Additive manufacturing acoustic signal feature processing method based on Markov transfer field

Abstract

The invention discloses an additive manufacturing acoustic signal characteristic processing method based on a Markov transfer field, and the method comprises the steps: building an acoustic signal data set containing the quality characteristics of a molten pool under different technological parameters through collecting acoustic signals generated in an electric arc additive manufacturing forming process; preprocessing the acoustic signal acquired by the acoustic signal detector; selecting a proper sound signal length, generating a two-dimensional state transition probability matrix by the signal through a Markov transition field, and mapping the two-dimensional state transition probability matrix into a two-dimensional image for representation; labeling the images and dividing the images into a training set, a verification set and a test set; and after being enhanced, the different data set images are respectively used for convolutional neural network feature extraction, related hyper-parameter adjustment and final model classification effect test, and finally arc additive manufacturing defect acoustic signal feature extraction and classification are realized. According to the method, the Markov transfer field is adopted to carry out state modeling on the specific time sequence change and the dynamic mode of the sound signal, so that the convolutional neural network can learn the specific characteristics of the sound signal, and finally defect identification is realized.

Description

Technical Field

[0001] The present invention belongs to the technical field of metal additive manufacturing, and in particular to a method for processing acoustic signal characteristics of additive manufacturing based on a Markov transfer field. Background Art

[0002] Compared with traditional casting, forging and subtractive manufacturing processes, Directed Energy Deposition with Arc (DED-Arc) has the advantages of low equipment cost, high material utilization and suitability for large-scale structure manufacturing. It has shown broad application prospects in aerospace, shipbuilding, energy equipment and other fields. During the AAM manufacturing process, the dynamic changes of welding heat source, material deposition and interlayer bonding directly affect the forming quality of parts. However, the manufacturing process involves complex thermal-mechanical coupling effects and unstable physical phenomena, such as fluctuations in the molten metal pool, welding spatter and material cracks, which can easily lead to internal defects such as surface pores, weld discontinuities and cracks. These defects will significantly reduce the mechanical properties and service safety of parts, so real-time detection and accurate identification of defects are crucial during the manufacturing process.

[0003] The existing arc additive manufacturing defect detection method is mainly based on molten pool image monitoring. However, the molten pool has poor environmental adaptability, is limited by viewing angles, and lacks comprehensiveness, which can easily lead to the omission of local defects. In contrast, acoustic signals are comprehensive, real-time, and non-contact, providing an accurate, real-time, and reliable means for defect monitoring. At the same time, recent studies have shown that acoustic signals, as an important process feature in the arc additive manufacturing process, can reflect changes in process parameters and the formation mechanism of defects, and are an effective way to identify defects and monitor quality.

[0004] At present, most of the analysis methods of acoustic signals are based on traditional time domain, frequency domain and time-frequency domain signal processing technologies, but these methods have limited feature extraction capabilities and insufficient robustness when dealing with complex nonlinear and random acoustic signals in the arc additive manufacturing process. Therefore, in the face of such nonlinear, random and dynamically changing acoustic signals.

[0005] The technical differences between this application and the prior art are as follows:

[0006] This patent proposes a method for arc additive manufacturing acoustic signal feature processing and defect recognition based on Markov transfer field. Markov transfer field is a spatiotemporal feature extraction method based on probabilistic graphical model, which can effectively capture the transfer law and spatial distribution characteristics of the signal in time series. Applying Markov transfer field to the acoustic signal feature processing of arc additive manufacturing can extract multi-dimensional information from the signal, providing more accurate feature expression and discrimination basis for defect recognition. This method is expected to significantly improve the real-time and accuracy of defect detection, thereby providing technical support for quality control of arc additive manufacturing process.

[0007] The technical differences between this application and the prior art are as follows:

[0008] Technical comparison with patent CN202410886291.2 "A method and system for arc additive manufacturing defect identification based on acoustic signal time-frequency diagram"

[0009] Patent CN202410886291.2 identifies defects through time domain diagrams, focusing on the time domain relationship in the acoustic signal, using wavelet transform. It reflects the waveform characteristics of the acoustic signal changing over time, that is, the change of amplitude over time. The time domain diagram directly displays the waveform of the acoustic signal, and can clearly observe the basic characteristics of the signal such as amplitude, period, and frequency.

