Method for defect detection in additively manufactured components using in situ acoustic emission analysis

Abstract

The invention relates to the fields of materials science and mechanical engineering and concerns a method for defect detection in additively manufactured components using in-situ acoustic emission analysis, employing machine learning. The object of the invention is to provide a method for defect detection in additively manufactured components using acoustic emission analysis, enabling improved, reliable, fast, and reproducible defect detection of additively manufactured components using acoustic emission analysis.The problem is solved using a method for defect detection in additively manufactured components via in-situ acoustic emission analysis. This involves acquiring and transforming acoustic emission data, then using a machine learning model to identify and localize characteristic signal patterns from the acoustic emission data. Subsequently, these characteristic and localized signal patterns are extracted from the original acoustic emission data and fed into a machine learning model. Finally, the extracted acoustic emission data are classified using a machine classification model. The proposed method can be used, for example, in laser-based additive manufacturing of components.

Description

[0001] The invention relates to the fields of materials science and mechanical engineering and concerns a method for defect detection in additively manufactured components using in situ acoustic emission analysis, which utilizes machine learning. The proposed method can be used, for example, in laser-based additive manufacturing of components.

[0002] Additive manufacturing processes are becoming increasingly important and the components produced with them have found their way into a wide variety of industries.

[0003] Well-known additive manufacturing processes include selective laser sintering, laser beam melting and electron beam melting.

[0004] In the additive manufacturing of a three-dimensional component, powdered materials are used, which are created as layers using a laser or electron beam and a coating device. The production of the three-dimensional component with this type of additive manufacturing process is achieved by creating several layers arranged one above the other on a type of platform, whereby a targeted energy input into the respective powdered layer material is required by the laser or electron beam.

[0005] Due to various influencing factors, such as insufficiently adapted, material-specific process parameters, different and inadequate qualities of the powder material or the use of difficult-to-process materials, various defects in the form of cracks, pores or insufficient material connection of the layers arranged on top of one another can occur during the additive manufacturing of the three-dimensional components, which are difficult to detect or determine during the manufacturing process.

[0006] This places particularly high demands on the mechanical properties and reliability of safety-relevant components, such as those used in aviation, mechanical and plant engineering or medical technology.

[0007] In-situ monitoring during additive manufacturing of a component is already known from the state of the art, but its application is limited in terms of location and time. The known methods pursue the approach of inspecting the additively manufactured component for potential defects after production.

[0008] Camera-based monitoring methods are known from SK Everton et al., Mater. Des. 95 (2016) 431-445, and exclusively observe the upper layer of the melt pool created during the manufacturing process. Thus, the method presented here can only detect irregular melting processes in the upper layer of the melt.

[0009] Acoustic methods for component monitoring during the additive manufacturing of three-dimensional components are also known.

[0010] For example, WO 2008 / 084 045 A1 discloses a method for monitoring undesired mechanical stresses in a polymer, wherein the acoustic emission generated by the polymer formation is detected by one or more acoustic sensors attached to a device.

[0011] DE 10 2017 124 100 A1 discloses a method and devices for the additive manufacturing of workpieces. For analysis, a test is performed during production using a testing method, for example, laser ultrasound. For analysis, the test result is compared with a test simulation result.

[0012] WO 98 / 31490 A2 discloses a method for determining deformations and the stresses acting on them using acoustic emission analysis. For this purpose, a ductile metallic material, preferably a weldable structural steel, is admixed with a particle-like material that is brittle at least in the temperature range below 400°C in order to increase the sound radiation behavior.

[0013] US 9 989 495 B2 discloses a method for monitoring an additive manufacturing process using a directed energy source. The method comprises measuring acoustic energy generated in or by the melt pool during processing to generate a measurement acoustic profile. Comparing the measurement acoustic profile with a baseline acoustic profile representative of a reference workpiece allows the determination of a difference between the measurement acoustic profile and the baseline acoustic profile, and thus the presence of a defect in the workpiece. Furthermore, the method comprises controlling at least one aspect of the additive manufacturing process in response to the measurement acoustic profile by taking a discrete action in response to the acoustic signal profile deviating from the baseline acoustic profile, wherein the discrete action is stopping the additive manufacturing process.

