Geometry-compensated in-SITU eddy current measurement systems and methods

Abstract

Systems and methods are disclosed for detecting, characterizing, and correcting defects and material property variations in additively manufactured components using geometry‑aware analysis of in‑situ sensor data. Multi‑channel sensor data acquired during a layer‑by‑layer build is spatially registered to build coordinates and organized along paths that follow geometric features such as edges, corners, walls, holes, or other evolving features. Sensor responses along the paths are processed to extract characteristic signatures associated with defects and non‑defect indications. Ground‑truth defect information from seeded defects and / or post‑production inspection is used to label the signatures and train supervised machine‑learning models. The trained models are applied to additional in‑situ data to identify defect locations, types, and severity during or after fabrication. The disclosure further includes techniques for estimating material properties from eddy current sensor measurements while compensating for geometric effects using electromagnetic modeling, deconvolution, or combinations thereof, enabling real‑time monitoring, post‑production analysis, and process control.

Description

[0001] GEOMETRY-COMPENSATED IN-SITU EDDY CURRENT MEASUREMENT SYSTEMS AND METHODS RELATED APPLICATIONS

[0002] The present international patent application claims priority under the Patent Cooperation Treaty to U.S. Provisional Patent Application No. 63 / 764,057, filed on February 27, 2025; U.S. Provisional Patent Application No. 63 / 786,664, filed on April 10, 2025; and U.S. Provisional Patent Application No. 63 / 870,533, filed on August 26, 2025, each of which is hereby incorporated by reference in its entirety.

[0003] TECHNICAL FIELD

[0004] The present disclosure relates to the fields of non-destructive evaluation (NDE) and in-situ inspection.

[0005] BACKGROUND ART

[0006] The subject matter discussed in the background section should not be considered prior art merely because of its mention in the background section. Similarly, a problem mentioned in the background section or associated with the subject matter of the background section should not be considered to have been previously recognized in the prior art. The subject matter in the background section merely represents different approaches, which in and of themselves, may also correspond to claimed embodiments.

[0007] In International Patent Application No. PCT / US2025 / 024420 having an International filing date of April 11, 2025, which is incorporated herein by reference in its entirety, Goldfine et al. (hereinafter “Goldfine I”) discloses techniques for analyzing in-situ sensing data collected during industrial and additive manufacturing processes using filtering performed along a selected direction that is not limited to the measurement coordinate axes.

[0008] Goldfine I recognizes that modern manufacturing and inspection systems may acquire large volumes of spatially registered, multi-dimensional sensing data across repeated scans or successive build layers. Conventional data analysis approaches typically process such data along fixed orthogonal directions, such as a scan direction or a build direction. While suitable in some circumstances, these approaches may be limited when geometric features or material transitions extend along non-orthogonal, curved, or evolving paths. In such cases, variations in sensorresponse caused by geometry or gradual process changes may obscure or complicate detection of localized defects or anomalies.

[0009] To address these limitations, Goldfine I discloses organizing measurement data along an arbitrary or preferred direction selected based on part geometry or expected sensor response behavior. The preferred direction may follow a feature of interest, such as an edge, comer, wall, hole, or other boundary that evolves across layers during fabrication, rather than being constrained to the direction of sensor motion or layer deposition.

[0010] In the disclosed approach, measurement data from multiple scans or layers are associated along a path extending through the data volume in the preferred direction. This path may be represented as a “pipe” having a defined cross-sectional area transverse to the preferred direction. Sensor responses within the cross-section are combined to produce a representative scalar value for each scan or layer, thereby forming a vector of responses along the preferred direction.

[0011] The preferred direction is selected such that, in the absence of defects, the sensor response along the path remains substantially consistent. By reducing variation attributable to geometry, deviations associated with localized anomalies become more apparent. The length of the response vector may be limited based on the expected spatial extent of defect signatures to facilitate subsequent processing.

[0012] Goldfine I further describes applying normalization, differencing, or shape-based filtering to the vectors formed along the preferred direction. These vectors may be compared to known signatures or processed using pattern-recognition techniques to identify and characterize defects. Accordingly, Goldfine I provides a framework for analyzing volumetric, layer-by-layer sensing data in a manner that accounts for evolving part geometry and addresses limitations associated with strictly axis-aligned data analysis.

[0013] In U.S. Patent 11,268,933, issued March 8, 2022, which is incorporated herein by reference in its entirety, Goldfine and Dunford (hereinafter “Goldfine II”) disclose a system and method for real-time measurement and feedback of metrology and metallurgical data during additive manufacturing (AM) part fabrication. A sensor is placed either in the rake / roller or following the rake / roller so that it has no impact on the process efficiency and can be used to provide real-time feedback and an archived digital map of the entire part volume. The solution provides non-contact sensing of AM layer's electrical conductivity in a high-temperature environment, metallurgical property verification, porosity imaging, local defect detection and sizing, local material temperature monitoring, and grain anisotropy imaging. Part geometry, theAM powder, and the laser / material interface are monitored in real-time. Real-time nonlinear control of the AM fabrication process is performed based on the sensor data.

[0014] In U.S. Patent 11,268,931, issued March 8, 2022, which is incorporated herein by reference in its entirety, Goldfine and Dunford (hereinafter “Goldfine III”) disclose a method including acts of placing a segmented field sensor proximal to a material under test, the material under test having a metal powder layer and a prior processed material, the segmented field sensor having a first sensing element at a first spatial wavelength and a second sensing element at a second spatial wavelength, the first and second sensing elements being inductive loops; exciting the segmented field sensor with an excitation signal having a first frequency; first measuring a first response of the first sensing element due to the excitation signal, the first response being a single scalar quantity; second measuring a second response of the second sensing element due to the excitation signal; and estimating a proximity of the segmented field sensor to the material under test and a property of the metal powder layer of the material under test from the first and second responses and no other substantially simultaneous response of the first sensing element due to the excitation signal at the first frequency that is independent of the first response.

[0015] In U.S. Patent No. 11,435,317, issued September 6, 2022, which is incorporated herein by reference in its entirety, Goldfine et al. (hereinafter “Goldfine IV”) describes a process for enhancing detection of defects having characteristic shapes provided in a signature library. Signatures, which may be obtained from actual sensor measurements from a known defect and then correlated with sensor measurements. A large correlation may be an indication that a defect is present at the material location where the measurement data was collected. Goldfine IV further describes methods for obtaining signatures for the signature library, selecting an appropriate signature from the library for data, performing single and multichannel correlation, and flagging defect detections.

[0016] In International Application No. PCT / US24 / 55269 having an International Filing Date of November 8, 2024, which is incorporated herein by reference in its entirety, Goldfine et al. (hereinafter “Goldfine V”), further describe the use of signature libraries.

[0017] In International Publication No. WO 2023 / 192887 Al, published October 5, 2023, which is incorporated herein by reference in its entirety, Dunford and Washabaugh (hereinafter “Dunford I”) disclose an eddy current sensor with a remote current sense that has a drive conductor, current sense conductor, and one or more sense conductors. The drive conductor has first and second loop portions, the current sense conductor has a third loop portion, and the senseconductor has a sense loop portion. The first and third loop portions are proximal to each other to form the remote current sense. The sense loop portion and the second loop portion are proximal to each other to form a sense element. The remote current sense and sense element are suitably distant from one another to have separate environments of sensitivity. The sensor may be used by collecting transimpedance measurements from both the remote current sense and sense element under known conditions, and with the sense element under unknown conditions. These measurements are combined to provide a calibrated measurement result suitable for further analysis.

[0018] In U.S. Patent No. 6,188,218, issued February 13, 2001, which is herein incorporated by reference in its entirety, Goldfine et al. (hereinafter “Goldfine VI”) describe calibration of an eddy current sensor “in air.”

[0019] In U.S. Patent No. 10,324,062, issued June 18, 2019, which is herein incorporated by reference in its entirety, Denenberg et al. (hereinafter “Denenberg”) describe a fully parallel, multi-channel impedance instrument.

[0020] In U.S. Patent No. 6,784,662, issued August 31, 2004, which is herein incorporated by reference in its entirety, Schlicker et al (hereinafter “Schlicker”), describes an eddy current sensor array.

[0021] In U.S. Patent No. 7,467,057, issued December 16, 2008, which is herein incorporated by reference in its entirety, Sheiretov et al (hereinafter “Sheiretov”), describes material property estimation using non-orthogonal responsive databases.

[0022] SUMMARY

[0023] Systems and methods are disclosed for detecting, characterizing, and correcting defects and material property variations in additively manufactured components using in-situ sensor data analyzed in a geometry-aware manner. In some implementations, multi-channel in-situ sensor data acquired during a layer-by-layer manufacturing process is spatially registered to build coordinates and organized along spatial paths that follow geometric features of interest, such as edges, comers, walls, holes, or other evolving part features. Sensor responses associated along these paths are filtered and processed to extract characteristic signatures corresponding to defects and non-defect indications. Ground-truth defect information obtained from seeded defects and / or post-production inspection is used to label the extracted signatures and train supervised machine-learning models. The trained models are applied to additional in-situ sensor data to identify defect locations, types, and severity during or after fabrication, enablingimproved defect detectability and reduced false indications. The disclosure further includes systems and methods for estimating material properties from eddy current sensor measurements while compensating for geometric effects using electromagnetic modeling, deconvolution, or combinations thereof. The disclosed techniques support real-time monitoring, post-production analysis, process control, and quality assessment in additive manufacturing and other industrial inspection applications.

[0024] One aspect relates to a method for generating a defect detection model for an additive manufacturing process. The method comprises acts of receiving in-situ sensor data acquired during a first additive manufacturing build of a component; identifying a spatial path that follows a geometric feature of interest of the component; receiving ground-truth defect information for the component; labeling a portion of the in-situ sensor data as corresponding to a ground-truth defect in the ground-truth defect information to form labeled training data; training a machine-learning algorithm as the defect detection model using the labeled training data; and testing the defect detection model using additional in-situ sensor data.

[0025] In some embodiments, receiving the in-situ sensor data comprises receiving

[0026] multi-channel in-situ sensor data.

[0027] In some embodiments, receiving the multi-channel in-situ sensor data comprises receiving eddy current array sensor data acquired during the first additive manufacturing build.

[0028] In some embodiments, the eddy current array sensor data comprises real and imaginary components of transimpedance measurements organized by frequency and sensor channel.

[0029] In some embodiments, receiving the in-situ sensor data comprises receiving sensor data for a plurality of layers of the first additive manufacturing build.

[0030] In some embodiments, the in-situ sensor data further includes position information captured by at least one of a position encoder or a machine motor controller.

[0031] In some embodiments, the method further comprises, prior to labeling, spatially registering the in-situ sensor data to build coordinates.

[0032] In some embodiments, spatially registering comprises converting at least a portion of the in-situ sensor data into material property data tied to part geometry.

[0033] In some embodiments, converting comprises estimating at least conductivity and liftoff from raw sensor measurements.

[0034] In some embodiments, the method further comprises, prior to labeling, performing at least one of removal of corrupted data, handling missing data, timestamp synchronization with amachine controller, layer number indexing, scan path indexing, noise reduction, background subtraction, smoothing, or dynamic range normalization.

[0035] In some embodiments, the method further comprises, prior to labeling, repopulating the in-situ sensor data to account for data density differences in a scan direction by at least one of interpolation or a moving window average.

[0036] In some embodiments, identifying the spatial path comprises identifying a preferred analysis direction that follows at least one of an edge, a comer, a wall, a hole, a boundary, or another evolving geometric feature of the component.

[0037] In some embodiments, identifying the spatial path comprises defining a path through a data volume having a cross-sectional region transverse to the spatial path.

[0038] In some embodiments, the method further comprises combining sensor responses within the cross-sectional region to generate a representative value along the spatial path.

[0039] In some embodiments, identifying the spatial path comprises subdividing a region of interest into a plurality of adjacent paths to increase spatial resolution within the region of interest.

[0040] In some embodiments, the method further comprises, prior to labeling, filtering the in-situ sensor data primarily along the spatial path to obtain directionally organized data.

[0041] In some embodiments, the method further comprises extracting characteristic signatures from the directionally organized data, wherein labeling the portion of the in-situ sensor data comprises labeling at least one characteristic signature.

[0042] In some embodiments, the method further comprises storing the characteristic signatures with metadata including at least one of defect type, size, depth, porosity level, layer number, or process parameters.