[0010] The present invention analyzes the characteristics and dynamic changes in the acoustic signal in the spatial domain, using the Markov transition field, and encapsulates the transition probabilities of different states in the acoustic signal, that is, the probability of transition from one state to another, into different bins and sorts them into a state transition matrix. At the same time, this process also captures other dynamic changes in the acoustic signal, that is, some unstable background noise in the acoustic signal, including irregular noise such as molten pool vibration and slag splashing.

[0011] Unlike patent CN202410886291.2, which focuses on processing the time domain features in the acoustic signal, the Markov graph of the present invention reflects the dynamic change pattern (such as state transition) and state transition law in the acoustic signal. The Markov transition field extracts the temporal dependency and global features of the signal to form a spatial domain picture, so that the convolutional neural network can better learn the dynamic change characteristics of the acoustic signal during the printing process and the respective complex background noise, thereby better identifying and classifying defects. This technology also predicts the next state based on the Markov transition field, extracts the adjustment of the printing process parameters through feedback control to avoid adverse defects. Summary of the invention

[0012] In order to solve the above technical problems, the present invention proposes a method for processing the characteristics of additive manufacturing acoustic signals based on Markov transfer fields. The Markov transfer field can capture the implicit patterns and dependencies in the time series and convert the dynamic change characteristics of the time series data into static images, so that the convolutional neural network can effectively process these dynamic characteristics, thereby having better classification accuracy for defect generation images. At the same time, the Markov matrix retains the global characteristics of the time series, including not only the numerical value, but also the trend and pattern of state transition, laying the foundation for the subsequent prediction of the next time state defect through images and making corresponding corrective measures in a timely manner.

[0013] To achieve the above object, the technical solution adopted by the present invention is:

[0014] The method for processing acoustic signal characteristics of additive manufacturing based on Markov transfer field includes the following steps:

[0015] S1: Create an acoustic signal dataset; design an arc additive manufacturing process experiment with different defects including discontinuity, pores, poor melting and position offset by setting different process parameters, and use an acoustic signal detector to collect acoustic signals during the printing process;

[0016] S2: preprocessing the collected acoustic signal; the acoustic signal contains background noise and arc starting and closing noise, and the preprocessing includes removing DC, filtering and cutting off noise;

[0017] S3: converting the one-dimensional acoustic signal into a two-dimensional image; selecting a suitable sampling window, converting the preprocessed acoustic signal into a two-dimensional state transition probability matrix through the Markov transition field, and mapping the two-dimensional state transition probability matrix into a two-dimensional image;

[0018] S4: Label the images according to different defects and divide them into training set, test set and validation set. Enhance the images and build a convolutional neural network model. The training set is used to train the model to learn the features and patterns in the data. The validation set is used to verify and evaluate the performance of the trained model. The test set is used to finally test the generalization performance of the model. Finally, feature extraction of arc additive manufacturing acoustic signals and defect recognition are achieved.

[0019] As a further improvement of the present invention, step S1 specifically includes:

[0020] An acoustic signal detector is used to monitor the acoustic signal generated during the arc additive manufacturing printing process. The acoustic signal detector is installed next to the arc additive manufacturing printer in a side-by-side manner.

[0021] As a further improvement of the present invention, step S2 specifically includes:

[0022] S2-1: The DC removal process can divide the signal into different frequency bands through wavelet decomposition, remove the corresponding DC component frequency and then reconstruct the signal;

[0023] S2-2: The filtering process uses a wavelet denoising algorithm to filter the background noise during the printing process;

[0024] S2-3: During the truncation process, the amplitude of the arc opening and closing noise is large and concentrated. The noise is located at the beginning and end of the signal. After truncation, the acoustic signal of the printing process is left.