[0014] US 2021 / 0 039 324 A1 discloses the use of a sensor readout system with at least one fiber optic sensor connected to a processing unit as part of an additive manufacturing facility via at least one signal line. For in-situ and real-time quality control of an ongoing additive manufacturing process, the acoustic emissions are measured via the at least one fiber optic sensor, the signals are subsequently transmitted, and the measured signals are analyzed in the processing unit.

[0015] A system for in-situ detection of defects during an additive manufacturing process is also known from US 2021 / 0 365 016 A1. It comprises a printing substrate configured to support a structure during an additive manufacturing process, a material dispensing means for applying one or more material layers to the printing substrate to form a structure, a plurality of sensors configured to detect sound waves generated during the additive manufacturing process, and one or more processors communicatively connected to the plurality of sensors. These determine whether a defect is present in a structure generated by the additive manufacturing process based on the analysis of the sound waves. The method involves executing one or more control commands configured to modify the additive manufacturing process in response to the detection of a defect present in the structure.

[0016] A disadvantage of the known methods for defect detection in additively manufactured components using acoustic emission analysis is that, due to the very large amount of data, only a few defect events can be recorded and identified. The large amount of data collected during the manufacturing process primarily comprises non-event information, i.e., information about the defect-free additive manufacturing of the component.

[0017] Another disadvantage is that the defect-induced acoustic signals cannot be sufficiently distinguished from the background noise, so that reliable detection of anomalies or defects in the additively manufactured component cannot be achieved.

[0018] Another disadvantage exists in the application of autoencoder architectures, which are typically considered for anomaly detection. Autoencoder architectures rely on learning a representation of data associated with a defect-free part. This data is used for reconstruction and anomaly detection. Due to the very large amount of data collected during the additive manufacturing process using acoustic emission analysis, an autoencoder struggles to effectively learn the normal state, making it ineffective for detecting the rare but important defect signals.Furthermore, training an autoencoder in such an environment would require processing the vast amounts of data in which the anomalies—the signals indicating defect events—are not only rare but may also be obscured and overlaid by the presence of other data.

[0019] The object of the present invention is to provide a method for defect detection of additively manufactured components by means of acoustic emission analysis, which enables improved, reliable, fast and reproducible defect detection of additively manufactured components by means of acoustic emission analysis.

[0020] The object is achieved by the invention defined in the claims. Advantageous embodiments are the subject of the dependent claims, whereby the invention also includes combinations of the individual dependent claims in the sense of an AND connection, as long as they are not mutually exclusive.

[0021] To achieve the object according to the invention, a method for defect detection of additively manufactured components by means of in-situ acoustic emission analysis is provided, in which acoustic emission data are acquired and then transformed, characteristic signal patterns are subsequently determined and localized from the acoustic emission data using a machine learning model, the characteristic and localized signal patterns are subsequently extracted from the original acoustic emission data and fed to a machine learning model, and the extracted acoustic emission data are subsequently classified using a machine classification model.

[0022] Advantageously, the transformation of the recorded emission data Et is carried out using the formula Et=∑k=1Tf(xt−k+1), where x t the amplitude at time t, f is a function with the property f(x)=f(-x) for any x, and T is the period.

[0023] Also advantageously, the amplitudes of the acquired acoustic emission data are transformed by means of weighted summation, whereby it is particularly advantageous for the transformation to use the period T as an adaptive duration parameter and / or to include only values ​​with positive amplitude.

[0024] Furthermore, it is advantageous if the transformation of the acoustic emission data is performed by weighting the relative amplitudes of the characteristic signal patterns and non-defect event signals and subsequently by generating permutations of combined events using authentic noise profiles.

[0025] In an advantageous embodiment of the method, the acoustic emission data are filtered frequency-based before the transformation.

[0026] In a further advantageous embodiment, the transformed acoustic emission data are processed by downsampling.

[0027] It is also advantageous if the extraction of the transformed acoustic emission data is carried out using artificial neural networks, random forest methods, decision trees, Gaussian processes, k-means clustering or combinations thereof.