[0043] In some embodiments, receiving the ground-truth defect information comprises identifying defects by at least one of deliberately seeded defects introduced during design or fabrication, computed tomography inspection, or destructive metallographic inspection.

[0044] In some embodiments, the method further comprises mapping identified defect locations to build coordinates for registration with the in-situ sensor data.

[0045] In some embodiments, the method further comprises storing defect metadata including at least one of defect type, size, morphology, severity indicators, or spatial coordinates including layer number and scan vector.

[0046] In some embodiments, labeling comprises associating a ground-truth defect with one or more corresponding in-situ responses along the spatial path.In some embodiments, the method further comprises labeling at least one portion of the in-situ sensor data as corresponding to a non-defect indication to train the machine-learning algorithm to suppress false signals.

[0047] In some embodiments, validating the defect detection model comprises validating using in-situ sensor data acquired during a second additive manufacturing build different from the first additive manufacturing build.

[0048] In some embodiments, the method further comprises testing the defect detection model using in-situ sensor data from one or more builds not used to train the defect detection model.

[0049] In some embodiments, the method further comprises determining whether a performance requirement is met and, when the performance requirement is not met, diagnosing a failure mode and expanding a training dataset prior to retraining the machine-learning algorithm.

[0050] In some embodiments, training comprises partitioning labeled training data into training, validation, and test sets.

[0051] In some embodiments, training the machine-learning algorithm comprises training at least one of an artificial neural network, a convolutional neural network, a recurrent neural network, a temporal convolutional neural network, or a physics-informed model.

[0052] In some embodiments, the method further comprises, after validating, applying the defect detection model to additional in-situ sensor data to output a defect prediction identifying at least one of defect location, defect type, or defect severity.

[0053] In some embodiments, applying the defect detection model is performed during the additive manufacturing process for real-time monitoring of at least one layer or scan.

[0054] In some embodiments, the method further comprises, responsive to outputting the defect prediction, using a second sensor or modality to further characterize a defect.

[0055] In some embodiments, applying the defect detection model is performed after completion of the build to perform post-production analysis of an in-situ dataset.

[0056] Another aspect relates to a system for defect detection in an additive manufacturing process, comprising at least one processor operatively coupled to a non-transitory computer-readable storage medium storing instructions that, when executed by the at least one processor, cause the system to receive in-situ sensor data generated from a layer-by-layer additive manufacturing build of a component, obtain a machine-learning-based defect detection model generated according to the disclosed methods, and apply the machine-learning-based defect detection model to the received in-situ sensor data to identify one or more defect locations in the component.In some embodiments, the system further comprises additive manufacturing processing hardware and in-situ sensing hardware, and the machine-learning-based defect detection model is applied during fabrication.

[0057] In some embodiments, the in-situ sensor data comprises multi-channel sensor data acquired during the layer-by-layer build.

[0058] In some embodiments, the in-situ sensor data comprises eddy current array sensor data including real and imaginary components of transimpedance measurements organized by frequency and sensor channel.

[0059] In some embodiments, the instructions further cause the system to spatially register the received in-situ sensor data to build coordinates prior to applying the defect detection model.

[0060] In some embodiments, the instructions further cause the system to identify one or more preferred analysis directions that follow at least one geometric feature of interest and to filter the received in-situ sensor data along the one or more preferred analysis directions.

[0061] In some embodiments, filtering comprises associating sensor responses along a path having a defined cross-sectional region transverse to the path and combining sensor responses within the cross-sectional region to generate a representative value along the path.

[0062] In some embodiments, the instructions further cause the system, responsive to identifying a defect location, to stop the additive manufacturing build.

[0063] In some embodiments, the instructions further cause the system, responsive to identifying a defect location, to revise a build plan for at least one subsequent layer or scan to alter processing parameters.

[0064] In some embodiments, the instructions further cause the system, responsive to identifying a defect location, to cause a second sensor or modality to be used to further characterize the defect.

[0065] Another aspect relates to a system for estimating a material property of a component. The system comprises an eddy current sensor having at least one drive conductor and a plurality of sense elements, an impedance analyzer configured to excite the drive conductor and acquire raw complex transimpedance measurements, a non-transitory computer-readable storage medium storing instructions, and at least one processor operatively coupled to the storage medium, wherein execution of the instructions causes the system to receive the raw complex transimpedance measurements, access electromagnetic model data characterizing an interaction between the eddy current sensor and a geometric feature of the component, determine a geometric correction for the complex transimpedance measurements, apply the geometriccorrection to generate corrected transimpedance data, and estimate the material property of the component.

[0066] In some embodiments, the material property is electrical conductivity.

[0067] In some embodiments, the electromagnetic model data comprises precomputed modeled complex transimpedance responses across a plurality of sensor liftoff values and material conductivity values.

[0068] In some embodiments, determining the geometric correction comprises selecting or interpolating a correction from a lookup table indexed by at least one of distance-to-edge, orientation, curvature, liftoff, excitation frequency, or sense element.

[0069] In some embodiments, applying the geometric correction comprises subtracting a modeled geometric transimpedance contribution from the measured transimpedance.

[0070] In some embodiments, estimating the material property comprises performing inverse interpolation using corrected transimpedance data and a uniform-layer grid database.

[0071] In some embodiments, the instructions further cause the system to estimate sensor liftoff from the corrected transimpedance data.

[0072] In some embodiments, the geometric correction is applied when the sensor is within a threshold distance of the geometric feature.

[0073] In some embodiments, the electromagnetic model data characterizes interactions between the eddy current sensor and the geometric feature across a plurality of build layers.

[0074] In some embodiments, the electromagnetic model data is generated using a

[0075] two-dimensional or three-dimensional finite element model.

[0076] In some embodiments, the electromagnetic model data is generated by modeling interactions with and without the geometric feature.

[0077] In some embodiments, the electromagnetic model data is used to train a

[0078] machine-learning model to predict the geometric correction.

[0079] Another aspect relates to a system for estimating a material property of a component. The system comprises an eddy current sensor having at least one drive conductor and a plurality of sense elements; an impedance analyzer to excite the at least one drive conductor and to acquire from the plurality of sense elements raw sensor response data comprising complex transimpedance measurements; a non-transitory computer-readable storage medium storing computer-executable instructions; and at least one processor operatively coupled to the storage medium; wherein the computer-executable instructions, when executed by the at least one processor, cause the system to: receive the raw sensor response data including the complextransimpedance measurements acquired as the eddy current sensor is positioned relative to the component; access a sensor response function characterizing a spatial response of the eddy current sensor to a geometric feature of a conductive material; apply a deconvolution process to the raw sensor response data using the sensor response function to compensate for geometric distortion of the complex transimpedance measurements; and estimate, from the deconvolved sensor response data, the material property of the component.

[0080] In some embodiments, the material property is electrical conductivity.

[0081] In some embodiments, the sensor response function is empirically characterized using reference geometric features.

[0082] In some embodiments, the sensor response function is generated using electromagnetic modeling.

[0083] In some embodiments, applying the deconvolution process comprises performing frequency-domain deconvolution using a Fourier transform.

[0084] In some embodiments, the sensor response function is derived from a step response and converted to an impulse response.

[0085] In some embodiments, the sensor response function is selected from a library of response functions.

[0086] In some embodiments, the instructions further cause the system to estimate sensor liftoff from the deconvolved sensor response data.

[0087] In some embodiments, the sensor response function is reduced using principal component analysis.

[0088] In some embodiments, the deconvolution process is applied independently to measurements at a plurality of excitation frequencies.

[0089] In some embodiments, the deconvolution process is applied selectively when the sensor is within a threshold distance of the geometric feature.

[0090] BRIEF DESCRIPTION OF DRAWINGS

[0091] The accompanying drawings are not intended to be drawn to scale. In the drawings, each identical or nearly identical component that is illustrated in various figures may be represented by a like numeral. For purposes of clarity, not every component may be labeled in every drawing. In the drawings:

[0092] FIG. 1 A is block diagram of a system for inspecting a test object, according to some embodiments;FIGs. 1B-C show eddy current array sensors, according to some embodiments;

[0093] FIG. 2 is a block diagram of an instrument for measuring and analyzing sensor measurements, according to some embodiments;

[0094] FIG. 3 is an laser powder bed fusion (LPBF) build chamber and illustration of the orientation of an X, Y, Z coordinate system, according to some embodiments;

[0095] FIG. 4A is a flow diagram illustrating a process 400 for supervised machine learning for defect detection along preferred directions using in-situ data, according to some embodiments;

[0096] FIG. 4B is an illustration of the paths used for arbitrary direction filtering of a tubular component, according to some embodiments;

[0097] FIG. 4C is an expanded view of the paths used for arbitrary direction filtering of a tubular component, according to some embodiments;

[0098] FIG. 5 is a flow diagram illustrating a process 500 for correction of material properties at edges and / or complex geometries for eddy current in-situ LPBF data using electromagnetic modeling, according to some embodiments; and

[0099] FIG. 6 shows a flow diagram illustrating a process 600 for correction of material properties at edges and / or complex geometries for eddy current in-situ LPBF data using deconvolution, according to some embodiments.

[0100] DETAILED DESCRIPTION

[0101] Aspects of some embodiments relate to the use of a system 100 for inspecting a test object 130. System 100 is shown as a block diagram in FIG. 1 A. System 100 includes an instrument 110 and a sensor cartridge 140. In some embodiments, system 100 includes a scanner 150 for providing mechanical support for instrument 110, sensor cartridge 140, and / or test object 130 as well as facilitating relative motion between sensor cartridge 140 and test object 130.

[0102] Instrument 110 may be housed in a housing 107; in some embodiments the housing is substantially cylindrical in shape such as that described in U.S. Patent No. 10,416,118, Measurement system and method of use, by Goldfine et al. issued September 17, 2019 and herein incorporated by reference in its entirety (hereinafter “Goldfine VII”). Sensor cartridge 140 may have a rigid connector which interfaces both mechanically and electrically with an instrument side connector 105.

[0103] In some embodiments, sensor cartridge 140 is connected to instrument side connector 105 via cable 180, passing data, power, and communication along conductors 321. Cable 180, shown in FIGS. 4A-4D may be of arbitrary length in accordance with the requirements of theapplication. Although cable 180 is shown with only excitation signals 181 and response signals 183 passing through it, it should be appreciated that cable 180 may also convey other signals (including power). For example, power and / or measurement signals for position encoder 103 may be conveyed through cable 180. Similarly, power and / or control signals for actuator 101 may be conveyed through cable 180. Cable 180 may include a protective covering 325 that is flexible enough to enable inspection for the features of interest and durable enough to prevent kinking, impact damage, tearing and other such failure or damage modes. Cable 180 may also have rigid housings 323 to protect the connector and maintain connection. These housings may be of clamshell design clamping around the connectors 322. Cable 180 may also include mechanical strain relief 324 at each connecting end, and clamshell housing 323 may clamp the strain relief and protective covering to alleviate any mechanical strain on cable. Cable 180 may include mechanical connections 326 to remove strain from the connectors and conductors. Strain relief may extend beyond the clamshell housings to provide mechanical support to the cable where it is contained by the clamshells. Cable 180 conductors 321 may be discrete wires or flexible printed circuit. Cable 180 may have bends 327 in housing 323 to support use of flexible printed circuit construction.

[0104] In some other embodiments, sensor 120 is directly connected to instrument side connector 105. Sensor cartridge 140 in some embodiments also includes a flexible sensor 120, and a mechanical support 141 to which the sensor is attached. Sensor 120 may be attached to mechanical support 141 with glue, tape, double sided tape, or in any suitable way. In some embodiments, sensor 120 is integrally manufactured with mechanical support 141. For example, electrical traces of sensor 120 may be printed (or otherwise formed) directly on mechanical support 141 and possibly buried within mechanical support 141 by adding additional material of mechanical support 141 over such traces. Instrument 110 is configured to provide excitation signals 181 to sensor 120 and measure the resulting response signals 183 of sensor 120.

[0105] Response signals 183 may be measured and processed to estimate properties of interest, such as electromagnetic properties (e.g., electrical conductivity, permeability, and permittivity), geometric properties (e.g., layer thickness, sensor liftoff), material condition (e.g., fault / no fault, crack size, layer to layer bond integrity, porosity, residual stress level, temperature), or any other suitable property or combination thereof including properties of the fabricated part and the powder. (Sensor liftoff is a distance between the sensor and the closest surface of the test object for which the sensor is sensitive to the test object’s electrical properties.) This may include qualification of a repair procedure for the additive manufacturing process, by inspecting thedefect area before and after the repair process to guide the repair process by measuring features of the defect such as location and size estimation and to verify the repair was performed properly and without defects. This may include detection of defects such as excess porosity, voids, variation in metallurgical structure, or variation in surface roughness of built-up material.

