying drawings and claims.

[0024] BRIEF DESCRIPTION OF DRAWINGS

[0025] In order that the invention may be more clearly understood, embodiments will now be described in detail by way of example, with reference to the accompanying drawings, in which:

[0026] FIG. 1 illustrates a flow chart diagram depicting the controlled system process flow for pre-production in LPBF in accordance with one embodiment of the invention. FIG. 2 illustrates a flow chart diagram of the controlled system process flow for model training in LPBF in accordance with one embodiment of the invention. FIG. 3 is a flow chart diagram illustrating the controlled system process flow for production in an additive manufacturing environment in accordance with one embodiment of the invention.

[0027] FIG. 4 is a flow chart diagram illustrating the controlled system process flow for defect detection and parameter adjustment in additive manufacturing in accordance with one embodiment of the invention.

[0028] FIG. 5 is a block diagram illustrating the defect monitoring subsystem's training and outcome process in an additive manufacturing system in accordance with one embodiment of the invention.FIG. 6 is a block diagram illustrating the process control subsystem fortraining and outcome prediction in an additive manufacturing system in accordance with one embodiment of the invention.

[0029] FIG. 7 is a block diagram illustrating the sensor simulation subsystem for training and outcome in an additive manufacturing process in accordance with one embodiment of the invention.

[0030] FIG. 8 is a block diagram illustrating the framework for processing acoustic emissions data using a Variational Autoencoder model in accordance with one embodiment of the invention.

[0031] FIG. 9 is a block diagram illustrating the variational autoencoder framework for processing pyrometer data in an additive manufacturing system in accordance with one embodiment of the invention

[0032] FIG. 10 is a block diagram illustrating the dependencies between various data sources and machine learning models in the additive manufacturing process in accordance with one embodiment of the invention.

[0033] Similar references are used in different figures to denote similar components

[0034] DETAILED DESCRIPTION

[0035] Terminology used in the disclosure includes:

[0036] Acoustic Emissions (AE) refer to the phenomenon where transient elastic waves are generated by the rapid release of energy from localized sources within a material.

[0037] Additive Manufacturing (AM) is a process of creating three-dimensional objects by successively adding material layer by layer, based on digital models. This technology allows for the production of complex and intricate designs that would be difficult or impossible to achieve with traditional subtractive manufacturing methods.Coupon: In the context of additive manufacturing, a "coupon" refers to a standardized test piece or sample that is produced to evaluate the quality and characteristics of the manufacturing process.

[0038] Deep Convolutional Neural Network (DCNN) refers to a type of artificial neural network specifically designed for processing structured grid data, such as images. It consists of multiple layers of convolutional operations that automatically and adaptively learn spatial hierarchies of features from input data, enabling it to effectively capture complex patterns and structures.

[0039] Direct Energy Deposition (DED) is an additive manufacturing process that involves the focused deposition of material onto a substrate using a high-energy source, such as a laser, electron beam, or plasma arc. The material, often in the form of powder or wire, is simultaneously melted and deposited layer by layer to build or repair components.

[0040] Design of Experiments (DOE) is a systematic method used in research and industrial processes to plan, conduct, and analyze controlled tests or experiments. It involves the strategic selection of variables and conditions to efficiently explore the effects of these factors on a particular outcome, allowing for the identification of optimal conditions and the understanding of interactions between variables. Human Machine Interface (HMI) is a user interface or dashboard that connects a person to a machine, system, or device. It allows for interaction and communication between the user and the machine, often through graphical displays, touchscreens, or control panels, enabling the user to monitor and control the machine's operations effectively.

[0041] Laser Metal Deposition (LMD) is an additive manufacturing process that involves the focused deposition of material onto a substrate using a high-energy source, such as a laser. The material, often in the form of powder or wire, is simultaneously melted and deposited layer by layer to build or repair components. This process allows for precise control over material placement.Multivariate Imputation by Chained Equations (MICE) is a statistical technique used to handle missing data in datasets by creating multiple imputations for each missing value. It operates by iteratively modeling each variable with missing data as a function of other variables in the dataset, using a series of regression models to predict and fill in the missing values.

[0042] Multi-In-Multi-Out (MIMO) refers to a system architecture commonly used in communication and control systems where multiple input signals are processed to produce multiple output signals.

[0043] Machine Learning (ML) is a subset of artificial intelligence that involves the development of algorithms and statistical models that enable computers to perform tasks without explicit instructions. It relies on patterns and inference to learn from data, allowing systems to improve their performance on specific tasks over time. Machine learning is widely used in various applications, including predictive analytics, image recognition, and natural language processing.

[0044] Neural Network (NN) is a computational model inspired by the way biological neural networks in the human brain process information. It consists of interconnected nodes, or neurons, organized in layers, where each node performs a simple computation and passes its output to the next layer.

[0045] Physics-based custom kernel modifies the behavior of the training process by adjusting gradients, loss functions, or intermediate feature representations, ensuring that the learned model adheres to constraints derived from physical processes or domain expertise.

[0046] Powder Bed Fusion (PBF) is an additive manufacturing process that involves the use of a high-energy source, such as a laser or electron beam, to selectively fuse layers of powdered material to create a three-dimensional object. During the process, a thin layer of powder is spread across a build platform, and the energy source selectively melts and fuses the powder according to a digital model, layer by layer, until the entire object is formed.Powder Bed Fusion - Energy Beam (PBF-EB) is an additive manufacturing process that uses an energy beam, such as an electron beam or laser, to selectively fuse layers of powdered material to create a three-dimensional object. During the process, the energy beam selectively melts and fuses the powder according to a digital model, layer by layer, until the entire object is formed.

[0047] Laser Powder Bed Fusion (PBF-L I LPBF) is an additive manufacturing process that involves using a high-energy laser to selectively fuse layers of powdered material to create a three-dimensional object. During the process, the laser selectively melts and fuses the powder according to a digital model, layer by layer, until the entire object is formed.

[0048] Principal Component Analysis (PCA) is a statistical technique used to simplify the complexity in high-dimensional data while preserving trends and patterns. It achieves this by transforming the original variables into a new set of uncorrelated variables called principal components, which are ordered by the amount of variance they capture from the data.

[0049] A Proportional-Integral-Derivative (PID) controller is a control loop mechanism widely used in industrial control systems to regulate temperature, flow, pressure, speed, and other process variables. It combines three control actions: proportional, which responds to the current error; integral, which accumulates past errors; and derivative, which predicts future errors based on the rate of change. This combination allows the PID controller to maintain the desired setpoint by minimizing the error between the setpoint and the process variable.

[0050] Process Structure Property (PSP) means process-structure-property relationships for metal materials and is a primary process control mechanism for the present invention.

[0051] Pyrometer (PY) is a type of non-contact temperature measurement device used to determine the temperature of an object by detecting the thermal radiation emitted from its surface.Quasi-Sliding Mode Controller (Q-SMC) is a type of control system used in engineering to manage dynamic systems with uncertainties or disturbances. It is an adaptation of the sliding mode control technique, which aims to drive the system's state to a predefined sliding surface and maintain it there, ensuring robustness against model inaccuracies and external perturbations. The "quasi" aspect refers to modifications that reduce the chattering effect, a common issue in traditional sliding mode control, by smoothing the control action near the sliding surface.

[0052] Selective Laser Melting (SLM) is an additive manufacturing process that uses a high-energy laser to selectively fuse layers of powdered material to create a three-dimensional object. During the process, a thin layer of powder is spread across a build platform, and the laser selectively melts and fuses the powder according to a digital model, layer by layer, until the entire object is formed.

[0053] Variational Auto-encoder (VAE) is a type of generative model in machine learning that uses neural networks to learn a probabilistic mapping from input data to a latent space, and then back to the original data space. It consists of an encoder that compresses the input data into a lower-dimensional latent representation and a decoder that reconstructs the data from this latent space.

[0054] Descriptions of apparatus, systems and methods for in-situ monitoring and adaptive control of the additive manufacturing process, their features and operation, and their properties in accordance with embodiments of the invention will now follow. Possible and preferred features of the present invention will also be described with particular reference to the accompanying drawings. However, it is to be understood that the features of the invention illustrated in and described with reference to the relevant drawings are not to be construed as limiting on the scope of the invention.

[0055] In the field of additive manufacturing, particularly laser powder bed fusion (LPBF), the production of complex and high-precision parts is often hindered by several challenges. The current state of the art predominantly relies on open-loop processes where manufacturing parameters are predefined and lack real-timeadaptability. This static approach can lead to defects such as porosity, overheating, and lack of fusion, which are typically detected only after production. As a result, these defects necessitate costly rework or complete reprints, significantly impacting the efficiency and cost-effectiveness of the manufacturing process.