[0028] In an advantageous embodiment of the method, the extracted acoustic emission data are classified by generating hybrid training examples through linear combinations of characteristic signal patterns with non-defect event signals, wherein particularly advantageously a relative amplitude weighting is used by combining the characteristic signal patterns with non-defect event signals and subsequently several permutations of the combined signals by using authentic noise profiles.

[0029] Furthermore, it is advantageous if the classification is carried out using a temporal event shifting technique, which involves the systematic shifting of signal segments of the characteristic signal patterns.

[0030] It is also advantageous to use a machine learning model for classification that uses an autoencoder architecture and / or data augmentation process.

[0031] The present invention makes it possible for the first time to provide a method for defect detection of additively manufactured components by means of in situ acoustic emission analysis, which enables improved, reliable, fast and reproducible defect detection of additively manufactured components by means of in situ acoustic emission analysis.

[0032] The task is solved with a new method for defect detection of additively manufactured components using acoustic emission analysis with the application of machine learning. In this method, defects with crack formation that occur in situ during additive manufacturing of the components are recorded in a practical and effective real-time monitoring and quality control, thus influencing a cost-effective and time-saving manufacturing process.

[0033] The proposed method provides several technical advantages.

[0034] First, the formation of cracks is characterized by characteristic signal patterns in the acoustic emissions that occur during the additive manufacturing or cooling process. Within the scope of the invention, characteristic signal patterns are understood to be emission data recorded and graphically evaluated during acoustic emission analysis that characterize the formation of cracks in additively manufactured components. These characteristic signal patterns become more pronounced within a time frame of 5 to 20 milliseconds and exhibit a characteristic wedge-shaped form, with a sharp increase in amplitude followed by a gradual decrease in amplitude.

[0035] However, on larger timescales when recording acoustic emissions, the characteristic signal patterns are difficult to detect because the ambient noise in the manufacturing environment masks the characteristic crack signal patterns. The main ambient noise is generated by a variety of standard operating processes, including the resonances of the laser system, the gas pump, the recoater, and its movements, the electronic switches, and the gas flow dynamics.

[0036] According to the invention, it is proposed to transform the acoustic emission data first recorded in situ, whereby the characteristic signal patterns of crack formation are clearly identified and can be better recognized at larger time scales of data acquisition.

[0037] The transformation of the acquired signal data can advantageously be described by the following formula: Et=∑k=1Tf(xt−k+1), where xt is the amplitude of the waveform at time t, and f is an arbitrary function with the property f(x)=f(-x) for any x. The period T depends on the discretization frequency and is determined by the characteristic time of the crack duration.

[0038] The acoustic emission waveform undergoes a special transformation to enhance the characteristic signal patterns of crack formation. This accentuates and enhances the visibility of the characteristic signal patterns resulting from potential crack formation in the component, making them stand out more clearly against the uniform background noise. The transformation of the acquired emission data reduces the influence of background noise and highlights and enhances the representation of the characteristic signal patterns.

[0039] In addition, the transformation of the acquired acoustic emission data not only improves the visibility of defect-related events but also prepares the acquired acoustic emission data for efficient machine learning processes, thereby increasing the effectiveness of the methodology in detecting crack formation already during the additive manufacturing of the components.

[0040] The transformation of the acoustic emission data can be performed under different conditions. For example, it may be advantageous to normalize the acoustic emission data before entering the training phase of machine learning. Such normalization must be consistent and is performed repeatedly before the prediction phase to ensure the homogeneity of the data and the accuracy of the machine learning model.

[0041] To transform the acquired acoustic emission data, an adaptive duration parameter of period T can advantageously be used. In this case, the period T is adjusted based on the real-time variance or the frequency components of the acoustic emissions. If the system detects higher-frequency noise, the period T can be reduced to better capture these fluctuations. Conversely, the period T can be increased for slower vibrations.

[0042] In a further advantageous embodiment of the transformation, a weighted summation can be performed over the amplitudes of the acquired acoustic emission data by multiplying the function f(x) by constant parameters. In this way, greater importance is attached to the current amplitudes, allowing the system to react particularly quickly to a sudden signal indicating a defect due to crack formation.