[0106] Instrument 110 may include a processor 111, a user interface 113, memory 115, an impedance analyzer 117, and a network interface 119. Though, in some embodiments of instrument 110 other combinations of components may be included. While instrument 110 is drawn with housing 107, it should be appreciated that instrument 110 may be physically realized as a single mechanical enclosure; multiple, operably-connected mechanical enclosures, or in any other suitable way. For example, in some embodiments it may be desired to provide certain components of instrument 110 as proximal to sensor 120 as practical, while other components of instrument 110 may be located at greater distance from sensor 120.

[0107] Processor 111 may be configured to control instrument 110 and may be operatively connected to memory 115. Processor 111 may be any suitable processing device such as for example and not limitation, a central processing unit (CPU), digital signal processor (DSP), controller, addressable controller, general or special purpose microprocessor, microcontroller, addressable microprocessor, programmable processor, programmable controller, dedicated processor, dedicated controller, or any suitable processing device. In some embodiments, processor 111 comprises one or more processors, for example, processor 111 may have multiple cores and / or be comprised of multiple microchips. Processing of sensor data and other computations such as for control may be performed sequentially, in parallel, or by some other method or combination of methods.

[0108] Memory 115 may be integrated into processor 111 and / or may include “off-chip” memory that may be accessible to processor 111, for example, via a memory bus (not shown). Memory 115 may store software modules that when executed by processor 111 perform desired functions. Memory 115 may be any suitable type of non-transient computer-readable storage medium such as, for example and not limitation, RAM, a nanotechnology-based memory, optical disks, volatile and non-volatile memory devices, magnetic tapes, flash memories, hard disk drive, circuit configurations in Field Programmable Gate Arrays (FPGA), or other semiconductor devices, or other tangible, non-transient computer storage medium.

[0109] Instrument 110 may have one or more functional modules 109. Modules 109 may operate to perform specific functions such as processing and analyzing data. Modules 109 may be implemented in hardware, software, or any suitable combination thereof. Memory 115 ofinstrument 110 may store computer-executable software modules that contain computerexecutable instructions. For example, one or more of modules 109 may be stored as computerexecutable code in memory 115. These modules may be read for execution by processor 111. Though, this is just an illustrative embodiment and other storage locations and execution means are possible.

[0110] Instrument 110 provides excitation signals for sensor 120 and measures the response signal from sensor 120 using impedance analyzer 117. Impedance analyzer 117 may contain a signal generator 112 for providing the excitation signal to sensor 120. Signal generator 112 may provide a suitable voltage and / or current waveform for driving sensor 120. For example, signal generator 112 may provide a sinusoidal signal at one or more selected frequencies, a pulse, a ramp, or any other suitable waveform. Signal generator 112 may provide digital or analog signals and include conversion from one mode to another. Denenberg provides a discussion of an impedance analyzer that may be used in some embodiments. See, for example, the discussion in connection with FIG. 19a which provides a discussion on how impedance analyzer 117 can take a measurement. Goldfine VI provides further discussion on how such impedance measurements may be calibrated to remove certain systematic bias from the measurements, according to some embodiments.

[0111] In some embodiments, impedance analyzer 117 has a current sensor 109 that is used to measure a current leaving signal generator 112. Current sensor 109 may be any suitable sensor for measuring such current. For example, current sensor 109 may include a known series resistance in the drive current signal path and current sensor 109 may measure the voltage across such known resistance such that the current may be calculated using Ohm’s Law. As another example, current sensor 109 may measure the voltage induced on an inductive pick-up coil having a well known transimpedance.

[0112] Sense hardware 114 may comprise multiple sensing channels for processing multiple sensing element responses in parallel. As there is generally a one to one correspondence between sense elements and instrumentation channels these terms may be used interchangeably. It should be appreciated that care should be used, for example, when multiplexing is used to allow a single channel to measure multiple sense elements. For sensors with a single drive and multiple sensing elements such as the MWM®- Array eddy current array available from JENTEK® Sensors, Inc., the sensing element response may be measured simultaneously at one or multiple frequencies including simultaneous measurement of real and imaginary parts of the transimpedance (or mathematically equivalent measurements / representations such as the magnitude and phase of thetransimpedance or the in-phase and quadrature components of the transimpedance). Though, other configurations may be used. For example, sense hardware 114 may comprise multiplexing hardware to facilitate serial processing of the response of multiple sensing elements and for eddy current arrays. Some embodiments of sensor 120 use certain MWM-Array formats to take advantage of the linear drive and the ability to maintain a consistent eddy current pattern across the part using such a linear drive. Sense hardware 114 may measure sensor transimpedance for one or more excitation signals at one or more sense elements 123 of sensor 120. It should be appreciated that while transimpedance (sometimes referred to simply as impedance), may be referred to as the sensor response, the way the sensor response is represented is not critical and any suitable representation may be used. In some embodiments, the output of sense hardware 114 is stored along with temporal information (e.g., a time stamp) to allow for later temporal correlation of the data, and positional data correlation to associate the sensor response with a particular location on test object 130. Instrumentation may also operate in a pulsed mode with time gates used to provide multiple sensing outputs and multiple channels used to acquire data from multiple sensing elements. If these sensing elements 123 have different drive-sense gaps (distance between a drive construct 122 and the sense elements 123, then this is referred to as a segmented field sensor. Thus, sensor operation can be at a single frequency, multiple frequencies, or in a pulsed mode where the drive is turned on and off in a prescribed manner or switched between two or more modes of excitation.

[0113] Sensor 120 is shown as an eddy-current sensor, though other sensor types may be used with system 100. For example, in some embodiments, sensor 120 is one or more of an eddy current sensor, an optical sensor, an ultrasonic testing (UT) sensor, a thermographic sensor, and a radiography sensor.

[0114] FIGs. 1B-1C show some eddy current array embodiments of sensor 120. Sensor 120 has a drive construct 122, a sense element 123 (or multiple sense elements), each of which is discussed further herein. In some embodiments sensor 120 provides temperature measurement, voltage amplitude measurement, strain sensing or other suitable sensing modalities or combination of sensing modalities. In some embodiments, sensor 120 is an eddy-current sensor such as an MWM, MWM-Rosette, or MWM-Array sensor available from JENTEK Sensors, Inc., Marlborough, MA. A discussion of some MWM-Array sensors may be found, for example, in Schlicker. Sensor 120 may be a magnetic field sensor or sensor array such as a magnetoresistive sensor (e.g., MR-MWM-Array sensor available from JENTEK Sensors, Inc.), a segmented field MWM sensor, and the like. Segmented field sensors have sensing elements atdifferent distances from the drive winding to enable interrogation of a material to different depths at the same drive input frequency. Sensor 120 may have a single or multiple sensing and drive elements. Sensor 120 may be scanned across, mounted on, or embedded into test object 130.

[0115] In FIG. IB, sensor 120 is an eddy current array having an array of sensing elements 123 and a drive winding 121. Drive winding 121 has a single rectangular drive construct 122. Drive construct 122 has a linear segment along which sensing elements 123 are each equidistant. The distance 124, shown as the distance from the nearest linear drive segment to the center of sensor elements may be defined as a drive-sense gap. Other definitions of drive-sense gap may be found in literature or used - for example, the distance between the nearest linear drive segment and the nearest segment of the sense element coil.

[0116] In FIG. IB, sensor 120 has sensing elements 123 within the confines of rectangular drive construct 122. Sensor 120 shown in FIG. 1C is essentially identical to sensor 120 shown in FIG. IB except that sensing elements 123 outside the confines of rectangular drive construct 122. The drive-sense gap, distance 124, may be defined in the same way for both designs.

[0117] In FIG. ID, sensor 120 has a drive winding 121 where the drive construct 122 is a dual-rectangular drive construct. In some embodiments the dual rectangular drive constructs are connected such that current flows in the same direction in the two adjacent drive segments.

[0118] In some embodiments, the computer-executable software modules 109 may include a sensor data processing module that, when executed, estimates properties of test object 130. The sensor data processing module may utilize multi-dimensional precomputed databases that relate one or more frequency transimpedance measurements to properties of test object 130 to be estimated. The generation of suitable databases and the implementation of suitable multivariate inverse methods are described, for example, in Sheiretov. The sensor data processing module may take the precomputed database and sensor data and, using a multivariate inverse method, estimate material properties for the processed part or the powder. Though, the material properties may be estimated using any other analytical model, empirical model, database, lookup table, or other suitable technique or combination of techniques.

[0119] User interface 113 may include devices for interacting with a user. These devices may include, by way of example and not limitation, keypad, pointing device, camera, display, touch screen, audio input and audio output.

[0120] Network interface 119 may be any suitable combination of hardware and software configured to communicate over a network. For example, network interface 119 may beimplemented as a network interface driver and a network interface controller (NIC). The network interface driver may be configured to receive instructions from other components of instrument 110 to perform operations with the NIC. The NIC provides a wired and / or wireless connection to the network. The NIC is configured to generate and receive signals for communication over the network. In some embodiments, instrument 110 is distributed among a plurality of networked computing devices. Each computing device may have a network interface 119 for communicating with other computing devices forming instrument 110.

[0121] In some embodiments, multiple instruments 110 are used together as part of system 100. Such systems may communicate via their respective network interfaces. In some embodiments, some components are shared among the instruments. For example, a single computer may be used to control all instruments. In one embodiment multiple areas on the test object are scanned using multiple sensors simultaneously or in an otherwise coordinated fashion to use multiple instruments and multiple sensor arrays with multiple integrated connectors to inspect the test object surface faster or more conveniently.

[0122] Actuator 101 may be one or more actuators used to position sensor cartridge 140 with respect to test object 130 and ensure that the liftoff of the sensor 120 is in a desired range relative to the test object 130. For example, actuator 101 may drive the movement of mechanical components of scanner 150 that in turn move the sensor 120 relative to test object 130. Actuator 101 may be an electric motor, pneumatic cylinder, hydraulic cylinder, or any other suitable type or combination of types of actuators for facilitating movement of sensor cartridge 140 with respect to test object 130. Actuator 101 may be controlled by motion controller 118. Motion controller 118 may control actuator 101 to move sensor cartridge 140 and sensor 120 relative to test object 130.

[0123] Regardless of whether motion is controlled by motion controller 118 or directly by the operator, position encoder 103 and motion recorder 116 may be used to record the relative positions of sensor 120 and test object 130. This position information may be recorded with impedance measurements obtained by impedance analyzer 117 so that the impedance data may be spatially registered.

[0124] For some applications the performance of system 100 depends (among other things) on the proximity of sensor 120 to test object 130; that is to say the sensor liftoff may be critical to performance for such applications. For example, crack detection in an aerospace application may require cracks 0.5 mm (0.02 inches) in length be reliably detectable in test object 130 (e.g., a turbine disk slot). In order to achieve reliable detection of a small crack, sensor 120’ s liftoff mayneed to be kept to under 0.25 mm (0.010 inches). Further, for such an application, sensor 120 may preferably be a sensor array, thus the liftoff of each element in the array may need to be kept to under 0.25 mm (0.010 inches). (It should be appreciated that these dimensions are illustrative and the specific requirements will be dictated by the details of the application.) Measurements may be complicated when test object 130 has a complex curved surface that may change along a measurement scan path.

[0125] To permit high-performance operation at higher excitation frequencies, use of current sensor 109 to measure the current in drive winding 121 may not be sufficient. The inventors have recognized and appreciated that measurement performance may be improved by measuring the current in drive winding 121 closer to the portion of the drive conductor that is inductively coupling to sense element 123. Specifically, and as described in Dunford I, a current sense element located on sensor 120 can be used to much more accurately measure the current in drive winding 121 that is inductively coupling to sense element 123. This is contrasted with measurement of the drive current much further from sense element 123 using current sensor 109 which is typically within instrument housing 107. Although the electrical impedance of cable 180 may alter the current at the instrument, the local measurement can account for any variation of the current due to the cable.