[0056] Existing systems in the industry have attempted to integrate sensor data for monitoring purposes. However, these systems often fall short in providing comprehensive, real-time feedback that can be used to dynamically adjust process parameters. Typically, they focus on single sensor inputs and produce limited outputs, which do not adequately address the complexity of the LPBF process. Moreover, the absence of standardized testing methods and the reliance on manual parameter adjustments further hinder the ability to achieve high-quality, reproducible parts. The lack of dynamic, in-situ monitoring and control mechanisms limits the ability to make real-time adjustments, thereby affecting the overall efficiency and precision of the manufacturing process.

[0057] The present system addresses these challenges by introducing an advanced method for in-situ monitoring and adaptive control of the additive manufacturing process. This method integrates a sensor subsystem configured to capture process data from the manufacturing environment, including at least one sensor selected from the group consisting of pyrometer sensors, acoustic emissions sensors, infrared cameras, thermal imaging sensors, CT sensors, thermocouples and laser displacement sensors with preprocessing supported by variational autoencoders, producing a unified sensor data output. A variational autoencoder is trained using the sensory data, but when put into practical use, it is split into the encoder and decoder for the monitoring and sensor simulation subsystems respectively. The decision-making module, operatively connected to the sensor fusion module, generates outputs that include defect classification and quantification of defect and feature certainty. By employing a defect detection dictionary using a non-cryptographic hashing method and Hamming-based similarity measure, the system directly couples classification outputs with regression outputs. This enables a process control subsystem to dynamicallyadjust process parameters, such as laser power and scanning speed, based on the integrated sensor data and decision-making outputs, thereby enhancing the precision and reliability of the additive manufacturing process.

[0058] Controlled / Closed-loop pre-production processes in LPBF

[0059] When introducing a control system according to the present invention, multiple reducing factors come together that facilitate the faster training process and the introduction of quicker process structure maps (PSP).

[0060] In the diagram Fig. 1 illustrating one embodiment of controlled system process flow for pre-production in LPBF, the reduction can be seen primarily in the execution of manual parameters identification as the control model now facilitates the easy identification of the ideal PSP.

[0061] When looking at the system as illustrated in the embodiment of Fig. 1 , the machine learning control system that produces this closed loop has the following data flow for the identification of a new process parameter set:

[0062] Model Setup

[0063] 1. Print parts as specified in the standardized coupon format of 5x cube units with low, intermediate, ideal, intermediate-high and high process energy density settings. Collect the data including:

[0064] a. PSP used (Power, Speed, Hatch spacing, Layer thickness) b. Part actual density %

[0065] c. Defects observed

[0066] 2. Train the VAEs to reproduce the signals from each of the sensors after sensor data prep (see Fig. 7).

[0067] 3. Combine the trained VAE models into the generalizing model and train each of the defects as the output.

[0068] 4. Using the defects, the PSP and the density identified in the parts in step 1 , combine these into a model that reproduces the PSP.5. Using the PSP, defects and the density identified in the parts in step 1, combine these into a model that has the same output size as the latent space from the VAE Decoder input. Train this model with the signals produced from step 1.

[0069] The standardized coupon suite described above can be executed not only to generate data for initial model training, but also as a repeatable regimen for (i) validating models prior to deployment or prior to enabling closed-loop actions, (ii) maintaining models over time by periodically re-running coupons to detect drift and trigger recalibration / retraining / model updates, (iii) performing sensor maintenance checks (e.g., calibration drift, noise-floor changes, channel faults, coupling / mount changes), and (iv) performing machine / process maintenance checks (e.g., laser power drift, scan-path deviations, recoater behavior changes, shielding-gas anomalies).

[0070] PSP Setup

[0071] Using the produced models from the Model Setup complete the following:

[0072] 1. Change 1 of the PSP parameters OR the density

[0073] 2. Provide the dictionary value, corresponding defect and remaining PSP values to the control model OR the ideal dictionary with no defects.

[0074] 3. Use the output of the Process Control Subsystem to predict PSP values to change or correct. Take note of these to see if the values are saying the system will be significantly out of control.

[0075] 4. Use the Process Control System to predict a new signal output

[0076] 5. Using the new signal, predict the defect and dictionary value

[0077] 6. MAP a new defect with the PSP that was used to provide a full map 7. If required, predict a corrective output. Note that the PSP predicted should vary from the current mapping.

[0078] What Fig. 1 represents:• Dynamic Process Execution: The LPBF machine operates with an initial set of parameters that are dynamically altered by the control system during operation.

[0079] • Manual Material Selection: Feedstock (metal powder) is still sourced and determined to be a new or previously used material.

[0080] o If a new material is introduced, the system requires the execution of a control model training procedure.

[0081] As noted above Fig. 1 shows a controlled system process flow for pre-production in laser powder bed fusion (LPBF) according to one embodiment of the present invention. The method can be implemented by an additive manufacturing system to prepare for the production phase. The process involves steps such as identifying and sourcing feedstock, creating a CAD model, and validating process parameters. The method begins with identifying and sourcing the feedstock, specifically metal powder, which is necessary for the additive manufacturing process. This step ensures that the appropriate material is available for the subsequent production stages. The CAD model is created and serves as the digital blueprint for the object to be manufactured. This model plays an important role in guiding the laser in selectively fusing the powder layers according to the design specifications. The method then determines whether the material is new. If the material is new, the process proceeds to the next step where the method checks whether the new material is supported in the literature or if manual parameter identification is required.

[0082] Reduced manual parameter identification is executed for materials not supported in the literature and may involve determining the optimal process parameters through experimentation and analysis.

[0083] The method is designed to execute a model training procedure to refine the process parameters for new material, ensuring the process parameters meet the necessary requirements for successful production. The method identifies the process parameters, which include settings such as laser power and scanning speed, tailored to the specific material and design requirements. The CAD modelis sliced according to the layer thickness derived from the feedstock properties. This step prepares the model for layer-by-layer construction during the manufacturing process. The sliced CAD model is loaded into the system, ready for the production phase. The defined process parameters determined earlier are loaded into the system, ensuring that the manufacturing process adheres to the specified settings. The powder feedstock is loaded into the system, completing the preparation for the additive manufacturing process and then the process can be initiated, marking the transition from pre-production to the actual manufacturing phase, where the object is constructed layer by layer according to the CAD model and process parameters.

[0084] In another aspect, the present invention provides a machine learning-based system for in-situ monitoring and adaptive control of an additive manufacturing process that can be configured to operate with a variety of sensor types beyond pyrometers and acoustic emissions sensors, such as infrared cameras or laser displacement sensors, to capture diverse process data. The sensor fusion module may be implemented using different neural network architectures, such as convolutional neural networks (CNNs) or recurrent neural networks (RNNs), to optimize the integration of sensor data depending on the specific application requirements. The decision-making module could be adapted to include additional outputs, such as predictive maintenance alerts or energy consumption metrics, enhancing the system's utility in various manufacturing environments. The defect detection dictionary within the decision-making module might employ alternative hashing methods or similarity measures, such as cosine similarity, to improve the accuracy and speed of defect classification. Furthermore, the process control subsystem could be designed to adjust a broader range of process parameters, including ambient temperature or material feed rate, to accommodate different materials or manufacturing conditions. This flexibility allows the system to be tailored for specific industries, such as aerospace or automotive, where precision and adaptability are of great importance.

[0085] Basic Process Flow for Model Training on a New Material (With Control System)In general Fig. 2 represents:

[0086] • Iterative Al Model Training for Process Optimization:

[0087] o The system iterates over test builds (coupons) with varied geometries to capture how LPBF process signatures shift due to geometric complexity, thermal effects, and scan path interactions. Geometries may include overhangs, horizontal channels (enclosed features), thin walls, supports, and gyroids.

[0088] o Each coupon is evaluated to build a dataset that accounts for geometry-dependent process deviations.

[0089] o Data is used to train an Al model that predicts process outcomes and refines parameter selection dynamically.

[0090] • Structured Data Acquisition & Validation:

[0091] o A control sample is printed to establish a reference baseline for quality.

[0092] o Al models are trained based on acquired process data from experimental builds.

[0093] o The trained Al model undergoes a validation cycle to assess accuracy.

[0094] o Pre-deployment validation can be performed using the standardized coupon suite (and / or representative geometries) to verify defect detection performance prior to production use.