[0043] It can also be advantageous to first process the emission data with a frequency-based filter before transformation. Using a bandpass filter can isolate specific frequency ranges that, for example, are most likely to indicate a defect with crack formation. This step increases the method's sensitivity to defects while reducing or even completely suppressing the influence of extraneous noise.

[0044] After transforming the acoustic emission data, it can advantageously be additionally provided that the transformed acoustic emission data is processed using the process of downsampling. Downsampling further reduces the resolution or data rate by discarding certain data points of the signal pattern.

[0045] The main advantages of downsampling from a machine learning perspective are that, by reducing the amount of data, processing times for algorithms are further reduced, leading to faster prediction cycles. Another advantage is that downsampling eliminates redundant or noisy data, thus optimizing the quality of the dataset, especially for machine learning models. The reduced dataset allows for efficient memory utilization, which is crucial for real-time applications or scenarios with limited computing resources.

[0046] After the transformation process and possible downsampling, the acoustic emission data is processed by a machine learning model, which detects and isolates the characteristic signal pattern that characterizes a crack in the additively manufactured component. The machine learning model is responsible for determining the temporal coordinates of these events. These coordinates enable the extraction of potential crack events from the original waveforms and the acoustic emission data. The extraction is then applied to the previously transformed data, using a machine learning model that specifically detects the presence of crack-related sound events within the extracted segments of the recorded acoustic emissions.

[0047] For effective training of the extracting machine learning model based on a robust classification algorithm, the invention provides for the compilation of a synthetic dataset. This dataset is divided into two classes: a first class contains a large number of characteristic and transformed signal patterns, and the second class contains no events of characteristic signal patterns.

[0048] To expand the richness of the dataset with a sufficient number of events, the technique of data augmentation can be advantageously employed. This involves excising segments of the transformed acoustic emission data that inherently contain characteristic signal patterns and then strategically embedding them into segments that are originally event-free. This method augments the dataset with a considerable variety of event-containing characteristic emission data, resulting in a comprehensive training system. Since most of the printing and subsequent cooling process in additive manufacturing of components occurs without characteristic signal patterns and thus without crack events, compiling a dataset with event-free signal patterns is straightforward.However, the method according to the invention stipulates that a data set to be considered complete covers a range of different operating conditions and regimes during the additive manufacturing process. Specifically, this means that the acoustic data set includes all background noises, such as the switching on and off of the gas pump, the movement of the coater, and the switching on and off of the laser beam, as well as crack noises.

[0049] After the process step of extracting characteristic acoustic emission data, the next step is to categorize these data into definitive classifications: defects caused by crack formation and noise. Noise refers to all acoustic signals that do not indicate crack formation in the component.

[0050] It has been found that acoustic signals from suddenly occurring cracks typically exhibit a unique wedge-shaped signal pattern. This characteristic signal pattern is characterized by an initial sharp increase in amplitude followed by a gradual decrease. Conversely, noise bursts can take on a variety of appearances, including sharp, transient peaks or longer sawtooth formations.

[0051] The method according to the invention uses classification algorithms based on machine learning methods. Advantageous embodiments of the algorithms implemented for the event extraction method can be artificial neural networks, random forest methods, decision trees, Gaussian processes, k-means clustering, or combinations thereof.

[0052] It is also conceivable to use an autoencoder architecture, which is known as an effective method for detecting defect events. Using an autoencoder architecture, the machine learning model is trained on data sets without crack events to establish a baseline of normality, which allows deviations from this baseline to be identified as anomalies or characteristic signal patterns within the meaning of the invention.

[0053] In an advantageous embodiment of the invention, it is proposed to use further data augmentation techniques to improve the classification of acoustic events and to distinguish between those resulting from crack formation and noise outbreaks in the additive manufacturing environment.

[0054] A possible data augmentation technique may advantageously involve the synthesis of hybrid event signals by linearly combining characteristic defect signal patterns with non-defect event signals, where the non-defect signals contain the real noise pulses occurring in the manufacturing process. This generates a diverse set of training examples that better reflects the actual conditions under which characteristic signal patterns can be masked by background noise.