[0126] FIG. 2 shows embodiments of instrument 110 with a specific focus on data collection and analysis. It should be appreciated that other aspects of instrument 110 discussed in connection with FIG. 1 A or elsewhere may also be part of such an embodiment.

[0127] Prior to using instrument 110 to collect and analyze sensor data as part of system 100, instrument 110 may be configured for a specific measurement application. An instrument control module 230 may be used to configure instrument 110 for a specific measurement application. Instrument control module 230 may utilize a session file 210 to store an instrument configuration 211, a measurement sequence instructions 212, and an interpolation configuration 213.

[0128] Instrument configuration 211 may store information identifying the type of sensor to be used, the excitation frequencies and their respective amplitudes, specific grids within precomputed database 203 for impedance data interpolation, the type of calibration to be used, the modules that are used as part of the measurement such as the specific signatures within signature library 205 for data analysis, and other information for configuring instrument 110 for a measurement application. The calibration typically uses an air calibration or an air with a one point reference measurement calibration. The term “air calibration”, as used herein, is calibrationaway from any conducting or magnetic materials (e.g., in a gas, air, vacuum, or in the presence of non-conducting, non-magnetic materials). This requires that conducting or magnetic materials are far enough away from the sensor so as to not impact the sensor response. For an air calibration itself, a measurement of the sensor response in air is used to adjust the measurement impedances to known and reproducible values. This approach does not require the use of reference standards for the instrument adjustment, but measurements on a reference part or material is recommended for verification of the calibration itself. To reduce channel -to-chann el variations in the sense element responses and improve consistency of the conductivity measurement, a second measurement point can be used as part of the calibration. This second measurement is usually for a reference material with known electrical properties. This provides consistency with other standard procedures for conductivity measurements. Note that one or more reference point measurements could be used but this tends to be less robust than including a measurement response in air since the reference part measurement for calibration requires knowledge of the conductivity of the reference material. The instrument configuration 211 typically also includes information about the data acquisition rate and the configuration of auxiliary information that could be associated with each measurement such as position encoder information, temperature, strain gages, etc.

[0129] Measurement sequence instructions 212 may define the sequence of actions that are to take place for a measurement. Instructions 212 may specify motor control, triggers, changes to the instrument configuration, and prompt user actions. For example, instructions 212 may indicate that after initializing a measurement, a first motor is to move at a certain speed during measurement collection and, after reaching an end point, measurement is to stop. As another example, after a first measurement is taken the instructions 212 may indicate the user is to be prompted to take an action (e.g., lay a non-conducting layer between the test object and the sensor to increase sensor liftoff) and then wait until a user initiated trigger is received. As yet another example, after taking first measurements the instructions may cause instrument 110 to be reconfigured to an alternate instrument configuration (e.g., having different excitation frequencies or other configuration properties).

[0130] Measurement sequence instructions 212 may also include definitions of the views to be presented to the end user. These views may be read by graphics generation module 270 to affect the graphical presentation to the user. Note that the graphics generation could also be in the form of data tables.In some embodiments, an inverse interpolation module 220 is used to process impedance data 201 obtained from sensor 120 by impedance analyzer 117. Inverse interpolation module 220 utilizes a grid database 203 to estimate physical properties from impedance data 201.

[0131] Physical properties estimated may include properties such as layer and gap thicknesses, electrical conductivity as a function of spatial position, and magnetic permeability as a function of spatial position. For example, the physical properties estimated by inverse interpolation module 220 for a sensor scanning a coated substrate material may include (i) liftoff, (ii) coating thickness, (iii) coating electrical conductivity, and (iv) substrate electrical conductivity.

[0132] Secondary properties may also be estimated using the output of the inverse interpolation module, such as layer thicknesses, gaps between layers, the size of a chamfer, and further to estimate defect sizes or to identify types of anomalies.

[0133] Interpolation configuration 213 of session file 210 may be used to specify aspects of the inverse interpolation. For example, in some embodiments a hierarchical approach can be used to increase numerical stability and accuracy of the multiple unknown inversion. Property effects can be systematically separated from one another by using specific excitation frequencies and / or segmented fields to estimate the properties they are most sensitive to. For example, a coating conductivity property may be estimated using only a high frequency excitation measurement, and then both the high and a low frequency used to determine coating thickness and substrate conductivity (with the coating conductivity in this second step assigned the value determined from the high frequency alone). This may be useful for example for inspection through a bushing. In one embodiment of this invention, the use of multiple frequencies and the inverse interpolation module along with the ability to scan the internal surface of the busing in a hole, is used to detect cracks in the stackup / layers / skins through a bushing. This can utilize other aspects of this invention, after accounting for the thickness and properties of the busing to estimate the conductivity or magnetic permeability of the “substrate” which in this case is the properties of the structural layers, aircraft skin, material that is inspected through the busing.

[0134] Further discussion of the operation of inverse interpolation module 220 may be found in Sheiretov.

[0135] In some embodiments, instrument 110 is also equipped with a forward model module 240 for precomputing grids for grid database 203 using a sensor-material model. The model may be a physics-based model, an empirical model based on prior measurements, or any other suitable type of model for creating measurement grids. In some embodiments, forward model module 240 is not made a part of instrument 110 and only grids are stored in grid database 203of instrument 110. For example, forward model module 240 may be a software application run on a computer to produce grids which are then stored in grid database 203.

[0136] In some embodiments, instrument 110 includes a signature definition module 250 for defining characteristic responses (“signatures”) of a feature to be enhanced or suppressed in measurement data. In some embodiments signatures and their use may be similar to those described in Goldfine IV and Goldfine V. Signature definition module 250 may allow a user to identify signatures and store them in a signature library; alternatively or additionally, signatures may be identified in an automated or semi -automated way. For example, a crack defect signature may appear in the electrical conductivity response measured by a sensor scanning over the crack. In the case of a sensor array, the response may be observed on a single or multiple adjacent channels. A signature may be identified as a single channel response or a multi-channel response. Signature definition module 250 may standardize signatures prior to storing them in library 205. For example, signatures may be standardized to a specific number of points or a specific amplitude range. Signatures may also include metadata that provide additional information about the signature such as the size of the defect the signature was obtained from.

[0137] Detection and sizing module 260 may be used to detect and size defects in measurement data from a test object using signatures from signature library 205. Module 260 may evaluate the correlation between a measurement and a signature. If the correlation exceeds a threshold a detection may be flagged. The threshold may be set based on the detection and false alarm requirements of the application. Signature library 205 may contain multiple signatures that may be tested against measurement data. The signature having the greatest similarity with the measurement may also be used to size a detected defect. For example, the defect size may be estimated to be the same as the size of the defect the signature.

[0138] Module 260 may also be used to suppress features that are not of interest such as fasteners or through holes. For example, a through hole in a plate typically has a significant effect on the estimated electrical conductivity of the substrate material if a planar model is used to estimate conductivity. The shape of the conductivity response with respect to position as the sensor is scanned over the hole depends upon the actual electrical conductivity of the substrate material, the excitation frequency, and the geometry (e.g., sense element size and spatial wavelength) of the sensor. However, for a given sensor array, because the conductivity response of the through hole is consistent, it may be removed from the conductivity estimate. For example, module 260 may identify a highly correlated through hole signature with the conductivity response from measurement. The conductivity response may then be updated toremove the signature. This will flatten the conductivity response and may also allow for the hole location to be accurately estimated from the measurement data. While this example discussed suppressing the response for processed data such as the estimated conductivity of the material this approach can also be used for unprocessed data such as the sensor impedance or transinductance.

[0139] Graphics generation module 270 may provide a graphical representation to the user to assist the user in the data collection and / or analysis process. Module 270 may present such a graphical presentation on a video display integral to and / or separate from instrument 110.

[0140] Information may be presented as tables, A-scans, B-scans, C-scans, or any suitable way. In some embodiments, module 270 configures the graphical environment based on instructions 212. In this way a consistent presentation of information can be provided to the user.

[0141] Report Generation Module 280 may be included to facilitate review of measurement results outside of the graphical environment of instrument 110. For example, report generation module 280 may produce a report of measurement data in pdf, docx, rtf, xlsx, or other suitable format. Session file 210 may specify the report format which may be used by module 280 to generate reports for measurement data.

[0142] In some embodiments, the output includes a decision with regards to the future disposition of the test object. Modules 270 and / or 280 may present such a decision. Examples include pass / fail decisions on the quality of a component, or the presence of flaws. As another example, it may be determined whether the test object may be returned to service, repaired, replaced, scheduled for more or less frequent inspection, and the like. If it is determined that the application was not determinative, instrument 110 may re-perform the procedure(if automated), or advise the user to re-perform the procedure. A procedure may need to be re-performed, for example, if all requirements of the procedure were not met. For example, the procedure may require the liftoff of the sensor to be below a threshold amount over the inspection surface and require re-performance if the liftoff requirement is not met.

[0143] Software implementations are focused on increasing data analysis, speed and large file handling. In one such embodiment, software is upgraded from 32 bit to 64 bit to improve the speed of data analysis and rendering of multiple images for viewing analyzed data. In some embodiments, multithreading, vector processing, or other methods for implementing rapid data analysis for multivariate inverse methods, intelligent filtering, or AI / ML implementations are utilized to improve the speed for data analysis and viewing. Inspection speed may also be affected by the scan speed of motors and the data acquisition rate. The scan speed of motors anddata acquisition rate are set to provide a minimum number of data points on a prescribed defect size (e.g., 50x50 mil comer cracks, or 50x25 mil midwall cracks) where the minimum number of data points required is determined from empirical data taken at two or more scan speeds / data rate combinations. Scan speeds and data rate should also take into account the type of data analysis to be implemented (e.g., MIM, intelligent filtering, AI / ML). Intelligent filtering is the use of signature libraries and shape filtering as described, for example, in Dunford I.

[0144] In some embodiments the measurement results are used to control a process. For example, a property measurement may be fed back into a control circuit that controls a process.

[0145] Attention is now turned to FIG. 3, which shows a Laser Powder Bed Fusion (LPBF) system 300 with sensor 120 attached to recoater 303. The x-direction is taken to be the scan direction 302 and the y-direction is perpendicular to the scan direction and in the plane of the ET array. Sensor 120 may be a part of system 100 and system 100 may be used with system 300 in any suitable way. If the sense elements of the array are aligned perpendicular to scan direction 302, this corresponds to the y-direction. Again, the z-direction is then normal to scan direction 302. Recoater 303 passes over build plate 301 on which the test object is built by the LPBF process.

[0146] In LPBF additive manufacturing, ensuring the quality of printed components is critical, and in-situ monitoring techniques can provide defect detection without costly post-production inspection. A defect, in the context of a production component, refers to any unintended geometric, microstructural, or material property deviation from the intended design that may adversely affect performance, reliability, or service life of the component. Among in-situ data modalities, eddy current array sensing offers distinct advantages by enabling the estimation of material properties during the build process as well as providing subsurface information (below the most recently printed layer). To leverage these advantages the described method utilizes the in-situ eddy current data to effectively identify and characterize defects, thereby supporting improved process control and part reliability.

[0147] FIG. 4A shows a flow chart illustrating a process 400 for supervised machine learning for defect detection along preferred directions using in-situ data. Process 400 may be implemented using system 100 or in any suitable way. Process 400 is generally described in connection with in-situ data collected during a laser powder bed fusion (LPBF) fabrication process where the test object 130 (FIG. 1) is a component being fabricated. Though, process 400 may be applied to any suitable application. Recall that the LPBF process involves fabricating a part or component inside a build chamber by placing a powder onto a build plate and using alaser to selectively heat portions of the powder according to a prespecified build pattern. After the heating process, the build plate is lowered, powder is reapplied by passing a recoater or spreader across the build plate, and then applying the laser heating. This is the fabrication process for each layer of the built-up material of the component or part. Process 400 involves filtering or analyzing measurement data in a direction that follows a feature of interest to detect indications of interest, and using that indication information to train an Al / ML model. In this context the indication could be a property variation that does not adversely affect performance of the fabricated material or a defect that adversely affects the performance. Separate sets of data are then used to validate and test the Al / ML model.

[0148] Process 400 begins at step 401 with receiving in-situ data that was acquired during an additive manufacturing build. This includes a full set of inspection data for each of the relevant layers of the build process. The relevant layers may be for a completed build for post-fabrication analysis or may include only a subset of layers or the most recent layer for an ongoing build. In some embodiments, the raw data received is the real and imaginary parts of the transimpedance measurements obtained from an eddy current sensor array, organized by frequency and sensor channel number. This data may also contain position information that was captured by a position encoder, machine motor controller, or other suitable position measurement method.