[0095] o Ongoing re-validation: the standardized coupon suite can be scheduled periodically to detect model drift due to machine changes, material lot changes, sensor changes, or process / environment changes; results can trigger model updates and re-qualification. o Maintenance checks: repeating the standardized coupon suite can also serve as a reference build to verify sensor health andmachine / process health, and to flag maintenance actions when deviations are detected.

[0096] • Closed-Loop Learning & Deployment:

[0097] o Once validated, the trained Al model replaces traditional DOE-based manual tuning, reducing the need for repeated experimental builds. Fig.. 2 shows a controlled system process flow for model training in laser powder bed fusion (LPBF) according to one embodiment of the present invention. The method is designed to optimize the training of Al models for additive manufacturing by systematically evaluating and refining process parameters and model performance. The method can be implemented by an additive manufacturing system to ensure that Al models are accurately trained and validated for production use. The method begins with identifying a PSP map for the new material. This step involves understanding the relationship between the process parameters, the structure of the material, and the resulting properties, which is used for setting the foundation for model training. Next the method identifies process parameters for three sets of varying quality levels. This involves selecting different combinations of parameters, such as laser power and scanning speed, to explore their effects on the quality of the manufactured parts. Then a CAD model is sliced according to layer thickness derived from the feedstock properties. This step prepares the model for layer-by-layer construction during the manufacturing process, ensuring that the design specifications are met. The method executes production for data acquisition by running the additive manufacturing process with the selected process parameters to produce test parts, which are used to gather data for model training. The quality of the produced test parts is validated by inspecting the parts for defects and measuring their properties to ensure they meet the desired specifications and quality standards. The method determines whether all coupon variants have been produced. If this is not the case, the process returns to the step to identify process parameters for three sets of varying quality levels and repeat the production and validation steps. The method then executes the production of a control sample for validation by producing a sample part using the establishedprocess parameters to serve as a benchmark for evaluating the Al model's performance. The quality of the control sample undergoes evaluation and involves assessing the sample to ensure compliance with the required quality standards, serving as a reference for model accuracy. Al models are trained using the data acquired from the production runs by using machine learning techniques to develop models that can predict and control the manufacturing process. The method then evaluates whether the Al model performs accurately. If the model does not perform accurately, the process undergoes further training and refinement. If the Al model performs accurately, the Al model is enrolled in production by integrating the trained model into the manufacturing process to enhance control and quality assurance. Optionally, the trained model's data hash map is enrolled in the defect library by storing the model's data for future reference and use in defect detection and process optimization.

[0098] In general Fig. 3 illustrates a controlled system flow in accordance with one embodiment of invention. In particular Fig. 3 illustrates:

[0099] • Real-Time Monitoring & Adaptive Control: This closed-loop system integrates real-time sensing and machine learning (ML)-driven process control to detect defects and dynamically adjust parameters based on observed conditions.

[0100] o Unlike open-loop LPBF, this system monitors temperature and acoustic emissions per layer and adapts process parameters if defects are detected.

[0101] • Integrated Sensor-Based Quality Assessment: The system fuses sensor data at the raw data-level to monitor the production process.

[0102] o Pyrometers measure melt pool temperature, ensuring consistent thermal conditions.

[0103] o Acoustic emissions (AE) sensors detect irregularities, such as lack of fusion, cracking, or keyhole porosity.Layer-Wise Deposition with Feedback Mechanism: The system has an integrated feedback mechanism where new process parameters are generated and applied between layers if defects were detected.

[0104] Major Advantage Over Open-Loop LPBF:

[0105] • Defects are corrected during the build, reducing scrap rates and rework costs.

[0106] • Dynamic process control adapts settings in response to real-time sensor data, ensuring higher part quality and process repeatability.

[0107] • Post-build data is stored for continuous improvement, enabling smarter future builds. In some embodiments, this post-build information is retained as part of a traceable build record linking synchronized raw sensor streams, process parameters, derived features, and model outputs.

[0108] As noted above Fig. 3 illustrates a controlled system process flow for production in an LPBF system according to one embodiment of the present invention. This method is implemented by an additive manufacturing system to manage the production phase, ensuring precise control over the manufacturing process. The method involves a series of steps, including initiating the process, reading process parameters, and monitoring various aspects of the production environment, such as monitoring temperature (pyrometer) and monitoring acoustic emissions. Additionally, the process includes reading slice for current layer, exposing layer, depositing layer, and determining if the part is complete. Finally, the system compiles and exports AE data, temperature data, and Hashmap functions. At step 1 , the process is initiated, marking the beginning of the production phase. This step sets the stage for subsequent actions by preparing the system to execute the defined manufacturing sequence. At step 2, the system reads the process parameters. These parameters include important settings such as laser power and scanning speed, which are necessary for maintaining the desired quality and consistency of the manufactured parts. At step 3, the system reads the slice for the current layer. This involves accessing the digital model to determine the specific layer geometry and instructions for the laser to follow during the fusionprocess. At step 8, the system deposits the layer by raising the dispenser and moving the recoater. This step ensures that a fresh layer of powder is evenly spread across the build platform, ready for selective laser melting. At step 4, the system exposes the layer to the laser. This action selectively fuses the powder according to the slice data, forming the desired geometry for the current layer. At step 5, the system monitors the temperature using a pyrometer. This non-contact temperature measurement ensures that the process remains within the optimal thermal range, preventing defects such as overheating or lack of fusion. At step 6, the system monitors acoustic emissions. This involves detecting transient elastic waves generated during the process, which can provide insights into the integrity and quality of the part being manufactured. At step 9, the system evaluates whether the part is complete. If the part is complete, the process moves to the next steps; otherwise, the system loops back to continue the layer-by-layer construction. At step 9, if the part is complete (step 6), the system removes the part (step 7) from the build platform. This step marks the end of the manufacturing process for the current part. At step 10, the system compiles and exports acoustic emissions data. This data plays an important role in post-process analysis and quality assurance. At step 11 , the system compiles and exports temperature data. This information helps in verifying that the process parameters were maintained within the desired range throughout production. At step 10, the system compiles and exports HashMap functions. These functions are used to correlate process data with specific manufacturing outcomes, aiding in process optimization and defect detection.

[0109] Fig. 4 shows the decision-making process within the controlled / closed-loop LPBF system. Fig. 4 illustrates:

[0110] • Layer-by-Layer Defect Detection and Adaptive Control: The system as illustrated continuously monitors sensor inputs from temperature (pyrometer) and acoustic emissions to detect defects and take corrective actions.

[0111] o Modifies laser power and / or scan speed to correct identified issues.• Iterative process validation: The system validates new process parameters by simulating a sensor output and validating it against the defect monitoring subsystem.

[0112] Major Advantage Over Open-Loop LPBF:

[0113] • Instead of printing with fixed parameters and checking for defects only after the build, this system corrects errors dynamically, improving repeatability and efficiency.

[0114] • Adaptive learning reduces trial-and-error, leading to fewer part failures and faster process optimization.

[0115] Fig. 4 illustrates the controlled system process flow for production in an additive manufacturing environment. This method is designed to ensure the quality and precision of the manufacturing process by iteratively monitoring and adjusting process parameters. The method can be implemented by an additive manufacturing system equipped with multiple sensor input, a defect monitoring subsystem, a process control subsystem (modify process parameters), and a simulation subsystem (sensor simulation). At step 1, the method begins with receiving sensor inputs from the manufacturing environment. This input plays a role in capturing real-time data about the process conditions, which will be used to detect any potential defects. At step 2, the defect monitoring subsystem analyzes the sensor inputs to determine if there are any defects present in the manufacturing process. This subsystem is responsible for identifying anomalies or deviations from the expected process parameters. At step 3, the method evaluates whether a defect is detected. If no defect is found, the process proceeds to the next step without any modifications. This decision point ensures that only necessary adjustments are made, maintaining process efficiency. At step 3, if a defect is detected, the process control subsystem is activated to predict and modify the process parameters. This subsystem dynamically adjusts settings such as laser power and scanning speed to correct the identified defect and optimize the manufacturing process. At step 5, the simulation subsystem performs sensor simulation to predict the outcomes of the adjusted process parameters. This stepplays a significant role in validating the effectiveness of the modifications before they are applied to the actual manufacturing process. The method iterates through these steps up to a predefined number of times, ensuring continuous monitoring and control of the manufacturing process to achieve the desired quality and precision, including steps such as sensor input, determining if there is a defect, proceeding, modifying process parameters through the process control subsystem, and utilizing the simulation subsystem for sensor simulation.