[0025] As a further improvement of the present invention, step S3 specifically includes:

[0026] S3-1: Set the appropriate sliding window to m and the sliding step size to n to extract the acoustic signal with a certain overlap, in order to expand the data set and enable the convolutional neural network to fully learn its features;

[0027] S3-2: For the extracted one-dimensional signal, the one-dimensional signal is converted into a two-dimensional image based on the Markov function in the python function library downloaded by PyCharm.

[0028] As a further improvement of the present invention, step S4 specifically includes:

[0029] S4-1: dividing the two-dimensional image data set into a training set, a validation set, and a test set according to different defects and different proportions;

[0030] S4-2: The image enhancement method includes random cropping, horizontal flipping, normalization and random erasing;

[0031] S4-3: The convolutional neural network model uses Efficientnet-B0 as the backbone network to extract and learn features of the input training set; uses the AdamW optimizer to replace the original SGD optimizer; uses the CosineAnnealingLrUpdater to replace the original StepLrUpdater; inputs the training set into the convolutional neural network model for training, and obtains a preliminary image classification model after the convolutional neural network model is trained;

[0032] S4-4: The image classification model is verified through a validation set to verify the performance of the model, and hyperparameters are adjusted according to the performance to achieve the ideal performance of the model to obtain the final image classification model. The performance parameter indicators include Precision, Recall and F1-Score. The indicator formula is as follows:

[0033]

[0034] In the formula, precision represents the accuracy rate; recall represents the recall rate; True Positive is the number of samples predicted to be class i and actually class i; False Negative is the number of samples actually class i but predicted to be other classes;

[0035] S4-5: Input the test set into the final image classification model to identify and classify defects.

[0036] The advantages of the present invention are:

[0037] 1. The present invention converts one-dimensional acoustic signals into two-dimensional images through Markov transfer field (MTF) technology, thereby effectively mapping the dynamic change characteristics of acoustic signals into static images, so that convolutional neural networks can directly process complex time series data. This method can not only capture the implicit patterns and dependencies in the time series, but also retain the global characteristics of the time series through the Markov matrix, including the trend and pattern of state transition, thereby achieving more accurate defect classification.

[0038] 2. The present invention adopts a non-contact acoustic signal monitoring method, and collects acoustic signals through an acoustic signal detector installed on the side axis, without directly intervening in the arc additive manufacturing process, thus avoiding the influence of high temperature, high humidity and metal splash on the sensor. Compared with the traditional molten pool monitoring method, the acoustic signal collection process of the present invention is more adaptable to complex working conditions, while providing real-time and complete signal defect detection capabilities.

[0039] 3. The present invention further improves the generalization ability of the model by expanding the data set through sliding windows and combining image enhancement methods such as random cropping, horizontal flipping, normalization and random erasing. In the design of convolutional neural networks, EfficientNet-B0 is used as the backbone network, and the AdamW optimizer and CosineAnnealingLrUpdater strategy are used to replace traditional methods, which greatly improves the training efficiency and classification accuracy of the model.

[0040] 4. The present invention can not only identify common defects such as discontinuities and pores in the printing process, but also has the ability to predict the next time state defects, providing technical support for real-time correction and quality control of arc additive manufacturing processes. This method is also suitable for process defect detection in various additive manufacturing methods such as laser cladding, laser powder bed melting, and electron beam additive manufacturing, and has good promotion value. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 It is a flow chart of the method for processing acoustic signal characteristics and identifying defects in arc additive manufacturing based on Markov transfer field of the present invention;

[0042] Figure 2It is a schematic diagram of a printing device and an acoustic signal detector of the arc additive manufacturing acoustic signal feature extraction and defect identification method provided by an embodiment of the present invention;

[0043] The components are as follows:

[0044] 1. Central controller; 2. Sound acquisition card; 3. Sound signal detector; 4. Electric welding gun head; 5. Wire feeding and gas feeding port; 6. Six-axis robotic arm; 7. Argon shielding gas cylinder; 8. Welding machine; 9. Control cabinet; 10. Q235 substrate; 11. Workbench.