[0055] The synthesis of hybrid event signals can be performed by adjusting relative amplitude weights between defect and non-defect event signals in the combinations to simulate different degrees of noise and ensure that a realistic range of scenarios is represented in the training data.

[0056] Furthermore, the synthesis of the hybrid event signal data may involve generating a variety of permutations of combined events by using authentic noise profiles, thereby significantly enriching the training dataset and assisting the machine learning model in accurately identifying crack events even when they match authentic noise signatures.

[0057] The proposed hybrid event synthesis method leverages the inherent variability of real acoustic emissions during the additive manufacturing process, harnessing this natural complexity to train more robust machine learning models. Using actual noise pulses instead of synthetic noise ensures that the augmented data retains the unpredictability and characteristics of the operating environment, increasing the model's predictive reliability.

[0058] Another advantageous data augmentation technique can involve temporal shifting of event data. This exploits the temporal shift of crack events within their respective windows. Since the event extraction process can variably position crack-typical signals within a time window, this temporal shifting of signal events aims to apply systematic temporal shifts to crack events in the training data, ensuring that the classification model is not constrained to a fixed temporal position of the event signal. Furthermore, the robustness of the model is improved by training it to recognize characteristic signal patterns regardless of their position within the acquired event window, accounting for the variability introduced by automatic event extraction.

[0059] In summary, the mentioned data augmentation techniques represent a comprehensive strategy that significantly improves the ability of the machine learning model to accurately classify events in the challenging environment of additive manufacturing, where signal ambiguity due to overlapping acoustic sources is prevalent.

[0060] The method according to the invention is explained in more detail below using an exemplary embodiment. The accompanying figures show Fig. Flowchart of the machine learning model for detecting cracks in the component during its powder bed-based additive manufacturing using a classification method, Fig. Typical waveforms of acoustic emissions indicating crack-related events, Fig. Segment of a recorded acoustic data stream as a function of time, Fig. Transformed version of the same segment to highlight the signals indicating a crack event, transformed and data-reduced signal patterns, Fig. Downsampled version of the transformed signal, Fig. Acoustic events from the Fig. shown waveform, (a) (c) correspond to noise pulses, while waveform (b) is caused by a crack event, and Fig. Schematic representation of the event extraction process. Example

[0061] During additive manufacturing, the quality of a component is tested in situ using acoustic emission analysis. The acoustic emissions generated in the immediate vicinity of the production process and by the component itself are recorded as signal data using an acoustic sensor that is partially connected to the underside of the build platform. Fig. shows such a time-dependent acoustic data stream, which represents all the noises during a manufacturing process. The transformation of the recorded signal data is described by the following formula: Et=∑k=ITf(xt−k+1), where x t is the amplitude of the waveform at time t, and f is any function with the property f(x)=f(-x) for any x. The period T depends on the discretization frequency, which is determined by the characteristic time of the crack duration. At a discretization frequency of 2 MHz, the period T is set equal to 1500 time steps, which corresponds to approximately 1 millisecond.

[0062] The goal here is to more clearly emphasize the differences between the regular background noise and the additional energy contributions during crack formation. The transformation of the acquired acoustic emission data is performed using a summation technique, which is calculated as a moving average over amplitude values ​​or their functionally transformed counterparts according to the formula in (1). Fig. shows a graphical representation of a Fig. shown waveform, which is subject to the transformation process using the given formula with f(x)=x 2 The transformation reveals three prominent peaks within the waveform, indicated by the arrows. These peaks, which stand out from a more uniform background noise level, are defined as characteristic signal patterns within this method. The original waveform, which is Fig. does not include these events visibly within their untransformed acoustic emission data. A closer look at the section corresponding to the middle event in Fig. corresponds and in Fig. confirms that this is a characteristic event signal triggered by a crack. The remaining events are consistent with noise signals, as shown in the Fig. The discretization frequency of the original waveform was set to 2 MHz. At this frequency, the data within a window of 275 milliseconds comprise 750,000 points. The transformed acoustic emission data in Fig. clearly show that the data volume was significantly reduced while maintaining the integrity of the characteristic signal patterns. Downsampling was used to retain every 1500th data point and remove the remaining data points.