[0116] In the embodiment illustrated the following components play a part in the Machine Learning Control System (referred to as sub-systems) in Fig.s 5-10:

[0117] 1. A Variable Auto-Encoder (VAE) designed for each sensor, from which: a. The Encoder is used to generalize the signals received from each of the sensors in a physics aware manner. Due to the nature of the sensors:

[0118] i. The acoustic emissions VAE incorporates both time and frequency-domain analysis.

[0119] ii. The pyrometer VAE incorporates only time-domain analysis. b. The Decoder assists in the simulation of a signal given an input that is produced from a predictive control model.

[0120] 2. A machine learning model that uses data-level sensor fusion to facilitate:

[0121] a. The classification of sensor and process condition data to identify the presence of defects (i.e. cracks and pores).

[0122] b. The production of a unique signature for the dictionary hash map.

[0123] 3. A machine learning model that facilitates the prediction of a control parameter set.

[0124] 4. Machine learning models that facilitate the simulation of sensor signals given a set of process conditions.

[0125] 5. A synchronized data acquisition and processing pipeline that ingests multimodal sensor streams and machine / process telemetry, performs timealignment / synchronization, supports buffering / streaming for near real-time inference, and archives data for post-build analysis.

[0126] 6. A traceability mechanism (e.g., a structured build record) that links synchronized raw sensor data, process parameters / telemetry, derived features, model outputs (including confidence / certainty metrics), and control actions and / or inspection labels when available.

[0127] 7. Sensor acquisition and signal-conditioning integration (including concurrent capture of multiple channels and time-stamped digitization) sufficient to support repeatable multi-modal sensor fusion.

[0128] 8. A model lifecycle workflow supporting repeatable data preparation, training, validation, deployment, and periodic re-validation / model maintenance over time (including drift detection and triggered updates).

[0129] 9. Optional automated integrity / health checks that verify data completeness, synchronization behavior, and throughput / latency prior to relying on model outputs.

[0130] In operation, the different components of the Machine Learning Control System integrate as follows:

[0131] 1. The VAEs are trained separately from the control architecture and are deployed through their component parts as follows:

[0132] a. Encoder-derived dimensionality-reduction models are used within the defect monitoring subsystem.

[0133] b. Decoder-derived reconstruction / simulation models are used within the sensor simulation subsystem.

[0134] 2. The defect monitoring subsystem comprises the sensor fusion module and the decision-making module. Within that subsystem, sensor data and contextual process data are fused to produce a unified representation from which defect classification output and a defect probability score are generated.a. If a defect is not detected, the process proceeds with the existing control parameter set.

[0135] b. If a defect is detected, the event is logged and the process control loop is initiated.

[0136] 3. The process control subsystem uses outputs of the defect monitoring subsystem together with process conditions data to predict a revised control parameter set.

[0137] 4. The sensor simulation subsystem then generates synthetic sensor signals corresponding to the revised control parameter set so that the revised control parameter set can be validated before being applied to the manufacturing process.

[0138] Utilizing a sensor fusion machine learning model, a monitoring model produces a classification of a defect type and a defect probability score, and a defect detection dictionary uses a non-cryptographic hashing method together with a Hamming-based similarity measure to generate a defect detection indication. These outputs provide relevant information for a control model that repeatedly and accurately adjusts the process to ensure repeatable parts. Finally, with a standard physical model set, the system can be trained on a limited variation test utilizing a simulation model that allows for hyper-tuned monitor and control systems for the additive manufacturing space targeted to a specific material or material family.

[0139] Standardized coupon builds can be used throughout the model lifecycle — not only for initial model training, but also for validation prior to deployment, ongoing model maintenance (including drift detection and triggered updates), and periodic maintenance checks on sensors and on the machine / process.

[0140] Fig. 5 illustrates a deep learning-based data fusion framework for in situ defect detection for LPBF machines according to one embodiment of the present invention. Illustrated here are components for training the neural network model, the supporting neural networks, and the outcome of the trained model. It consists of four components organized in three stages:1. Data Pre-processing - Pyrometer and acoustic emissions sensor data is collected in addition to data related to the process conditions (e.g. laser power, scanning speed, hatch spacing). A 3D computed tomography (CT) scan provides porosity labels as ground truth for training the neural network model

[0141] 2. Supporting Neural Networks - Models designed for the dimensionality reduction of the raw pyrometer and acoustic emissions sensor data are extracted from the encoder portion of their respective VAEs.

[0142] 3. Deep Learning Model - Sensor data from the sensor subsystem, together with the contextual process data, are fused within the defect monitoring subsystem by one neural network comprised of:

[0143] a. A physics aware data encoder that analyzes the input from each data source, leveraging the dimensionality reduction models to process the respective sensor data prior to fusing the input data. b. A multi-out process interpreter that combines the encoder streams and analyzes the data to provide multiple output functions.

[0144] 4. Decision Making - The model predicts the probability of the presence of crack propagation and porosity, assigning the region as either Good or Bad based on the likelihood of a either type of defect.

[0145] 5. Dictionary - The model provides a dictionary of the probability values that correspond to the defect classification output.

[0146] This framework enhances real-time defect detection in LPBF through sensor fusion at the raw data level. Leveraging the dimensionality reduction models for each respective sensor, extracted from the corresponding VAE, reduces the complexity of the model and thereby, the training time.

[0147] Fig. 5 shows a block diagram illustrating the defect monitoring subsystem's training and outcome process in an additive manufacturing system according to one embodiment of the invention. The figure includes data inputs such as 3D CT, AE data, Pyrometer data, Process conditions data (PCD), Part quality label (PQL). Insupporting neural networks include AE dimensionality reduction model (ADRM) and Pyrometer dimensionality reduction model (PDRM). In the Neural Network Model there are Physics aware data encoder (PADE) and Multi-out process interpreter (MOPI). Decision making is based on Predicted possibility for the presence of a crack (PPFTPOAC), Predicted possibility for the presence of a pore (PPFTPOAP), Defect detection library regressor (DDLR), cracks (absent / good or present / bad), and pores (absent / good or present / bad). These components interact to process data, reduce dimensionality, encode data, interpret processes, and make decisions regarding defect detection. 3D CT is responsible for providing comprehensive imaging data of the manufactured part, which is important for assessing the overall quality and identifying potential defects. This data is used together with the Part quality label to establish a baseline for evaluating the manufacturing process. AE data captures acoustic emissions from the manufacturing environment, which are indicative of material behavior and potential defects. This data is processed by the AE dimensionality reduction model to reduce complexity while retaining important features for further analysis. Pyrometer data provides temperature measurements from the manufacturing process, which are important for understanding thermal conditions and their impact on material properties. The Pyrometer dimensionality reduction model (PDRM) processes this data to simplify the information while preserving important details necessary for defect detection. Process conditions data (PCD) includes various parameters of the manufacturing environment, such as temperature, pressure, and material feed rate. This data is integrated with other inputs, including AE data and pyrometer data, to provide a comprehensive view of the manufacturing process. The Physics aware data encoder (PADE) encodes the processed data into a format suitable for analysis by the Multi-out process interpreter (MOPI). This interpreter generates multiple outputs, including the Predicted possibility for the presence of a crack (PPFTPOAC) and the Predicted possibility for the presence of a pore (PPFTPOAP), which are used to assess the likelihood of defects. The Defect detection library regressor (DDLR) utilizes a library of known defect patterns to refine the predictions made by the Multi-out process interpreter (MOPI). The DDLRassists in classifying the outputs into categories such as Good (cracks), Bad (cracks), Good (pores), and Bad (pores) providing a clear indication of the part's quality.

[0148] Fig. 6 shows a deep learning-based framework for the prediction of corrective process parameters for LPBF machines. Illustrated here are the components required for training the neural network model, the supporting dictionary, and the outcome of the trained model. It consists of four components organized in three stages:

[0149] 1. Data Pre-processing - Process conditions are collected (e.g. laser power, scanning speed, hatch spacing) and linked to the defect classification data output by the defect monitoring subsystem.

[0150] 2. Supporting Dictionary - The probability values that correspond to the defect classification output from the defect monitoring subsystem.

[0151] 3. Neural Network Model - Defect classification data, output from the defect detection library regressor, and the process conditions data are analyzed by one neural network in the form of a multi-layer perceptron.

[0152] 4. Predictions - The model provides a predicted control parameter set, specifically, a value for the laser power and scanning speed. All other process parameters are kept constant.