[0045] Figure 3 is a schematic diagram of a method for preprocessing a one-dimensional time-domain acoustic signal provided by an embodiment of the present invention;

[0046] Figure 4 It is a schematic diagram of a method for converting a one-dimensional time-domain acoustic signal into a two-dimensional space-domain Markov transfer field diagram provided by an embodiment of the present invention;

[0047] Figure 5 This is a flow chart of defect recognition and classification using a convolutional neural network provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0048] Please refer to Figure 1 As shown, an embodiment of the present invention provides a method for processing acoustic signal characteristics and identifying defects in arc additive manufacturing based on Markov transfer field, the method comprising:

[0049] S1: Create an acoustic signal dataset; by setting different process parameters, design the acoustic signals of three different defects in the arc additive manufacturing printing process: discontinuity, pores, and normal, and use an acoustic signal detector to collect them.

[0050] S2: Preprocessing the collected acoustic signal; the acoustic signal contains background noise and arc starting and closing noise, and the preprocessing includes DC removal, filtering and noise truncation.

[0051] S3: Convert the one-dimensional acoustic signal into a two-dimensional image; select a suitable sampling window, convert the preprocessed acoustic signal into a two-dimensional state transfer probability matrix through the Markov transition field, and map the two-dimensional state transfer probability matrix into a two-dimensional image.

[0052] S4: Label the images according to different defects and divide them into training set, test set and validation set. Enhance the images and build a convolutional neural network model. The training set is used to train the model to learn the features and patterns in the data. The validation set is used to verify and evaluate the performance of the trained model. The test set is used to finally test the generalization performance of the model. Finally, feature extraction of arc additive manufacturing acoustic signals and defect recognition are achieved.

[0053] Please refer to Figure 2As shown, a schematic diagram of a printing device and an acoustic signal detector for an arc additive manufacturing acoustic signal feature extraction and defect identification method provided in an embodiment of the present invention, the method comprising:

[0054] S1: The printing device and the acoustic signal detector of the arc additive manufacturing acoustic signal feature extraction and defect identification method include a central controller 1, a sound acquisition card 2, an acoustic signal detector 3, an electric welding gun head 4, a wire feeding and gas feeding port 5, a six-axis robot arm 6, an argon shielding gas cylinder 7, a welding machine 8 and a control cabinet 9, a Q235 substrate 10, and a workbench 11; the device is installed and connected as shown in the figure.

[0055] S2: In the process of arc additive manufacturing, the dynamic fluctuation of the molten pool, welding spatter and other phenomena will generate specific acoustic signals. First, a high-precision acoustic signal detector is used to collect the acoustic signals in the arc additive manufacturing process. The collected acoustic signals are sound pressure data, which are collected using the DAQ module of Labview. The sampling frequency is 40000Hz, and the sound pressure data is saved as a TDMS file. The detector is installed in a side-axis manner to avoid interference that may affect the sensor, such as high temperature, high humidity and metal spatter. During the acquisition process, by adjusting the process parameters of arc additive manufacturing, such as current, voltage, shielding gas flow and scanning speed, acoustic signals with three typical defect characteristics are designed: discontinuous defects, pore defects and normal state signals.

[0056] S3: After the signal acquisition is completed, it is divided into three data sets according to discontinuous defects, pore defects and normal states, providing basic data for subsequent signal preprocessing and feature extraction.

[0057] Please refer to Figure 3 As shown, a schematic diagram of a method for preprocessing a one-dimensional time-domain acoustic signal provided by an embodiment of the present invention, the method comprising:

[0058] S1: Remove DC. Decompose the signal through wavelet transform and divide it into multiple frequency bands. For the low-frequency band (DC component), remove its frequency component and reconstruct the signal through inverse wavelet transform. This can eliminate the DC noise caused by the start and end of the arc, ensuring that the remaining signal can better reflect the acoustic characteristics of the actual manufacturing process.