[0149] At step 403, the in-situ data is preprocessed and aligned to build coordinates. This step involves converting raw sensor data to spatially and temporally registered ML-ready material property data tied to part geometry and eventual truth inspection. This ensures that every data point can be traced to build layer, scan vector, and process parameters, as applicable. In some embodiments, the scan data is the real and imaginary parts of the transimpedance measurements obtained from the sensor. In this case, a preprocessing step involves converting the raw impedance data to material property information, such as material conductivity and lift-off. In some embodiments, a multivariate inverse method is used in combination with a database of precomputed sensor responses to estimate the material properties. In another embodiment, an Al / ML model is trained to map impedance to material property values, based on an air calibration, and this model is used to estimate the material properties, Additional typical preprocessing tasks included in step 403 are removal of corrupted data, handling of missing data such as partial acquisitions, repopulation of data to account for data density differences, timestamp synchronization with the machine controller, and layer number and scan path indexing. If applicable, metadata tagging (laser power, speed, hatch spacing, host material, etc.) of the data is useful at this step. Additional preprocessing steps relate to noise reduction which may includetime based and spatial filtering to reduce background variation or identify conditions of interest, background subtraction, smoothing filters, and dynamic range normalization. At Step 403, the geometric registration of the data to build coordinates is performed, which may involve mapping the sensor data to machine coordinates, part geometry, and / or scan vectors, as applicable. These preprocessing steps are not a complete list but are meant as examples of typical tasks that may be applied in certain embodiments. Since the acquired data density is not necessarily uniform in the scan direction, for example from non-uniform scanner rates as the recoater accelerates and decelerates. This step may also involve converting the measurement data into a specified spatial density in the scan direction as well as the array channel direction. This conversion could involve interpolation between data points or applying a moving window average.

[0150] At step 405, preferred filtering directions are identified. The preferred directions could be provided through human input, or via a software module that automatically identifies the features using predefined requirements. The filtering direction may follow a feature in the test object, such as a corner or edge that is linear or curved. This is illustrated in FIG. 4B where several of these analysis paths or “pipes” 421 are shown near the comer of a tubular structure 420. FIG. 4C shows an expanded view of the corner of the tubular structure 420 where the corner region 423 has been subdivided to contain many individual paths to provide a small cross-sectional area 425 compared to the area of the corner region to provide a higher spatial resolution in the region of interest. This illustrates that for finer spatial analysis, additional pipes can be used that follow a path that is adjacent to the original pipes. This is similar to the generation of meshes for finite element methods. Additional implementations of these pipes are described in Goldfine I. In one embodiment, the data is from a bolt hole scan, and the preferred direction is in the plunge direction to follow the axial direction or near other edges. In another embodiment, the data is from risers with wires, and the preferred direction follows the wires along the length and assesses the spacing between the wires and the presence of cracks in the wires themselves. In some embodiments, Step 405 may involve locating such features of interest as edges, corners, holes, wall thickness transitions, and / or areas of the test object where stress is expected to concentrate during the design use of the test object (and thus is particularly susceptible to damage). These located features are then used to identify the preferred filtering directions. In one embodiment, the interaction of the sensed energy source with a build feature is used to determine the preferred directions. In another embodiment, the preferred direction increments are periods of time, in addition to or separate from spatially spaced increments. In some embodiments, the preferred directions are selected to follow geometric features such asedges and corners and to fill the space away from edges and corners so as to improve the computational efficiency, similar to how finite element meshes are finer near geometric features and coarser away from them. It should be appreciated that the varying spatial dimensions associated with a fine or coarse mesh may require interpolation of the received inspection data in the scan direction and / or the array channel direction. It should also be appreciated that step 405 is not limited to identifying a single preferred filtering direction, and may identify multiple directions such as, for example, all comers and edges as well as the volume within the built-up material. Furthermore, the identified filtering directions can be divided into regions that are more likely to form defects than other regions and regions that are more critical to component performance if defects are formed. Given the ranking of regions based on severity criteria that includes both likelihood to form defects and the criticality of locations, in one implementation the resolution of the preferred filter pipes would be adjusted higher in more critical or more likely defect areas. In one such embodiment, only these higher ranked areas are processed in real-time to enable adjustment of the build process to repair detected defects. In all such embodiments, the preferred filtering directions are to be utilized with the preprocessed aligned data.

[0151] At Step 407, defect locations and characteristics are identified and documented to establish high-fidelity ground truth data. Though, in some embodiments, the information and results of this step are collected or generated outside of process 400 and process 400 receives that information at this step. This step catalogues actual defects present in the build parts, including details such as their type, size, and morphology, and mapping their location to build coordinates. Note that other indications, not defects, may also be of interest. Defect data may be sourced from complementary approaches to ensure comprehensive coverage of relevant features of interest. In one embodiment, deliberately or intentionally seeded defects during the part design and fabrication are identified. Examples of these intentionally introduced defects include embedded voids of controlled size and location in the CAD model, varied laser power, speed, or hatch spacing, and layer thickness variations in targeted regions to induce predictable features of interest. In one such embodiment, empirical studies are performed, in some cases using design of experiment, to minimize the number of builds required to generate a representative range of defects and material changes (e.g., porosity, micro structure) for both good and bad quality parts. In another such embodiment, defects of at least one family of shapes (e.g. horizontal cylinders) are included in a build at known locations with varied sizes / dimensions. In another such embodiment, the defects are filled with powder and processed differently to mimic metallurgicaldefects. In another such embodiment, defects are intentionally placed at internal edges, corners or other such features in a part to enable training on relevant defects at critical locations. In another such embodiment, multiple families of defects with different shapes are used to train the algorithm. In one such embodiment, a sufficient number of families are included to allow a generic defect to be characterized and sized. In another embodiment, defects are identified through post-production inspection methods such as Computed Tomography (CT), destructive metallographic cross-sectioning, or other non-destructive or destructive evaluation techniques. Post-production inspection is recommended, when feasible, as it serves to verify intentionally seeded defects, quantitatively characterize their attributes (e.g., actual size, morphology, etc.), and detect any unintended defects introduced during fabrication. For each identified defect, regardless of approach, descriptive metadata is recorded and stored in a structured format, and spatial coordinates are mapped to build coordinates for registration with in-situ datasets.

[0152] Descriptive metadata typically includes information such as detect type (porosity, void, crack, etc.), size metrics (volume, diameter, etc.), morphology (aspect ratio, sphericity, etc.) severity indicators (clustering, proximity to critical features, etc.), and spatial coordinates (layer number, scan vector, etc.). In some embodiments, the built-up part with seeded defects is fatigued and subsequently undergoes post-production inspection to identify the effects of the fatigue. In such an embodiment, this data is also included in the metadata as a time-based element for life prediction.

[0153] At Step 409, the in-situ data, which was preprocessed and aligned to build coordinates in Step 403, is filtered and analyzed primarily along the preferred directions that were identified in Step 405. During Step 409, the data is analyzed to identify the characteristic process signatures that correspond to the defect locations identified in Step 407, and these signatures are labeled with metadata for Al / ML defect prediction models. The objective of this step is to associate each ground-truth and / or seeded defect with its corresponding in-situ response(s). This step systematically scans through all preferred filtering directions, with an emphasis on any regions of interest or areas with ground-truth defect coordinates. In one embodiment, the signatures stored for each defect are material property variations along a vector in space and / or time, and may include data at multiple frequencies and / or along multiple channels. The defect signatures are stored along with metadata such as defect type, size, depth and other geometric features, porosity level, layer number, laser power, speed, hatch spacing, host material, and may also include all sensor array information, frequency, channel, LPBF machine type, etc. This data also includes the vector or trajectory of the signature in space and / or time. In some embodiments, thisstep also identifies and stores signals of non-defects. That is, signals in the in-situ data that appear to be features of interest, but were not validated in the post-production inspection. In one embodiment, this information is used to further train the Al model to suppress false signals in the subsequent steps. In another embodiment, data augmentation may be used at this stage to artificially expand the diversity and quantity of the defects without needing more physical builds. Depending on the format of the in-situ data, the data augmentation could include addition of noise and time jitter, rotation of defect signatures or responses, addition of scaling and / or amplitude noise. In some embodiments, physics-based data augmentation is performed by utilizing a numerical model of the process and perturbing parameters in the model to generate plausible changes in part outcome. In one such embodiment, the empirical studies are combined with the electromagnetic modeling or other such numerical models to expand the training set.

[0154] At Step 411, supervised learning models are trained to learn the relationship between in-situ defect signatures and ground-truth and / or intended defect data. Step 411 begins by partitioning the labeled data into training, validation and test sets. It is preferable that entire builds are reserved for the test set. In some embodiments, the test set should contains at least 10-20% of the total available data. This step may utilize different types of Al / ML models depending on the type of in-situ data and the goal. In some embodiments, where the data is complex, deep learning models such as artificial neural networks (ANNs), convolutional neural networks (CNNs), recurrent neural networks (RNNs) or temporal CNNs, or hybrid or physics-informed els may be used. The training process involves feeding the labeled examples into the model repeatedly and adjusting its internal settings to improve its predictions. The training may take place on CPUs, GPUs, a combination of the two, or in any other suitable way. After training, the model’s performance may be checked through validation on a separate portion of the data that was not used for training. This helps confirm how well the model generalizes to new situations and guides any final adjustment to the model. The end result of Step 411 is a reliable model that can take new in-situ data during a build and output predictions about where defects are likely to appear, what kind they might be, how severe they could be, and helping to catch issues early and improve part quality. In some embodiments, as more data is available, it may be used to continually improve the training of the algorithm.

[0155] At step 413, the model is tested by applying it to independent sets of data that were not used in the training or validation. This data is typically from separate builds and may include different geometries, materials, or process conditions than those used during training and validation. The primary goal of this step is to confirm that the model can accurately detect andcharacterize potential defects from unseen in-situ sensor readings. The requirements that are typically checked include high detection reliability for actual defects while keeping false calls low to avoid unnecessary interruptions or rework. This may utilize performance metrics such as Probability of Detection (POD), Repeatability and Reproducibility (R&R), and / or other statistical assessment of performance. Other key expectations are good performance across varying conditions such as different layer thicknesses, laser settings, or part geometries, as well as rapid computational speed for potential real-time use during a build.

[0156] At step 415, an evaluation decision is made. If the testing phase meets the predefined requirements (e.g. a successful POD study) on the independent test sets, one may consider the model effectively ready for production use. If the model fails to meet the requirements, the process proceeds to step 417. At Step 417, the failure mode of the model must be diagnosed (insufficient representation of specific defect types, overfitting, etc.). Based on this analysis, the training dataset may be expanded to include additional data from other builds or targeted samples that better represent the conditions under which the performance was inadequate. If examples from the original test set are used to update the model, those examples must be removed from the test set and assigned to the training and / or validation pool and cannot be used as subsequent test data. After updating the dataset, the process returns to Step 411 to retrain and validate the Al / ML model.

[0157] Once the Al / ML model has been thoroughly trained, validated, and tested on representative data, it is ready for production deployment in the additive manufacturing process. In some embodiments, this model is deployed on a CPU, GPU, a combination of the two, or on another suitable platform. The Al / ML model may be integrated into the workflow in various ways. For example, in one embodiment, it is used for real-time monitoring and is continuously fed in-situ sensor data as each layer or scan completes, or in batches. This enables real-time detection of potential defects, allowing either the stopping of bad parts to save time and money or to enable the process to be altered to improve the outcome of the part being built. In one such embodiment, once a defect has been located, a second sensor or modality may be used for further characterization of the defect. For example, the second sensor or modality may be used to determine the size, length, depth or to track an event (follow a specific feature). In another embodiment, the Al / ML model processes the complete in-situ dataset after the build finishes, performing post-production analysis. Regardless of whether the model is deployed during the build or afterwards, or both, the model reduces the reliance on time-consuming and expensive post-production inspection.Attention is now turned to some systems and methods for improving material property estimates using electromagnetic modeling and / or deconvolution using system theory. For eddy current array data, some methods have been previously disclosed for rapidly estimating material properties from measured impedance responses (e.g., see Sheiretov). These methods may utilize a multivariate inverse method in combination with a precomputed database of sensor responses to estimate the material properties. The precomputed database may be generated using a uniform layer model. This model assumes the material under inspection is an infinite, uniform layer with constant electromagnetic properties (e.g. conductivity and permeability) and contains no edges, corners, or other complex geometric features. The uniform model-based grids provide highly accurate metallurgical property estimates in regions sufficiently far from geometric features. However, LPBF printed components are well known for their complex features, which can cause eddy current paths to be truncated and distorted (effects that are not reflected in the model or database). Therefore, to provide accurate material property estimates in the presence of such geometric features, two correction methods are disclosed, shown in FIG. 5 and FIG. 6, and described below.