[0153] Fig. 6 illustrates the process control subsystem fortraining and outcome prediction in an additive manufacturing system. The method can be implemented by the process control subsystem to optimize manufacturing parameters based on defect detection and process conditions data. The subsystem comprises data preprocessing, a dictionary, a neural network model, and predictions. The method begins with data pre-processing, where defect classification and process conditions data are prepared for further analysis. This step involves organizing and cleaning the data to ensure the data is suitable for input into the subsequent stages of the subsystem. The method utilizes a dictionary specifically a defect detection library regressor, to analyze the pre-processed data. This step involves applyingregression techniques to identify patterns and correlations within the defect detection data, which are important for accurate predictions. The method employs a neural network model, specifically a multi-layer perceptron, to process the outputs from the dictionary. This step involves using the neural network to learn from the data and generate predictions based on the identified patterns and correlations. The method generates predictions resulting in a new control parameter set. This step involves determining optimal settings for manufacturing parameters, such as laser power and scanning speed, based on the predictions from the neural network model. These parameters are then adjusted to enhance the precision and reliability of the additive manufacturing process.

[0154] In one embodiment, the sensor subsystem within the system for sensor fusion can be configured to include additional sensors beyond the pyrometer and acoustic emissions sensors, such as infrared cameras or thermal imaging sensors, to capture a broader range of process data from the additive manufacturing environment. This expanded sensor array can enhance the system's ability to detect subtle variations in temperature and material behavior, thereby improving the accuracy of defect detection. In another embodiment, the neural network data pipeline can be designed to incorporate advanced machine learning techniques, such as deep learning models or reinforcement learning algorithms, to optimize the fusion of sensor data and improve the system's adaptability to different manufacturing conditions. This approach allows the system to learn and adjust its parameters dynamically, ensuring robust performance across various additive manufacturing setups. Additionally, the variational autoencoder models within the pipeline can be tailored to accommodate different sensor types and configurations, providing flexibility in the system's deployment across diverse manufacturing environments. In yet another embodiment, the decision module can be equipped with a user interface that allows operators to input specific manufacturing goals or constraints, enabling the system to prioritize certain outputs, such as defect classification or feature identification, based on real-time production requirements. This customization capability ensures that the system can be adapted to meet the specific needs of different manufacturing processes, enhancing its utility andeffectiveness. Furthermore, the defect detection dictionary can be implemented using alternative hashing methods or data structures, such as Bloom filters or hash tables, to optimize the speed and efficiency of defect classification and certainty quantification, thereby improving the system's responsiveness and accuracy in real-time applications.

[0155] The system for in-situ monitoring and adaptive control of an additive manufacturing process can be configured with a sensor subsystem that includes multiple pyrometer sensors and acoustic emissions sensors strategically placed around the manufacturing environment to capture a comprehensive set of process data. This configuration enhances the system's ability to detect defects and monitor features with greater accuracy. In another embodiment, the sensor fusion module may utilize a neural network data pipeline that incorporates advanced variational autoencoder models tailored for each sensor type, allowing for more precise data integration and output generation. The decision nodes in this embodiment could be designed to operate with a higher degree of parallel processing capability, enabling the system to handle complex data streams and produce a defect classification output and a defect detection indication. Additionally, the process control subsystem could be equipped with advanced algorithms that dynamically adjust a wider range of process parameters to optimize the manufacturing process in real-time. This embodiment could be particularly beneficial for manufacturing environments that require high precision and adaptability to varying material properties and geometries. Geometries may include overhangs, horizontal channels (enclosed features), thin walls, supports, and gyroids. Furthermore, the system could be implemented in a modular fashion, allowing for easy upgrades and integration with existing manufacturing setups, thereby providing a scalable solution that can be customized to meet specific production needs.

[0156] The method for detecting defects in an additive manufacturing process utilizes a sensor subsystem comprising multiple pyrometer sensors and acoustic emissions sensors, each strategically positioned to capture comprehensive process data from various angles and depths within the manufacturing environment. This configuration enhances the accuracy and reliability of the defect detection processby providing a more detailed and holistic view of the manufacturing conditions. In another embodiment, the sensor fusion module employs advanced neural network architectures, such as deep convolutional neural networks (DCNNs), to process and integrate the sensor data, thereby improving the precision of the unified output. This embodiment allows for the detection of subtle defects that may not be apparent with simpler fusion techniques. Additionally, the non-cryptographic hashing algorithm used to generate the defect detection dictionary can be implemented using different hashing methods, such as locality-sensitive hashing, to optimize the speed and efficiency of defect classification and certainty determination. Furthermore, the method can be adapted to accommodate various additive manufacturing materials, including metals, polymers, and ceramics, by calibrating the sensor fusion module and defect detection dictionary to account for the specific properties and behaviors of each material type. This adaptability ensures that the method remains effective across a wide range of manufacturing scenarios and material compositions.

[0157] The system for defect detection in an additive manufacturing process can be configured with a sensor subsystem that includes multiple pyrometer sensors and acoustic emissions sensors, strategically positioned to capture comprehensive process data from various angles and locations within the manufacturing environment. This configuration enhances the system's ability to detect defects with greater accuracy. In another embodiment, the sensor fusion module may utilize advanced neural network architectures, such as deep convolutional neural networks (DCNNs), to process and integrate the sensor data, thereby improving the robustness and reliability of the unified sensor data output. The decisionmaking module could be implemented with a flexible software architecture that allows for the integration of additional sensor types, such as thermal cameras or laser profilometers, to further enhance defect detection capabilities. Additionally, the defect detection module might employ a variety of non-cryptographic hashing algorithms, each optimized for different types of defects, to generate a more comprehensive defect detection dictionary. This module could also be designed to operate in real-time, providing immediate feedback and defect detectionindications to the output module, which could be configured to trigger automated corrective actions or alerts to operators. These embodiments demonstrate the system's adaptability to different manufacturing environments and requirements, ensuring effective defect detection across a wide range of additive manufacturing processes.

[0158] Fig. 7 illustrates an embodiment of a deep learning framework for the simulation of sensor data given a set of process conditions in LPBF machines according to the present invention. Illustrated here are the components for training the neural network model, the supporting neural networks, and the outcome of the trained model. It consists of four components organized in three stages:

[0159] 1. Data Pre-processing - Process conditions are collected (e.g. laser power, scanning speed, hatch spacing) and linked to the sensor data (pyrometer and acoustic emissions) collected during production act. The sensor data acts as the ground truth used in training the neural network model.

[0160] 2. Supporting Neural Networks - Models designed for the reconstruction of the raw pyrometer and acoustic emissions sensor data are extracted from the decoder portion of their respective VAEs.

[0161] 3. Neural Network Models - The process conditions data is analyzed by two neural networks, one dedicated to the pyrometer and the other acoustic emissions sensor, each leverages their respective signal reconstruction models for improved performance.

[0162] 4. Signal Simulation - Each model reconstructs a simulated sensor signal based on the process conditions.

[0163] This framework validates process parameters output by the process control subsystem by generating synthetic acoustic-emissions and pyrometer signals from process conditions using decoder-derived models..

[0164] Fig. 7 illustrates the sensor simulation subsystem for training and outcome in an additive manufacturing process. This subsystem comprises AE data, Pyrometer data, and Process conditions data (PCD), which are processed through variouscomponents including AE signal reconstructors, Pyro signal reconstructor, Data encoders, Simulator for AE signals and Simulator for Pyrometer signals. AE data serves as an input that captures acoustic emissions from the manufacturing environment. This data is used, together with process conditions data, within the sensor simulation subsystem to support generation of synthetic acousticemissions signals.. The generated acoustic-emissions signals are used to validate predicted process parameters by evaluating the simulated signals through the defect monitoring subsystem. The reconstructor signals may further be utilized in the neural network models for training purposes. Pyrometer data provides temperature measurements from the manufacturing process. This data is used, together with process conditions data, within the sensor simulation subsystem to support generation of synthetic pyrometer signals. The generated pyrometer signals are used in validating predicted process parameters as with the generated acoustic-emissions signals. Process conditions data (PCD) encompasses various environmental and operational parameters of the manufacturing process. This data is for contextualizing the sensor data and is used together with the AE data and Pyrometer data to provide a comprehensive input for the neural network models. The Data encoders encode the processed sensor data into a format suitable for neural network processing. These encoders facilitate the transformation of raw data into structured inputs that the neural networks can efficiently utilize for training and simulation. The AE signal reconstructor (ASR) and Pyrometer signal reconstructor (PSR) are decoder-derived reconstruction / simulation support models used within the sensor simulation subsystem to generate synthetic acoustic-emissions and pyrometer signals. These components ensure that the data fed into the simulators and is of high fidelity, thereby enhancing the accuracy of the simulation results. Simulator for AE signals (SFAS) and Simulator for Pyrometer signals (SFPS) are the final components in the subsystem. These simulators utilize the reconstructed and encoded data to simulate the acoustic and temperature signals, respectively. The simulations offer insights into the manufacturing process, aiding in the training of neural network models and enhancing the predictive capabilities of the system.Fig. 8 illustrates a VAE-based framework for processing acoustic emissions data in an additive manufacturing system. The process consists of three main stages:

[0165] 1. Data Pre-processing - Acoustic emissions data is collected and prepared for further analysis. The data is altered to create additional variants to improve training accuracy by independently introducing jitter, time warp, and applying filters. Note that these are only applied when training the model.