[0059] S2: Filtering and denoising, using the wavelet denoising algorithm to filter the signal, using the wavelet denoising module built into Labview to filter out the interference noise in the sound signal, with the specific parameters configured as db04 wavelet, soft threshold and four-level decomposition, filtering out the background noise generated by the environmental noise (such as wind, equipment vibration, etc.) in the arc additive manufacturing process. This method can effectively retain the useful components of the signal while suppressing the interference of irrelevant noise.

[0060] S3: Cut off the noise. The noise during arc starting and closing is usually large and concentrated, usually located at the beginning and end of the signal. By observing the start and end time periods of the one-dimensional time domain acoustic signal, these unnecessary arc opening and closing noise components are cut off to retain the effective signal during the welding process. After these preprocessing steps, the final signal will have a higher signal-to-noise ratio, providing clear signal data for subsequent feature extraction.

[0061] Please refer to Figure 4 As shown, a schematic diagram of a method for converting a one-dimensional time-domain acoustic signal into a two-dimensional space-domain Markov transfer field diagram provided by an embodiment of the present invention, the method comprising:

[0062] S1: Import necessary libraries in PyCharm, import TdmsFile, import numpy, MTF library and matplotlib.pyplot library, among which TdmsFile is used to open the acoustic signal file saved in Tdms format, and the data of a specific channel is read and converted into a NumPy array. Slice the data, select a sliding window with a data length of m = 20000, and extract 20000 data points each time; the sliding step length n = 4000, skip 4000 data points and continue to process new data segments, and finally obtain a sufficient number of processed data segments, which are all independent time series segments.

[0063] S2: The data segment is converted from a one-dimensional acoustic signal to a two-dimensional spatial domain image through the introduced MTF library, and 20,000 data points are quantized and divided into 8 different quantile bins. The quantization process is quantized using a quantile strategy, and the frequency of transferring a data point in a quantile bin to other quantile bins (including the quantile bin where the data originally resides) is constructed into a transfer frequency matrix W, forming a Markov transfer field M; wherein the formulas of the transfer frequency matrix and the Markov transfer field are as follows:

[0064] ∑w ij =1(i,j∈1…Q)

[0065]

[0066] In the formula, w ij It represents the frequency of transfer from bin i to bin j.

[0067] S3: Use Matplotlib to set up a rainbow-hued color map to display the image, set the origin of the image to the lower left corner, hide the coordinate axes to make the image display clearer, and finally map the Markov transfer field into a spatial domain image for display.

[0068] Please refer to Figure 5As shown, a convolutional neural network defect recognition and classification flow chart provided by an embodiment of the present invention, the method includes:

[0069] S1: The obtained spatial domain images are divided into training set, validation set and test set in a ratio of 8:1:1. The training is performed to enhance the image, and the enhancement methods include random cropping, horizontal flipping, normalization and random erasing, and then input into the convolutional neural network for training. The convolutional neural network model uses Efficientnet-B0 as the backbone network for feature extraction and learning of the input training set; the AdamW optimizer is used to replace the original SGD optimizer; the CosineAnnealingLrUpdater is used to replace the original StepLrUpdater; the training set is input into the convolutional neural network model for training, and a preliminary image classification model is obtained after the convolutional neural network model is trained; the validation set is input into the image classification model for verification, and the hyperparameters are adjusted according to the recognition and classification effects of the acoustic signal features in the image; and the test is performed after the final parameters are adjusted.

[0070] S2: The image classification model is verified through the validation set. The hyperparameters are adjusted according to the performance to achieve the ideal performance of the model and obtain the final image classification model. The performance parameter indicators include Precision, Recall and F1-Score. The indicator formula is as follows:

[0071]

[0072] In the formula, precision represents the accuracy rate; recall represents the recall rate; True Positive (TP) is the number of samples predicted to be class i and actually class i; False Negative (FN) is the number of samples actually class i but predicted to be other classes; the verification results are as follows:

[0073] category Precision Recall F1-Score Detection time Discontinuous 98.84 96.59 97.70 0.013 normal 96.97 100.0 98.46 0.014 Pores 100.0 99.06 99.53 0.014

[0074] S3: Implement the image classification model to extract, identify and accurately classify the defect acoustic signal features in the image.