[0063] Fig. illustrates the transformed acoustic emission data, which now consists of only 500 data points. This significantly reduces the data volume compared to the original data set of 750,000 data points, while retaining the visibility of the relevant events. As shown in Fig. As shown, the peak of the cracking event has a short duration of about 2 milliseconds, corresponding to about 4,000 data points. However, further reducing the original waveform to a lower resolution would reduce this segment to about 3 data points, making the cracking event undetectable.

[0064] The process step for extracting the characteristic signal patterns is described in Fig. Illustrated step by step. This methodology is based on a sliding window mechanism with a typical time span of 400 to 700 milliseconds. A machine learning model with an autoencoder architecture is used for identification, which was trained, validated, and tested with a previous dataset. To train the model, the original dataset was extended as follows to obtain a so-called hybrid dataset: First, characteristic crack signal patterns and non-defect event signals are isolated, followed by a time-shifted combination of the two signal types with additional weighted relative amplitudes. This generation of permutations of combined events was performed using authentic noise profiles.The machine learning model developed in this way is now able to record data sets of acoustic emissions and to examine them for the occurrence of cracks by means of classification during the production of components using powder bed-based additive manufacturing methods.

[0065] The machine learning model examines each temporal segment that falls within the sliding window according to Fig. lies, for characteristic signal patterns that indicate defect events. If the machine learning model does not identify the presence of an event, a temporal shift of the sliding window is initiated, which can be either a transition over the entire length or an incremental shift to maintain data continuity. If the machine learning model determines the presence of a defect event with a characteristic signal pattern, the event extraction procedure involves determining the time step corresponding to the highest amplitude within the transformed waveform segment. This enables precise localization of the event within the original acoustic emission stream, which is crucial for subsequent process steps. As described in Fig.As shown, if the machine learning model has identified a defect event, the segment associated with the defect event is extracted and replaced with a median value representative of the entire window segment. The modified window is then re-evaluated by the machine learning model to determine the presence of additional events. This process repeats, removing detected events and replacing the median value, until the machine learning model confirms that no further events are present within the window.

Claims

[1] Method for defect detection of additively manufactured components using in-situ acoustic emission analysis, in which acoustic emission data are recorded and then transformed, characteristic signal patterns are then determined and localized from the acoustic emission data using a machine learning model, the characteristic and localized signal patterns are then extracted from the original acoustic emission data and fed to a machine learning model, and the extracted acoustic emission data are then classified using a machine classification model. [2] Method according to claim 1, wherein the transformation of the detected emission data Et is carried out using the formula Et=∑k=ITf(xt−k+1) takes place, whereby x t the amplitude at time t, f is a function with the property f(x)=f(-x) for any x, and T is a period. [3] A method according to claim 1, wherein amplitudes of the acquired acoustic emission data are transformed by means of weighted summation. [4] Method according to claim 3, in which a period T is used as an adaptive duration parameter for the transformation and / or only values with positive amplitude are included. [5] A method according to claim 1, wherein the transformation of the acoustic emission data is performed by weighting the relative amplitudes of the characteristic signal patterns and non-defect event signals and subsequently generating permutations of combined events using authentic noise profiles. [6] A method according to claim 1, wherein the acoustic emission data are frequency-based filtered before the transformation. [7] A method according to claim 1, wherein the transformed acoustic emission data are processed by downsampling. [8] A method according to claim 1, wherein the extraction of the transformed acoustic emission data is carried out using artificial neural networks, random forest methods, decision trees, Gaussian processes, k-means clustering or combinations thereof. [9] The method of claim 1, wherein the extracted acoustic emission data are classified by generating hybrid training examples by linear combinations of characteristic signal patterns with non-defect event signals. [10] A method according to claim 9, wherein relative amplitude weighting is applied by combining the characteristic signal patterns with non-defect event signals and subsequently several permutations of the combined signals are applied by using authentic noise profiles. [11] A method according to claim 1, wherein the classification is performed using a temporal event shifting technique which involves systematically shifting signal segments of the characteristic signal patterns. [12] The method of claim 1, wherein a machine learning model is used for classification, wherein an autoencoder architecture and / or data augmentation process is used.

Images