[0158] FIG. 5 shows a flow chart illustrating a process 500 for correction of material properties at edges and / or complex geometries for eddy current in-situ LPBF data using electromagnetic modeling. Process 500 may be implemented using system 100 or in any suitable way. Process 500 is generally described in connection with in-situ data collected during a laser powder bed fusion (LPBF) fabrication process where the test object 130 (FIG. 1) is a component being fabricated. Though, process 500 may be applied to any suitable application.

[0159] Process 500 begins at step 501 with receiving in-situ data that was acquired during an additive manufacturing build. This includes a full set of inspection data for each of the relevant layers of the build process. The relevant layers may be for a completed build for post-fabrication analysis or may include only a subset of layers or the most recent layer for an ongoing build. In some embodiments, the raw data received is the real and imaginary parts of the transimpedance measurements obtained from an eddy current sensor array, organized by frequency and sensor channel number. This data may also contain position information that was captured by a position encoder, machine motor controller, or other suitable position measurement method.

[0160] At step 503, the material properties of each layer are estimated across the xy-plane in regions sufficiently far from edges and / or complex features where uniform layer modeled responses are highly accurate. In some embodiments this step may be omitted, for example, if the printed component does not contain any such regions. In some embodiments, where the rawdata is the real and imaginary parts of the transimpedance measurements obtained from the sensor, this data is converted to material property information using a multivariate inverse method in combination with a database of precomputed sensor responses to estimate the material properties.

[0161] At step 505, electromagnetic models of the eddy current field interference at edges and / or complex geometries are run to determine a correction for the impedance measurements and subsequent property estimates. This may be accomplished using 2D or 3D finite element modeling (FEM) software, or other suitable method. Either the entire build file geometry may be imported into the software, or only the geometric regions of interest. The eddy current array, or relevant portions of the eddy current array, is included in the model as well as air regions around the array and part. A fine mesh is utilized near edges and in sense element regions and may be coarser elsewhere. The induced voltages / impedances at each sense element are determined from the model and stored for each measurement frequency and scan position. Since the in-situ data is obtained for each build layer (e.g., every 40 microns, in some cases), the model is run at various layers that represent the different ways the sensor array may encounter a geometric feature as the part is built up. The model may also be run across a range of lift-offs, material conductivities, and sensor configurations, as well as varying in-plane material orientations and normal to plane geometric variations. For example, an overhang located a millimeter below the inspection plane may influence the eddy current sensor response and should therefore be accounted for in the modeling. The data is stored to allow position mapping back to the in-situ data. In some embodiments, the FEM model is run both with and without the part geometry and the difference between these is stored as a correction factor. In one embodiment, the electromagnetic model is set up and run manually for specific build(s) or features of interest. In another embodiment, an automated supervisor is utilized. The automated supervisor may determine and / or receive the features of interest, load the geometries into the electromagnetic model and set up other relevant parameters, run the model, and export the data back to the processing algorithm. In one embodiment, the FEM is used to determine the distance from each edge at which point the uniform layer model is no longer valid and a correction needs to be applied to the impedance response. In another embodiment, an automated supervisor determines whether an electromagnetic model needs to be run, or whether a correction can be estimated from the output of previous models. For example, if the response of the array interaction with an edge of the material has been modeled at 30 degrees and 60 degrees, the supervisor may determine to estimate the correction at 45 degrees, rather than run a new model.At Step 507, the corrections that were calculated in Step 505 are applied to the in-situ data. For each inspection point near an edge or feature of interest (that is, in a region where the uniform layer model is no longer valid), subtract or otherwise compensate the modeled geometric impedance correction from the measured signal. Then, invert the corrected signal using the unform layer model measurement grids to obtain improved material property estimates. In some embodiments, a correction lookup table or response surface as a function of distance to edge, orientation, and local curvature may be pre-computed and stored.

[0162] At Step 509, corrections that were computed and stored during step 505 are leveraged to train an Al / ML model. The objective of this step is to approximate the EM corrections without requiring execution of the full EM model for each new geometry. By learning the relationship between the impedance corrections and the model setup parameters (e.g., geometric parameters, material properties, sensor configuration), the trained model can predict appropriate adjustments for similar geometries much more rapidly than an EM model. The physics-based model remains the authoritative reference, while the Al / ML model serves to increase efficiency for real-time or large-scale applications.

[0163] FIG. 6 shows a flow chart illustrating a process 600 for correction of material properties at edges and / or complex geometries for eddy current in-situ LPBF data using deconvolution. Process 600 may be implemented using system 100 or in any suitable way. Process 600 is generally described in connection with in-situ data collected during a laser powder bed fusion (LPBF) fabrication process where the test object 130 (FIG. 1) is a component being fabricated. Though, process 600 may be applied to any suitable application.

[0164] Process 600 begins at step 601 with receiving in-situ data that was acquired during an additive manufacturing build. This includes a full set of inspection data for each of the relevant layers of the build process. The relevant layers may be for a completed build for post-fabrication analysis or may include only a subset of layers or the most recent layer for an ongoing build. In some embodiments, the raw data received is the real and imaginary parts of the transimpedance measurements obtained from an eddy current sensor array, organized by frequency and sensor channel number. This data may also contain position information that was captured by a position encoder, machine motor controller, or other suitable position measurement method.

[0165] At Step 603, the material properties of each layer are estimated across the xy -plane in regions sufficiently far from edges and / or complex features where uniform layer modeled responses are highly accurate. This step is optional if the printed component does not contain any such regions, or if it is more practical to apply the deconvolution process in Step 605 acrossthe entire part. In some embodiments, where the raw data is the real and imaginary parts of the transimpedance measurements obtained from the sensor, this data is converted to material property information using a multivariate inverse method in combination with a database of precomputed sensor responses to estimate the material properties.

[0166] At Step 605, deconvolution of the measured signal using known probe step response (or its derivative, the impulse response) is used to correct for edge effects and better estimate the material properties near part boundaries. For an eddy current array, the measured impedance response Z(x,y) can be modeled as the convolution of the true material property distribution p(x,y) and a response function h(x,y) that describes how the sensor array responds at different relative positions to the material. By adequately characterizing the response function h(x,y) in advance, deconvolution may be used to recover a corrected property map p(x,y). In some embodiments, the process includes empirically collecting and characterizing sensor response functions by scanning relevant geometric features. In some embodiments, the geometric features are specifically designed for this empirical data collection. In such embodiments, thick parts may be used to obtain step functions going up and down walls at various angles. The same will be done for thin walls and very thin walls. From the step response, an impulse response may be obtained by taking the numerical derivative, with appropriate smoothing to control noise. For ultra-thin walls / features or narrow slots, the response itself may approximate direct impulse-like functions. In some embodiments, the empirical sensor response captures include cylindrical parts with thick and thin walls of various diameters, and / or responses for specific features for a known build. As more empirical data is gathered, the families of sensor responses are stored in a labeled library. In some embodiments, the empirical data is replaced with or expanded by electromagnetic modeling generation of the sensor response impulse function. In some embodiments, once the sensor impulse response has been empirically characterized, frequencydomain deconvolution is applied to the in-situ data using Fourier transforms, or other such method, to recover an estimate of the underlying property distribution. In some embodiments, a scaling factor may need to be applied after the deconvolution. In some embodiments, principal component analysis is used to consolidate the empirical response families into a generalizable model that fills the space of possible probe behaviors, making the deconvolution more adaptive and reliable across varying scan conditions. In all such embodiments, the deconvolution is applied to each of the sense element’s scan data from the in-situ data using the empirically characterized (or PCA-reduced) impulse response to obtain corrected material properties.In some embodiments, a hybrid method of process 500 and process 600 is utilized with an automated supervisor.

[0167] Note that the impedance correction methods described in process 500 and process 600 are also used to correct lift-off estimates. In one embodiment, the layer height of the printed component is monitored through the corrected lift-off and the power or laser speed or other process feature is adjusted to attain the desired layer height.

[0168] In one embodiment, the corrected conductivities are used to estimate the temperature of the metal and in turn this is used to control the timing and laser patterns to improve outcomes. In one such embodiment, the goal is to avoid or delay sintering the next layer, if the prior layer temperature is too high.

[0169] In some embodiments, after the material properties have been corrected by process 500 and process 600, a correlation table may be used to correlate the electromagnetic material properties (conductivity, permeability, etc.) to other properties such as residual stress, texture, porosity, grain length, and roughness. In some embodiments, this property information may be used to predict the fatigue life of the printed component.

[0170] Attention is now turned to additional aspects of some embodiments. The following discussion describes further implementations, variations, and extensions that may be employed alone or in combination with the systems and methods described above.

[0171] In some embodiments, preferred analysis directions, paths, or pipes are identified for organizing and processing in-situ sensor data. In one such embodiment, an interaction between a sensed energy source and a build feature is used to determine a location or trajectory to be followed. The trajectory of a detected edge or feature may be compared to an intended position of the edge or feature defined by a build file or design model to assess deviations during fabrication.

[0172] In some embodiments, a preferred analysis direction follows spatial positioning along a geometric feature, while in other embodiments the preferred analysis direction follows a progression in time. For example, in bolt hole inspections, a preferred direction may be oriented in a plunge or axial direction to follow the length of the hole or adjacent edges. In other embodiments, such as for wires embedded in risers, a preferred direction follows the wires along their length to assess spacing between wires and to detect defects such as cracks within the wires. In further embodiments, preferred direction analysis is applied to engine slot scanning along or off an edge of the slot.In some embodiments, information from a build file or intended geometry is combined with eddy current sensing data to generate a three-dimensional rendering of a component. The three-dimensional rendering may include spatially registered material property estimates, defect indications, or other sensing results, enabling visualization of both internal and external features relative to the intended geometry.

[0173] In some embodiments, a component-based training approach is used in which responses to complex defects are constructed from training on less complex defect geometries. In some embodiments, machine learning models are trained primarily along preferred paths, including internal and external edges, and filtering or machine learning is performed along pipes or paths oriented in a build direction rather than primarily in an x-y plane.

[0174] In some embodiments, machine learning models are trained to correct estimated material properties such as electrical conductivity, optionally incorporating electromagnetic models as part of the training. In one embodiment, conductivity is estimated using a layered media model, and permeability is subsequently correlated to estimate residual stress. In another embodiment, machine learning is trained to map impedance measurements to conductivity based on air calibration data.

[0175] In some embodiments, to increase the effective data resolution in the z-direction, the eddy current array acquires data at multiple frequencies simultaneously and uses a model-based approach to estimate material properties in multiple layers with 25 micron resolution or smaller. In some embodiments, the excitation frequencies are selected such that the depth of penetration for the highest two frequencies is on the order of or smaller than the grain length in that direction. In some embodiments, sensor data is repopulated to account for data density variations, scan wobble, or non-uniform scan velocity. In some embodiments, remote current sensing is utilized to improve measurement resolution. In some embodiments, once conductivity and lift-off data are determined using grids, machine learning, or both, correlation tables are used to relate electromagnetic properties to porosity, texture, grain length, or surface roughness, which may be used to assess part quality or predict fatigue life.

[0176] In some embodiments, conductivity is initially estimated using a uniform layer model in regions away from geometric features, and a supervisory process determines where electromagnetic models are applied in regions with complex geometry to improve conductivity estimates. In some embodiments, conductivity estimates are improved using machine learning and are further used to estimate material temperature, which in turn is used to control timing, laser power, scan patterns, or other process parameters. In one such embodiment, control is usedto avoid sintering a subsequent layer when a prior layer temperature exceeds a threshold. In another embodiment, layer height is monitored using lift-off or machine-leaming-corrected liftoff, and process parameters are adjusted to achieve desired layer thickness.

[0177] In some embodiments, principal component analysis is applied to families of response functions corresponding to different geometric features to generate reduced-order representations used in deconvolution. In some embodiments, hybrid approaches combine system-theoretic deconvolution with electromagnetic modeling and supervisory logic to determine an appropriate correction method. In some embodiments, empirically derived response functions are replaced or supplemented by response functions generated using electromagnetic modeling.