[0166] 2. Neural Network Model - A VAE neural network is trained on the acoustic emissions data. The VAE consists of an encoder and a decoder:

[0167] a. The encoder learns a lower-dimensional latent representation of the input signals.

[0168] b. The decoder reconstructs the original signals from this latent space.

[0169] 3. Supporting Model Generation - After training the VAE:

[0170] a. The encoder is extracted for use as dimensionality reduction model that compresses signal information while preserving essential features.

[0171] b. The decoder is extracted for use as a signal reconstruction model, allowing for the generation or restoration of acoustic emissions signals.

[0172] This framework enables efficient processing of acoustic emissions data by leveraging the trained VAE for both feature compression and signal synthesis. The former is applied to the framework for defect detection whereas the latter is applied to the framework for sensor simulation.

[0173] Fig. 8 shows a framework for processing acoustic emissions data using a variational autoencoder model. This method is designed to enhance the analysis of acoustic emissions data by employing a neural network structure that includes encoders. The method illustrated can be implemented within an additive manufacturing system to improve the detection and classification of defects through advanced data processing techniques but can be applied outside ofadditive manufacturing. The method begins with the data pre-processing phase, where raw acoustic emissions data is collected. This step ensures that the data is in a suitable format for further processing and analysis. The method may introduce jitter to the acoustic emissions data by adding small, random variations to the data to simulate real-world conditions and improve the robustness of the model. A time warp can be added to the acoustic emissions data by adjusting the timing of the data points to account for potential variations in the data collection process, enhancing the model's ability to handle temporal discrepancies. The method applies filtering to the acoustic emissions data to remove noise and irrelevant information from the data, ensuring that only the most pertinent features are retained for analysis. The method processes the pre-processed data through a neural network, using an encoder. The encoder compresses the data into a lowerdimensional representation, capturing the features for defect detection. The method then utilizes a decoder within the neural network model for signal reconstruction to reconstruct the data from the compressed form. This ensures that the model can accurately recreate the original data, validating the effectiveness of the encoding process. The method supports model generation by employing a neural network model for dimensionality reduction of signals. This model further refines the data, reducing complexity while preserving important information. A neural network model is utilized for signal reconstruction for accurately reconstructing the original signals from the reduced data, ensuring that the integrity of the data is maintained throughout the process.

[0174] Fig. 9 shows a VAE-based framework for processing pyrometer data in an additive manufacturing system. The process consists of three main stages:

[0175] 1. Data Pre-processing - Pyrometer data is collected and prepared for further analysis.

[0176] 2. Neural Network Model - A VAE neural network is trained on the pyrometer data. The VAE consists of an encoder and a decoder:

[0177] a. The encoder learns a lower-dimensional latent representation of the input signals.b. The decoder reconstructs the original signals from this latent space.

[0178] 3. Supporting Model Generation - After training the VAE:

[0179] a. The encoder is extracted for use as dimensionality reduction model that compresses signal information while preserving essential features.

[0180] b. The decoder is extracted for use as a signal reconstruction model, allowing for the generation or restoration of pyrometer signals.

[0181] This framework enables efficient processing of pyrometer data by leveraging the trained VAE for both feature compression and signal synthesis. The former is applied to the framework for defect detection whereas the latter is applied to the framework for sensor simulation.

[0182] FIG. 9 shows a variational autoencoder framework for processing pyrometer data in an additive manufacturing system although the present invention has applications to other systems requiring a framework for processing pyrometer data. The method can be implemented by a neural network-based system to enhance the processing and analysis of temperature data captured during the manufacturing process. The framework comprises steps including data preprocessing, encoding, decoding, and supporting models for signal analysis. The method begins with data pre-processing of pyrometer data by preparing the raw temperature data captured by the pyrometer sensor for further analysis. The preprocessing may include filtering noise, normalizing the data, and extracting relevant features to ensure the data is suitable for input into the neural network model for dimensionality reduction of signals and the neural network model for signal simulation. The method proceeds to the encoding phase within the neural network. The encoder is responsible for compressing the pre-processed pyrometer data into a lower-dimensional latent space. This transformation allows the system to capture patterns and structures in the data while reducing complexity, facilitating efficient storage and processing. The method then decodes the encoded data. The decoder reconstructs the original pyrometer data from the latent representation, ensuring that important information is preserved. This step plays a role in validatingthe accuracy and reliability of the encoding process, as this demonstrates the system's ability to recover the original data from the compressed form. The method incorporates supporting models for signal analysis that include a neural network model for dimensionality reduction of signals and a neural network model for signal simulation. The dimensionality reduction model further refines the data by reducing complexity, while the signal simulation model generates synthetic data to simulate various manufacturing scenarios, enhancing the system's adaptability and robustness.

[0183] In Fig. 10 the diagram outlines the process flow of ML model training and the dependencies between the ML models and the various data sources. The process consists of three main stages:

[0184] 1. VAE model training - The ML models for the pyrometer and acoustic emissions data are trained independently using the data from the respective sensors as both the input and the ground truth.

[0185] 2. Defect monitoring and sensor simulation subsystems - The ML models for the two subsystems can be trained independently only after the VAEs for both sensors have been successfully trained.

[0186] a. The defect monitoring system is dependent on the encoder portion of the VAEs for the two sensors as they are an inherent part of its neural network architecture for feature extraction. The model uses data from each sensor in addition to the process conditions data as the input, the part quality data obtained through analysis of the parts after production is used as the ground truth for training. b. Both sensor simulation models, one dedicated to the pyrometer and the other on the acoustic emissions sensor, are dependent on the decoder portion of their respective VAEs as they are an inherent part of the neural network architecture for the generation of the synthetic signals. The models use the process conditions data as the input and the corresponding sensor data as the ground truth for training.3. Process control subsystem - The ML model can be trained independently from all other models and subsystems. It uses neither the encoder nor decoder from either of the VAEs. The model uses part quality data, process conditions data, and hashmap library data that is linked to the presence of defects as the input and uses process parameters extracted from the process conditions data known to present no defects.

[0187] Fig. 10 shows the dependencies involved in machine learning model training for an additive manufacturing process. The figure includes Acoustic Emissions Data, Pyrometer Data, Part Quality Data, Process Conditions Data, and Hashmap Library Data. These data sources interact with various systems, including Acoustic Emissions VAE, Pyrometer VAE, Defect Monitoring System, AE Simulation System, Pyrometer Simulation System, and Process Control System. Acoustic Emissions Data plays a role in providing information on transient elastic waves generated during the manufacturing process. This data is processed by Acoustic Emissions VAE, which utilizes variational autoencoder models to analyze and extract meaningful features from the acoustic emissions. The processed data is then integrated into the Defect Monitoring System, which uses the information to identify potential defects and monitor the quality of the manufacturing process. Pyrometer Data captures temperature measurements from the manufacturing environment, which are for maintaining appropriate thermal conditions during the additive manufacturing process. This data is processed by Pyrometer VAE, which applies variational autoencoder models to refine and interpret the temperature data. The refined data is subsequently utilized by the Pyrometer Simulation System to simulate and predict temperature-related outcomes in the manufacturing process. Part Quality Data (PQD) provides insights into the quality and characteristics of the manufactured parts. This data is important for assessing the effectiveness of the manufacturing process and is integrated into the Defect Monitoring System (DMS). The system uses PQD to evaluate the presence and severity of defects, ensuring that the parts meet the required quality standards. Process Conditions Data PCD encompasses various parameters that define the manufacturing environment, such as laser power and scanning speed. This datais processed by the Process Control System, which dynamically adjusts these parameters to optimize the manufacturing process. The PCS ensures that the process conditions are maintained within the desired range to achieve high-quality results. Hashmap Library Data HLD is used to store and retrieve information related to defect detection and classification. HLD is provided to the Process Control System, together with Part Quality Data and Process Conditions Data, to support prediction and refinement of the control parameter set.