[0075] The above description is only a preferred embodiment of the present invention and does not constitute any other form of limitation to the present invention. Any modification or equivalent change made based on the technical essence of the present invention still falls within the scope of protection required by the present invention.

Claims

1. A method for processing acoustic signal characteristics of additive manufacturing based on Markov transfer field, characterized by: The following steps are involved: S1: Create an acoustic signal dataset; design an arc additive manufacturing process experiment with different defects including discontinuity, pores, poor melting and position offset by setting different process parameters, and use an acoustic signal detector to collect acoustic signals during the printing process; S2: preprocessing the collected acoustic signal; the acoustic signal contains background noise and arc starting and closing noise, and the preprocessing includes removing DC, filtering and cutting off noise; S3: Convert one-dimensional acoustic signals into two-dimensional images; Under a suitable sampling window, the preprocessed acoustic signal is converted into a two-dimensional state transition probability matrix through the Markov transition field, and the two-dimensional state transition probability matrix is ​​mapped into a two-dimensional image; S4: Label the images according to different defects and divide them into training set, test set and validation set. Enhance the images and build a convolutional neural network model. The training set is used to train the model to learn the features and patterns in the data. The validation set is used to verify and evaluate the performance of the trained model. The test set is used to finally test the generalization performance of the model. Finally, feature extraction of arc additive manufacturing acoustic signals and defect recognition are achieved.

2. The method for processing acoustic signal characteristics of additive manufacturing based on Markov transfer field according to claim 1, characterized in that: Step S1 specifically includes: An acoustic signal detector is used to monitor the acoustic signal generated during the arc additive manufacturing printing process. The acoustic signal detector is installed next to the arc additive manufacturing printer in a side-by-side manner.

3. The method for processing acoustic signal characteristics of additive manufacturing based on Markov transfer field according to claim 1, characterized in that: Step S2 specifically includes: S2-1: The DC removal process can divide the signal into different frequency bands through wavelet decomposition, remove the corresponding DC component frequency and then reconstruct the signal; S2-2: The filtering process uses a wavelet denoising algorithm to filter the background noise during the printing process; S2-3: During the truncation process, the amplitude of the arc opening and closing noise is large and concentrated. The noise is located at the beginning and end of the signal. After truncation, the acoustic signal of the printing process is left.

4. The method for processing acoustic signal characteristics of additive manufacturing based on Markov transfer field according to claim 1, characterized in that: Step S3 specifically includes: S3-1: Set the appropriate sliding window to m and the sliding step size to n to extract the acoustic signal with a certain overlap, in order to expand the data set and enable the convolutional neural network to fully learn its features; S3-2: For the extracted one-dimensional signal, the one-dimensional signal is converted into a two-dimensional image based on the Markov function in the python function library downloaded by PyCharm.

5. The method for processing acoustic signal characteristics of additive manufacturing based on Markov transfer field according to claim 1, characterized in that: Step S4 specifically includes: S4-1: dividing the two-dimensional image data set into a training set, a validation set, and a test set according to different defects and different proportions; S4-2: The image enhancement method includes random cropping, horizontal flipping, normalization and random erasing; S4-3: The convolutional neural network model uses Efficientnet-B0 as the backbone network to extract and learn features of the input training set; uses the AdamW optimizer to replace the original SGD optimizer; uses the CosineAnnealingLrUpdater to replace the original StepLrUpdater; inputs the training set into the convolutional neural network model for training, and obtains a preliminary image classification model after the convolutional neural network model is trained; S4-4: The image classification model is verified through a validation set to verify the performance of the model, and hyperparameters are adjusted according to the performance to achieve the ideal performance of the model to obtain the final image classification model. The performance parameter indicators include Precision, Recall and F1-Score. The indicator formula is as follows: In the formula, precision represents the accuracy rate; recall represents the recall rate; True Positive is the number of samples predicted to be class i and actually class i; False Negative is the number of samples actually class i but predicted to be other classes; S4-5: Input the test set into the final image classification model to identify and classify defects.

Images