[0178] In some embodiments, sensor data is used during process development to reduce the time required to establish suitable processing parameters, and subsequently used during production to ensure continued quality. In some embodiments, real-time data is processed using multithreading, parallel processing, or other techniques to enable rapid detection of defects, stopping of defective parts, or adjustment of processing parameters to improve part outcomes.

[0179] In some embodiments, after a defect or feature is identified, a second sensor or inspection modality is used to further characterize the defect, including determining size, length, depth, or tracking the feature over time or space. In some embodiments, after a defect is identified using ultrasonic testing, an energy source is selectively focused or controlled to improve tracking of the defect during fabrication or inspection.

[0180] In some embodiments, material temperature is estimated from conductivity as a function of depth and evaluated at different spatial or temporal offsets relative to an active process, including before and after energy deposition. In some embodiments, sensor data is normalized using reference measurements obtained from a machine frame or other reference structure to reduce temperature-related or system-level effects. In some embodiments, normalization using the machine frame is applied to remove temperature-induced variations in sensor response.

[0181] In some embodiments, prediction algorithms are used to estimate a future state of a component in which an interaction between an energy source and a geometric feature is no longer changing over time or space. In one such embodiment, non-destructive evaluation techniques are applied to a geometric feature whose material condition evolves due to manufacturing or environmental effects to predict a time to a prescribed future state or to predict a material condition at a specified time. In some embodiments, multiple anomalies areintentionally fabricated, subjected to fatigue loading, and inspected to evaluate crack initiation and growth behavior.

[0182] In some embodiments, information from an intended build, such as data derived from a build file, computer-aided design (CAD) model, or planned scan strategy, is combined with in-situ eddy current sensing data to generate a three-dimensional rendering of a component. The rendering may spatially correlate estimated material properties, defect indications, or other sensed parameters with the intended geometry of the component, thereby facilitating visualization of internal and external features and enabling comparison between intended and as-built conditions.

[0183] In some embodiments, excitation frequencies for the eddy current sensor are selected such that an effective sensing resolution in a build direction is finer than a nominal layer thickness of an additive manufacturing process. By appropriate frequency selection, segmented field configuration, or both, sensitivity to material properties at different depths may be enhanced, enabling resolution of property variations within a single build layer or across multiple adjacent layers.

[0184] In some embodiments, after a defect or feature of interest is identified using a first sensing modality, such as ultrasonic testing, an energy source associated with a manufacturing or inspection process is selectively focused or controlled to improve tracking of the defect or feature in space and / or time. This may include adjusting process parameters, scan paths, sensing trajectories, or inspection timing to follow the evolution of the feature during continued fabrication, inspection, or post-processing.

[0185] In some embodiments, material temperature is estimated from electrical conductivity measurements as a function of depth within the component. Conductivity-based temperature estimation may be performed at multiple depths and at different spatial or temporal offsets relative to an active manufacturing process, including before, during, and after energy deposition. Such temperature estimates may be used to assess thermal history, monitor cooling behavior, or inform process control decisions.

[0186] In some embodiments, large, substantially planar eddy current sensor arrays are employed, each having a plurality of sensing elements arranged in two dimensions. One or more such sensor arrays may be positioned before and / or after a processing region and moved relative to the component in a build direction by controlling sensor lift-off. The sensor arrays may also be rotated based on knowledge of the intended build geometry and pattern recognition ofmeasured features to enable tracking of the process both upstream and downstream of an energy interaction region.

[0187] In some embodiments, the planar sensor arrays include a combination of continuously operated sensing elements and selectively multiplexed sensing elements. For example, one or more rows of sensing elements may be operated continuously while other rows are multiplexed on a periodic or event-driven basis, thereby enabling acquisition of high-resolution data without requiring all sensing channels to be operated in parallel at all times. Periodic multiplexing of additional sensing elements may be used to acquire a more complete spatial representation of the sensed region.

[0188] In some embodiments, sensor data is normalized using reference measurements obtained from a frame, structure, or other reference feature of an additive manufacturing machine. Such normalization may be used to reduce the influence of environmental or system-level variations on the measured sensor responses.

[0189] In some embodiments, normalization using reference measurements from the additive manufacturing machine frame is applied to compensate for temperature-related effects in the sensor data. By accounting for temperature-induced changes in the sensing system or surrounding environment, improved stability, repeatability, and accuracy of material property estimates may be achieved.

[0190] The disclosed methods and systems are not limited to any particular additive manufacturing process or equipment configuration. The techniques described herein may be used with a wide range of additive manufacturing processes, including laser powder bed fusion, electron beam melting, directed energy deposition, wire arc additive manufacturing, binder jetting, material extrusion, vat photopolymerization, material jetting, sheet lamination, cold spray deposition, and hybrid additive-subtractive processes, as well as variants and combinations thereof. Further, the disclosed methods and systems may be applied to other manufacturing, fabrication, repair, or inspection processes in which in-situ or process-integrated sensor data is acquired, including subtractive manufacturingjoining, coating, forming, heat treatment, and other industrial processes or inspection applications. Accordingly, the disclosed techniques are broadly applicable to manufacturing and inspection contexts beyond additive manufacturing.

[0191] Having thus described several aspects of at least one embodiment of this invention, it is to be appreciated that various alterations, modifications, and improvements will readily occur to those skilled in the art. Such alterations, modifications, and improvements are intended to be partof this disclosure and are intended to be within the spirit and scope of the invention. Accordingly, the foregoing description and drawings are by way of example only.

[0192] It should be appreciated that all mechanical and end electrical equipment will have functional limitations. Generally, the ideal behavior has been described so as to not unnecessarily distract from the general operation and description of the embodiments. Those of skill in the art will recognize and appreciate the need to consider both ideal and non-ideal behavior in designing specific embodiments just as with any electrical or mechanical device.

[0193] It should also be appreciated that the descriptions of components having the same name or same reference number appear in multiple drawings so as to avoid having to describe the common aspects of a component multiple times. It should be clear to those of skill in the art whether such descriptions made with reference to one embodiment are applicable to another embodiment.

[0194] Various aspects of the present invention may be used alone, in combination, or in a variety of arrangements not specifically discussed in the embodiments described in the foregoing and is therefore not limited in its application to the details and arrangement of components set forth in the foregoing description or illustrated in the drawings. For example, aspects described in one embodiment may be combined in any manner with aspects described in other embodiments. The above-described embodiments of the present invention can be implemented in any of numerous ways. For example, the embodiments may be implemented using hardware, software or a combination thereof. When implemented in software, the software code can be executed on any suitable processor or collection of processors, whether provided in a single computer or distributed among multiple computers.

[0195] Further, it should be appreciated that a computer may be embodied in any of a number of forms, such as a rack-mounted computer, a desktop computer, a laptop computer, or a tablet computer. Additionally, a computer may be embedded in a device not generally regarded as a computer but with suitable processing capabilities, including a Personal Digital Assistant (PDA), a smart phone or any other suitable portable or fixed electronic device.

[0196] Also, a computer may have one or more input and output devices. These devices can be used, among other things, to present a user interface. Examples of output devices that can be used to provide a user interface include printers or display screens for visual presentation of output and speakers or other sound generating devices for audible presentation of output.

[0197] Examples of input devices that can be used for a user interface include keyboards, and pointingdevices, such as mice, touch pads, and digitizing tablets. As another example, a computer may receive input information through speech recognition or in other audible format.

[0198] Such computers may be interconnected by one or more networks in any suitable form, including as a local area network or a wide area network, such as an enterprise network or the Internet. Such networks may be based on any suitable technology and may operate according to any suitable protocol and may include wireless networks, wired networks or fiber optic networks.

[0199] Also, the various methods or processes outlined herein may be coded as software that is executable on one or more processors that employ any one of a variety of operating systems or platforms. Additionally, such software may be written using any of a number of suitable programming languages and / or programming or scripting tools, and also may be compiled as executable machine language code or intermediate code that is executed on a framework or virtual machine.

[0200] In this respect, the invention may be embodied as a computer readable medium (or multiple computer readable media) (e.g., a computer memory, one or more floppy discs, compact discs, optical discs, magnetic tapes, flash memories, circuit configurations in Field Programmable Gate Arrays or other semiconductor devices, or other tangible computer storage medium) encoded with one or more programs that, when executed on one or more computers or other processors, perform methods that implement the various embodiments of the invention discussed above. The computer readable medium or media can be transportable, such that the program or programs stored thereon can be loaded onto one or more different computers or other processors to implement various aspects of the present invention as discussed above.

[0201] In this respect, it should be appreciated that one implementation of the above-described embodiments comprises at least one computer-readable medium encoded with a computer program (e.g., a plurality of instructions), which, when executed on a processor, performs some or all of the above-discussed functions of these embodiments. As used herein, the term “computer-readable medium” encompasses only a computer-readable medium that can be considered to be a machine or a manufacture (i.e., article of manufacture). A computer-readable medium may be, for example, a tangible medium on which computer-readable information may be encoded or stored, a storage medium on which computer-readable information may be encoded or stored, and / or a non-transitory medium on which computer-readable information may be encoded or stored. Other non-exhaustive examples of computer-readable media include a computer memory (e.g., a ROM, a RAM, a flash memory, or other type of computer memory), amagnetic disc or tape, an optical disc, and / or other types of computer-readable media that can be considered to be a machine or a manufacture.

[0202] The terms “program” or “software” are used herein in a generic sense to refer to any type of computer code or set of computer-executable instructions that can be employed to program a computer or other processor to implement various aspects of the present invention as discussed above. Additionally, it should be appreciated that according to one aspect of this embodiment, one or more computer programs that when executed perform methods of the present invention need not reside on a single computer or processor, but may be distributed in a modular fashion amongst a number of different computers or processors to implement various aspects of the present invention.

[0203] Computer-executable instructions may be in many forms, such as program modules, executed by one or more computers or other devices. Generally, program modules include routines, programs, objects, components, data structures, etc. that perform particular tasks or implement particular abstract data types. Typically, the functionality of the program modules may be combined or distributed as desired in various embodiments.

[0204] Also, data structures may be stored in computer-readable media in any suitable form. For simplicity of illustration, data structures may be shown to have fields that are related through location in the data structure. Such relationships may likewise be achieved by assigning storage for the fields with locations in a computer-readable medium that conveys relationship between the fields. However, any suitable mechanism may be used to establish a relationship between information in fields of a data structure, including through the use of pointers, tags or other mechanisms that establish relationship between data elements.

[0205] Also, the invention may be embodied as a method, of which an example has been provided. The acts performed as part of the method may be ordered in any suitable way.

[0206] Accordingly, embodiments may be constructed in which acts are performed in an order different than illustrated, which may include performing some acts simultaneously, even though shown as sequential acts in illustrative embodiments.

[0207] For the purposes of describing and defining the present disclosure, it is noted that terms of degree (e.g., “substantially,” “slightly,” “about,” “comparable,” etc.) may be utilized herein to represent the inherent degree of uncertainty that may be attributed to any quantitative comparison, value, measurement, or other representation. Such terms of degree may also be utilized herein to represent the degree by which a quantitative representation may vary from a stated reference (e.g., about 10% or less) without resulting in a change in the basic function ofthe subject matter at issue. Unless otherwise stated herein, any numerical values appearing in this specification may be modified by a term of degree thereby reflecting their intrinsic uncertainty.

[0208] Use of ordinal terms such as “first,” “second,” “third,” etc., in the claims to modify a claim element does not by itself connote any priority, precedence, or order of one claim element over another or the temporal order in which acts of a method are performed, but are used merely as labels to distinguish one claim element having a certain name from another element having a same name (but for use of the ordinal term) to distinguish the claim elements.

[0209] Also, the phraseology and terminology used herein is for the purpose of description and should not be regarded as limiting. The use of "including," "comprising," or “having,” “containing,” “involving,” and variations thereof herein, is meant to encompass the items listed thereafter and equivalents thereof as well as additional items.

Claims

CLAIMS1. A method for generating a defect detection model for an additive manufacturing process, the method comprising acts of:receiving in-situ sensor data acquired during a first additive manufacturing build of a component;identifying a spatial path that follows a geometric feature of interest of the component; receiving ground-truth defect information for the component;labeling a portion of the in-situ sensor data as corresponding to a ground-truth defect in the ground-truth defect information to form labeled training data;training a machine-learning algorithm as the defect detection model using the labeled training data; andtesting the defect detection model using additional in-situ sensor data.