[0188] The machine learning algorithm with a physics-based custom kernel is designed to simulate system responses in an additive manufacturing process by incorporating a variety of sensor data inputs, such as temperature and acoustic emissions, to enhance the accuracy of process predictions. This embodiment can be adapted to different materials, such as metals, polymers, or ceramics, by adjusting the kernel's parameters to account for the specific thermal and mechanical properties of each material. The custom kernel may be configured to simulate interactions at various scales, from microstructural changes to macroscopic thermal gradients, allowing for precise control over the additive manufacturing process. In another embodiment, the modular units within the machine learning algorithm can be tailored to specific manufacturing environments, such as high-temperature or vacuum conditions, by modifying the tensor weighting activation functions to reflect the environmental constraints. Additionally, the algorithm can be implemented in various configurations, such as a distributed system where different modules operate on separate computing nodes, enabling parallel processing and faster adaptation to new materials or process conditions. This flexibility allows the system to be deployed in diverse manufacturing settings, from small-scale research labs to large industrial production lines, ensuring optimal performance and defect detection across a wide range of applications.

[0189] In one embodiment, the modular machine learning algorithm is configured to adapt to a wide range of materials by utilizing a sensor data acquisition module that can be adjusted to collect data from various types of sensors beyond just pyrometers and acoustic emissions sensors, such as thermal cameras or laser profilometers,thereby enhancing the system's versatility in different manufacturing environments. The material-specific training module can be tailored to process data from diverse material families, including metals, polymers, and ceramics, allowing the generation of distinct response signatures that optimize the algorithm's performance for each material type. The physics-based custom kernel within the algorithm can be modified to incorporate different physics-driven insights, such as thermal conductivity or material expansion coefficients, which are important for accurately simulating the interactions within the manufacturing process. The modular unit interface is designed to allow independent adjustment of training parameters and neural network components, enabling the system to be fine-tuned for specific material characteristics, such as grain size or surface finish, thus improving the precision of defect detection and process control. Additionally, the output module can be configured to generate a variety of outputs, including real-time alerts for process anomalies or detailed reports on feature integrity, ensuring comprehensive monitoring and control across different manufacturing scenarios.

[0190] The method for standardized testing in an additive manufacturing process according to the invention involves using a variety of standardized test coupons, each with distinct geometries and process parameter variations tailored to different material types, such as metals, polymers, or composites. These coupons can be designed to test specific features like overhangs, thin walls, or enclosed structures, allowing for comprehensive evaluation of the manufacturing process across diverse material properties. The sensor data captured during the printing of these coupons can be processed using a neural network-based data pipeline that incorporates variational autoencoder models, which are adaptable to different sensor configurations, such as varying numbers of pyrometer and acoustic emissions sensors, to accommodate different manufacturing setups. The defect detection dictionary, implemented using a non-cryptographic hashing method, can be customized to include a range of defect types, providing flexibility in defect classification and defect detection. The iterative updating of the machine learning model can be performed using different calibration techniques, such as supervisedlearning with labeled defect data or unsupervised learning for anomaly detection, to rapidly produce a hyper-tuned model that is specifically adapted for the material or material family in question. This approach ensures that the method is versatile and can be applied to a wide range of additive manufacturing scenarios, enhancing the adaptability and precision of in-situ monitoring and control.

[0191] Having illustrated and described embodiments of the invention and certain possible modifications thereto, it should be apparent to those of ordinary skill in the art that the invention permits of further modification in arrangement and detail and is not restricted to the specific embodiments illustrated.

[0192] It will be appreciated that the above description related to the preferred embodiments by way of example only. Many variations on the invention will be obvious to those knowledgeable in the field, and such obvious variations are within the scope of the invention as described and claimed, whether or not expressly described.

Claims

CLAIMS1. A system for monitoring an industrial process, comprising: a sensor subsystem configured to capture process data from at least one sensor selected from the group consisting of pyrometer sensors, acoustic emissions sensors, infrared cameras, thermal imaging sensors, CT sensors, thermocouples and laser displacement sensors; and a defect monitoring subsystem operatively coupled to the sensor subsystem, the defect monitoring subsystem comprising: (i) a sensor fusion module configured to fuse the process data, optionally together with contextual process data, into a unified representation using a neural-network-based data pipeline comprising one or more variational autoencoder models; and (ii) a decision-making module configured to generate at least a defect classification output and a defect probability score from the unified representation, wherein the decision-making module employs a defect detection dictionary to compare the unified representation with stored defect entries and generate a defect detection indication.

2. A system of claim 1, further comprising a process control subsystem communicatively linked to the defect monitoring subsystem and configured to adjust one or more process parameters based on outputs of the defect monitoring subsystem.

3. A system according to claim 1 or 2, wherein the sensor subsystem comprises at least one pyrometer sensor and at least one acoustic emissions sensor.

4. A system of any one of claims 1 to 3 wherein the defect detection dictionary uses a non-cryptographic hashing method together with a Hamming-based similarity measure to compare the unified representation with stored defect entries and generate a defect detection indication.

5. A machine learning-based system for in-situ monitoring of an industrial process, comprising: a sensor subsystem configured to capture process data from at least one sensor selected from the group consisting of pyrometersensors, acoustic emissions sensors, infrared cameras, thermal imaging sensors, CT sensors, thermocouples and laser displacement sensors; and a defect monitoring subsystem operatively coupled to the sensor subsystem, the defect monitoring subsystem comprising:(i) a sensor fusion module configured to integrate the captured process data, optionally together with contextual process data, using a neural-network-based data pipeline to produce a unified representation; and(ii) a decision-making module configured to generate a defect classification output and a defect probability score, wherein the decision-making module employs a defect detection dictionary that compares the unified representation with stored defect entries and generate a defect detection indication.

6. A machine learning-based system of claim 5, wherein the sensor subsystem is configured to capture process data from at least two different sensors and further comprising a process control subsystem communicatively linked to the defect monitoring subsystem and configured to dynamically adjust process parameters based on the unified representation and outputs of the defect monitoring subsystem.

7. The machine learning-based system of claim 5 or 6 wherein the neural network-based data pipeline incorporates variational autoencoder models.

8. The machine learning-based system of any one of claims 5 to 6 wherein the defect detection dictionary uses a non-cryptographic hashing method together with a Hamming-based similarity measure to compare the unified representation with stored defect entries and generate a defect detection indication.

9. The machine learning-based system of claim 7, wherein the sensor fusion module comprises a plurality of variational autoencoder models, each tailored to process sensor data from one of the at least one pyrometer sensor and the at least one acoustic emissions sensor, the variational autoencoder models being configured with sensor-specific pre-processing and feature extraction parameters.

10. The machine learning-based system of claim 6, wherein the process control subsystem is configured to dynamically adjust process parameters selected from the group consisting of laser power and scanning speed based on the unified representation and the outputs of the defect monitoring subsystem.

11. The machine learning-based system of claim 5 or 6, wherein the sensor subsystem further comprises at least one additional sensor selected from the group consisting of an infrared camera, a thermal imaging sensor, and a laser displacement sensor to capture a broader range of process data.

12. A system for sensor fusion utilizing a neural network data pipeline, comprising: a sensor subsystem configured to capture process data, the sensor subsystem comprising at least one sensor selected from the group consisting of pyrometer sensors, acoustic emissions sensors, infrared cameras, thermal imaging sensors, CT sensors, thermocouples and laser displacement sensors; a neural network data pipeline configured to receive sensor data from the sensor subsystem, and fuse the sensor data, optionally together with contextual process data, into a unified output, the neural network data pipeline comprising a plurality of variational autoencoder models, each configured to process sensor data corresponding to one of the at least one sensors; a decision module operatively coupled to the neural network data pipeline and configured to generate a defect classification output and a defect probability score, wherein the decision module employs a defect detection dictionary that compares the unified output with stored defect entries and generate a defect detection indication.

13. The system of claim 12 wherein the sensor subsystem comprises at least one pyrometer sensor and at least one acoustic emissions sensor, wherein the decision module employs a defect detection dictionary that utilizes a non-cryptographic hashing method together with Hamming-based similarity measure to compare the unified output with stored defect entries and generate a defect detection indication.

14. The system of claim 13, wherein the sensor subsystem further comprises at least one additional sensor selected from the group consisting of an infrared camera, a thermal imaging sensor, and a laser displacement sensor.

15. The system of claim 14, wherein the plurality of variational autoencoder models are each configured with sensor-specific pre-processing and feature extraction parameters for processing sensor data from one of the at least one pyrometer sensor and the at least one acoustic emissions sensor.

16. A method for detecting defects in an additive manufacturing process, comprising: receiving process data from at least two different sensors configured to monitor the additive manufacturing process; fusing the received process data using a neural network-based sensor fusion module incorporating variational autoencoder models to produce a unified output including a defect classification value and a defect probability score; applying a non-cryptographic hashing algorithm to the unified output to generate a defect detection dictionary; determining a Hamming-based similarity measure by comparing the unified output with stored defect entries in the defect detection dictionary; and outputting a defect detection indication based on the determined Hamming-based similarity measure.

17. The method of claim 16, wherein the at least two different sensors are selected from the group consisting of a pyrometer sensor, an acoustic emissions sensor, an infrared camera, and a laser displacement sensor.

18. The method of claim 16 or 17, further comprising calibrating the unified output by correlating sensor data obtained from standardized test coupons with corresponding process parameters to refine the defect detection indication.

19. The method of claim 16, wherein the plurality of variational autoencoder models are each configured with sensor-specific pre-processing and feature extraction parameters.

20. The method of any one of claims 16 to 19, wherein determining the Hamming-based similarity measure further comprises comparing the defect detectiondictionary against a predefined threshold to generate the defect detection indication.

21. The method of claim 16, wherein outputting the defect detection indication comprises generating a defect classification and an associated defect probability score.

22. A physics-based custom kernel for a machine learning algorithm, configured to simulate system responses in an additive manufacturing process, the product comprising: a machine learning algorithm incorporating a custom kernel designed to simulate physical interactions within the additive manufacturing environment; the custom kernel configured to adjust tensor weighting activation functions based on physics-based insights, enabling the machine learning algorithm to predict process outcomes with enhanced accuracy; the machine learning algorithm further configured to provide training data for each unique material or material family, producing a response signature that facilitates rapid adaptation of the control system; and modular units within the machine learning algorithm that can be independently adjusted to accommodate specific material characteristics, thereby optimizing the interaction between system components and improving the precision of defect detection and process control.

23. The physics-based custom kernel of claim 22, wherein the custom kernel is configured to dynamically adjust tensor weighting activation functions in response to real-time process condition data, thereby enhancing the predictive accuracy of the machine learning algorithm.

24. The physics-based custom kernel of claim 22 or 23, wherein the machine learning algorithm is further configured to include a material-specific training module that generates a response signature based on selected material properties, enabling rapid adaptation of the control system to different material families.

25. The physics-based custom kernel according to any one of claims 22 to 24, wherein the modular units within the machine learning algorithm are independently adjustable via a modular unit interface to optimize the interaction between system components for improved precision in defect detection and process control.

26. The physics-based custom kernel according to any one of claims 22 to 25, wherein the custom kernel further simulates physical interactions within the additive manufacturing environment by incorporating physics-based insights derived from process parameters selected from the group consisting of laser power, scanning speed, hatch spacing, and layer thickness.

27. The physics-based custom kernel according to any one of claims 22 to 26, wherein the machine learning algorithm is further configured to integrate real-time sensor data from at least one sensor selected from the group consisting of a pyrometer sensor and an acoustic emissions sensor to refine the simulation of system responses.

28. The physics-based custom kernel according to any one of claims 22 to 27, wherein the machine learning algorithm further comprises an output module configured to correlate the generated response signature with predicted process outcomes, thereby facilitating adaptive control based on defect detection.

29. A modular machine learning system for material-specific training, comprising: a sensor data acquisition module configured to collect sensor data and process condition data; a material-specific training module configured to receive and process training data corresponding to a particular material or material family and to generate a material-specific response signature; a physics-based custom kernel integrated within said algorithm and configured to adjust tensor weighting activation functions based on physics-driven insights derived from the sensor data and process condition data; a modular unit interface configured to independently adjust training parameters and neural network components in response to the material-specific response signature; and an output module configured to generate a defect classification output and a defect probability score, wherein the outputmodule employs a defect detection dictionary implemented using a non-cryptographic hashing method together with Hamming-based similarity measure to generate a defect detection indication.

30. The modular machine learning system of claim 29, wherein the sensor data acquisition module is configured to collect sensor data from a plurality of sensors selected from the group consisting of a pyrometer sensor, an acoustic emissions sensor, an infrared camera, and a laser displacement sensor.

31. The modular machine learning system of claim 29 or 30, wherein the physics-based custom kernel is configured to dynamically adjust tensor weighting activation functions in response to real-time process condition data.

32. The modular machine learning system of claim 29, 30 or 31 , further comprising a calibration module communicatively linked to the output module, wherein the calibration module is configured to correlate sensor data with process parameters to refine the material-specific response signature.

33. The modular machine learning system of any one of claims 29 to 32, wherein the sensor data acquisition module further comprises a pre-processing unit configured to remove noise and extract features from the sensor data prior to processing by the material-specific training module.

34. A method for standardized testing for quick model turnaround in an additive manufacturing process, comprising: printing a standardized test coupon having predetermined geometries and preset process parameter variations; capturing sensor data from the additive manufacturing process during printing of said standardized test coupon using a pyrometer sensor and an acoustic emissions sensor; fusing the captured sensor data via a neural network-based data pipeline incorporating variational autoencoder models for each of the sensors to produce a unified sensor data output; generating a defect classification output and a defect probability score from the unified sensor data output; generating a defect detection indication by applying a defect detection dictionary implementedusing a non-cryptographic hashing method together with Hamming-based similarity measure; correlating the unified sensor data output and the defect detection results with the known process parameters of the standardized test coupon to calibrate a machine learning model for adaptive control; and iteratively updating the machine learning model based on said calibration to rapidly produce a hyper-tuned model adapted for a specific material or material family for in-situ monitoring and control.

35. The method of claim 34, further comprising printing said standardized test coupon using a test coupon printing module configured to produce the predetermined geometries and preset process parameter variations.

36. The method of claim 34 or 35, wherein capturing the sensor data further comprises capturing sensor data from at least one additional sensor selected from the group consisting of an infrared camera, a thermal imaging sensor, and a laser displacement sensor.

37. The method according to any one of claims 34 to 36, further comprising calibrating the unified sensor data output by correlating the captured sensor data with the known process parameters of the standardized test coupon to refine calibration of the machine learning model for adaptive control.

38. The method according to any one of claims 34 to 37, further comprising iteratively updating the machine learning model via a feedback loop that includes a calibration module communicatively linked to a process control subsystem to refine process parameter adjustments based on the captured sensor data.

39. The method according to any one of claims 34 to 38, wherein iteratively updating the machine learning model further comprises comparing the defect probability score and / or the Hamming-based similarity measure with a predefined threshold to determine if further adjustments are required, thereby producing a hyper-tuned model adapted for a specific material or material family.

40. A system for standardized testing and quick model turnaround in an additive manufacturing process, comprising: a test coupon printing module configured to produce standardized test coupons with predetermined geometries and preset process parameter variations; a sensor subsystem configured to capture sensor data from the additive manufacturing process during the printing of said standardized test coupons, the sensor subsystem comprising at least one pyrometer sensor and at least one acoustic emissions sensor; a sensor fusion module configured to integrate the captured sensor data into a unified sensor data output using a neural network-based data pipeline incorporating variational autoencoder models for each sensor; a decision-making module operatively coupled to the sensor fusion module and configured to generate defect classification output and a defect probability score from the unified sensor data output; a defect detection module configured to generate a defect detection indication by applying a defect detection dictionary implemented using a non-cryptographic hashing method together with Hamming-based similarity measure; a calibration module configured to correlate the unified sensor data output and the defect detection results with the known process parameters of the standardized test coupons to calibrate a machine learning model for adaptive control; and a model updating module configured to iteratively update the machine learning model based on said calibration to rapidly produce a hyper-tuned model adapted for a specific material or material family for in-situ monitoring and control.

41. The system of claim 40, wherein the test coupon printing module is configured to produce test coupons having predetermined geometries selected from the group consisting of cubes, cylinders, and other standardized shapes.

42. The system of claim 40 or 41, wherein the sensor subsystem further comprises at least one additional sensor selected from the group consisting of an infrared camera, a thermal imaging sensor, and a laser displacement sensor to capture a broader range of process data.

43. The system of claim 40, 41 or 42, wherein the sensor fusion module is configured to aggregate multisensory data from spatially distributed sensor locations in the manufacturing environment to provide a comprehensive unified sensor data output.

44. The system according to any one of claims 40 to 43, further comprising a data storage module operatively linked to the model updating module for storing historical sensor data, calibration parameters, and process control adjustments to facilitate continuous improvement of the adaptive control system.

Images