2. The method of claim 1, wherein receiving the in-situ sensor data comprises receiving multi-channel in-situ sensor data.

3. The method of claim 2, wherein receiving the multi-channel in-situ sensor data comprises receiving eddy current array sensor data acquired during the first additive manufacturing build.

4. The method of claim 3, wherein the eddy current array sensor data comprises real and imaginary components of transimpedance measurements organized by frequency and sensor channel.

5. The method of claim 1, wherein receiving the in-situ sensor data comprises receiving sensor data for a plurality of layers of the first additive manufacturing build.

6. The method of claim 1, wherein the in-situ sensor data further includes position information captured by at least one of a position encoder or a machine motor controller.

7. The method of claim 1, further comprising, prior to labeling, spatially registering the in-situ sensor data to build coordinates.

8. The method of claim 1, wherein spatially registering comprises converting at least a portion of the in-situ sensor data into material property data tied to part geometry.

9. The method of claim 8, wherein converting comprises estimating at least conductivity and liftoff from raw sensor measurements.

10. The method of claim 1, further comprising, prior to labeling, performing at least one of: removal of corrupted data, handling missing data, timestamp synchronization with a machine controller, layer number indexing, scan path indexing, noise reduction, background subtraction, smoothing, or dynamic range normalization.

11. The method of claim 1, further comprising, prior to labeling, repopulating the in-situ sensor data to account for data density differences in a scan direction by at least one of interpolation or a moving window average.

12. The method of claim 1, wherein identifying the spatial path comprises identifying a preferred analysis direction that follows at least one of an edge, a comer, a wall, a hole, a boundary, or another evolving geometric feature of the component.

13. The method of claim 1, wherein identifying the spatial path comprises defining a path through a data volume having a cross-sectional region transverse to the spatial path.

14. The method of claim 13, further comprising combining sensor responses within the cross-sectional region to generate a representative value along the spatial path.

15. The method of claim 1, wherein identifying the spatial path comprises subdividing a region of interest into a plurality of adjacent paths to increase spatial resolution within the region of interest.

16. The method of claim 1, further comprising, prior to labeling, filtering the in-situ sensor data primarily along the spatial path to obtain directionally organized data.

17. The method of claim 16, further comprising extracting characteristic signatures from the directionally organized data, wherein labeling the portion of the in-situ sensor data comprises labeling at least one characteristic signature.

18. The method of claim 17, further comprising storing the characteristic signatures with metadata including at least one of defect type, size, depth, porosity level, layer number, or process parameters.

19. The method of claim 1, wherein receiving the ground truth defect information comprises identifying defects by at least one of: deliberately seeded defects introduced during design or fabrication, computed tomography inspection, or destructive metallographic inspection.

20. The method of claim 19, further comprising mapping identified defect locations to build coordinates for registration with the in-situ sensor data.

21. The method of claim 1, further comprising storing defect metadata including at least one of defect type, size, morphology, severity indicators, or spatial coordinates including layer number and scan vector.

22. The method of claim 1, wherein labeling comprises associating a ground truth defect with one or more corresponding in-situ responses along the spatial path.

23. The method of claim 1, further comprising labeling at least one portion of the in-situ sensor data as corresponding to a non-defect indication to train the machine learning algorithm to suppress false signals.

24. The method of claim 1, wherein validating the defect detection model comprises validating using in-situ sensor data acquired during a second additive manufacturing build different from the first additive manufacturing build.

25. The method of claim 1, further comprising testing the defect detection model using in-situ sensor data from one or more builds not used to train the defect detection model.

26. The method of claim 25, further comprising determining whether a performance requirement is met and, when the performance requirement is not met, diagnosing a failure mode and expanding a training dataset prior to retraining the machine learning algorithm.

27. The method of claim 1, wherein training comprises partitioning labeled training data into training, validation, and test sets.

28. The method of claim 1, wherein training the machine learning algorithm comprises training at least one of an artificial neural network, a convolutional neural network, a recurrent neural network, a temporal convolutional neural network, or a physics-informed model.

29. The method of claim 1, further comprising, after validating, applying the defect detection model to additional in-situ sensor data to output a defect prediction identifying at least one of defect location, defect type, or defect severity.

30. The method of claim 29, wherein applying the defect detection model is performed during the additive manufacturing process for real-time monitoring of at least one layer or scan.

31. The method of claim 29, further comprising, responsive to outputting the defect prediction, using a second sensor or modality to further characterize a defect.

32. The method of claim 29, wherein applying the defect detection model is performed after completion of the build to perform post-production analysis of an in-situ dataset.

33. A system for defect detection in an additive manufacturing process, comprising:at least one processor; anda non-transitory computer-readable storage medium storing instructions that, when executed by the at least one processor, cause the system to:receive in-situ sensor data generated from a layer-by-layer additive manufacturing build of a component;obtain a machine-learning-based defect detection model generated according to the method of claim 1; andapply the machine-learning-based defect detection model to the received in-situ sensor data to identify one or more defect locations in the component.

34. The system of claim 33, further comprising:additive manufacturing processing hardware; andin-situ sensing hardware configured to measure the in-situ sensor data during the build of the component,wherein the at least one processor applies the machine-learning-based defect detection model to the in-situ sensor data in real time during the additive manufacturing build to identify defect locations as the component is being fabricated.

35. The system of claim 33, wherein the in-situ sensor data comprises multi-channel sensor data acquired during the layer-by-layer build.

36. The system of claim 33, wherein the in-situ sensor data comprises eddy current array sensor data including real and imaginary components of transimpedance measurements organized by frequency and sensor channel.

37. The system of claim 33, wherein the instructions further cause the system to spatially register the received in-situ sensor data to build coordinates prior to applying the machine learning based defect detection model.

38. The system of claim 33, wherein the instructions further cause the system to identify one or more preferred analysis directions that follow at least one geometric feature of interest and to filter the received in situ sensor data along the one or more preferred analysis directions to generate directionally organized data that is provided to the machine learning based defect detection model.

39. The system of claim 38, wherein filtering along the one or more preferred analysis directions comprises associating sensor responses along a path having a defined cross-sectional region transverse to the path and combining sensor responses within the cross-sectional region to generate a representative value along the path.

40. The system of claim 33, wherein the instructions further cause the system, responsive to identifying a defect location, to determine whether the defect location corresponds to a catastrophic defect and, when the defect location corresponds to a catastrophic defect, to stop the additive manufacturing build.

41. The system of claim 34, wherein the instructions further cause the system, responsive to identifying a defect location, to revise a build plan for at least one subsequent layer or scan to alter processing parameters to reduce or repair the defect as the component is being fabricated.

42. The system of claim 33, wherein the instructions further cause the system to, responsive to identifying a defect location, cause a second sensor or modality to be used to further characterize the defect.

43. A system for estimating a material property of a component, comprising:an eddy current sensor having at least one drive conductor and a plurality of sense elements;an impedance analyzer to excite the at least one drive conductor and to acquire raw sensor response data comprising complex transimpedance measurements from the plurality of sense elements;a non-transitory computer-readable storage medium storing computer-executable instructions; andat least one processor operatively coupled to the storage medium;wherein the computer-executable instructions, when executed by the at least one processor, cause the system to:receive the raw sensor response data including the complex transimpedance measurements acquired as the eddy current sensor is positioned relative to the component;access electromagnetic model data characterizing an interaction between the eddy current sensor and a geometric feature of the component;determine, from the electromagnetic model data, a geometric correction for the complex transimpedance measurements associated with the geometric feature;apply the geometric correction to the raw sensor response data to generate corrected transimpedance data; andestimate, from the corrected transimpedance data, the material property of the component.

44. The system of claim 43, wherein the material property is electrical conductivity.

45. The system of claim 43, wherein the electromagnetic model data comprises precomputed modeled complex transimpedance responses for the geometric feature across a plurality of sensor lift-off values and a plurality of material conductivity values.

46. The system of claim 43, wherein determining the geometric correction comprises selecting, interpolating, or both selecting and interpolating a correction from a lookup table or response surface indexed by at least one of distance-to-edge, edge orientation, local curvature, sensor liftoff, excitation frequency, or sense element identifier.

47. The system of claim 43, wherein applying the geometric correction comprises subtracting a modeled geometric transimpedance contribution from the complex transimpedance measurements to produce the corrected transimpedance data.

48. The system of claim 43, further comprising a grid database generated using a uniform-layer sensor-material model, wherein estimating the material property comprises performing an inverse interpolation using the corrected transimpedance data and the grid database to estimate at least electrical conductivity.

49. The system of claim 43, wherein the computer-executable instructions further cause the system to estimate sensor lift-off from the corrected transimpedance data, and to use the estimated sensor lift-off in estimating the material property.

50. The system of claim 43, wherein the computer-executable instructions further cause the system to determine that the eddy current sensor is within a threshold distance of the geometric feature and, responsive to determining that the eddy current sensor is within the threshold distance, to apply the geometric correction, and otherwise to estimate the material property without applying the geometric correction.

51. The system of claim 43, wherein the electromagnetic model data characterizes the interaction between the eddy current sensor and the geometric feature for a plurality of build layers of an additively manufactured component.

52. The system of claim 43, wherein the electromagnetic model data is generated using at least one of a two-dimensional finite element model or a three-dimensional finite element model of the eddy current sensor and the component geometry.

53. The system of claim 43, wherein the computer-executable instructions further cause the system to generate the electromagnetic model data by modeling the interaction both with and without the geometric feature and to determine the geometric correction based on a difference between the modeled complex transimpedance responses.

54. The system of claim 43, wherein the computer-executable instructions further cause the system to use the electromagnetic model data to train a machine-learning model to predict the geometric correction from inputs comprising at least one of geometric parameters of the component, sensor configuration parameters, excitation frequency, sensor lift-off, or estimated material properties, and to apply the machine-learning model to generate the geometric correction without executing an electromagnetic model for a new geometry.

55. A system for estimating a material property of a component, comprising:an eddy current sensor having at least one drive conductor and a plurality of sense elements;an impedance analyzer to excite the at least one drive conductor and to acquire from the plurality of sense elements raw sensor response data comprising complex transimpedance measurements;a non-transitory computer-readable storage medium storing computer-executable instructions; andat least one processor operatively coupled to the storage medium;wherein the computer-executable instructions, when executed by the at least one processor, cause the system to:receive the raw sensor response data including the complex transimpedance measurements acquired as the eddy current sensor is positioned relative to the component;access a sensor response function characterizing a spatial response of the eddy current sensor to a geometric feature of a conductive material;apply a deconvolution process to the raw sensor response data using the sensor response function to compensate for geometric distortion of the complex transimpedance measurements; andestimate, from the deconvolved sensor response data, the material property of the component.

56. The system of claim 55, wherein the material property is electrical conductivity.

57. The system of claim 55, wherein the sensor response function comprises an empirically characterized response obtained by scanning the eddy current sensor relative to one or more reference geometric features of known geometry.

58. The system of claim 55, wherein the sensor response function comprises a modeled response generated using at least one of a two-dimensional finite element model or a three-dimensional finite element model of the eddy current sensor and the geometric feature.

59. The system of claim 55, wherein applying the deconvolution process comprises performing a frequency-domain deconvolution using a Fourier transform representation of the sensor response function.

60. The system of claim 55, wherein the sensor response function is derived from a step response of the eddy current sensor relative to the geometric feature, and wherein an impulse response is obtained from the step response for use in the deconvolution process.

61. The system of claim 55, wherein the sensor response function is selected from a library of response functions corresponding to different geometric features, orientations, or sensor lift-off conditions.

62. The system of claim 55, wherein the computer-executable instructions further cause the system to estimate sensor lift-off from the deconvolved sensor response data and to use the estimated sensor lift-off in estimating the material property.

63. The system of claim 55, wherein the computer-executable instructions further cause the system to apply principal component analysis to a plurality of sensor response functions to generate a reduced-order representation used in the deconvolution process.

64. The system of claim 55, wherein the deconvolution process is applied independently to complex transimpedance measurements acquired at a plurality of excitation frequencies.

65. The system of claim 55, wherein the computer-executable instructions further cause the system to apply the deconvolution process selectively when the eddy current sensor is within a threshold distance of the geometric feature.

Images