Defect detection model training method, defect detection method, system and equipment

Abstract

The invention provides a defect detection model training method, a defect detection method, a defect detection system and defect detection equipment. The training method comprises the following steps: acquiring a sample powder spreading image of a powder bed in an additive manufacturing process and corresponding sample defect data; acquiring sample time sequence data corresponding to the sample powder spreading image in the laser melting process; and taking the sample powder spreading image and the sample time sequence data as input, taking corresponding sample defect data as output, and training to obtain a defect detection model. According to the method, the sample powder spreading image and the corresponding sample time sequence data are combined, the defect detection model is obtained through training, and the effectiveness and reliability of the defect detection model are guaranteed. And inputting the actual powder spreading image and the actual time sequence data into the defect detection model to output the actual defect data, so that real-time anomaly detection in the additive manufacturing process is realized, complementarity among different modes is fully utilized, the accuracy and reliability of defect detection are remarkably improved, and the production efficiency is improved.

Description

Technical Field

[0001] The present disclosure relates to the field of additive manufacturing technology, and particularly to a method for training a defect detection model, a defect detection method, a system and a device. Background Art

[0002] Additive manufacturing is an advanced technology for manufacturing three-dimensional objects by layer-by-layer material deposition, which is widely used in many fields such as aerospace, medical devices, and automotive manufacturing. Selective Laser Melting (SLM) technology is an important part of additive manufacturing technology and is currently the most widely used powder bed additive manufacturing technology. Compared with other traditional manufacturing methods, SLM technology is not restricted by the complex shape of parts during the manufacturing process, has a fast and efficient production process, a low manufacturing cost, and an environmentally friendly characteristic of no pollution. At the same time, the overall quality of the formed parts is also very excellent. Therefore, this technology shows irreplaceable advantages in manufacturing metal parts with complex structures compared with traditional production and processing techniques.

[0003] With the wide application of SLM, the importance of ensuring the quality of formed parts is increasing day by day. However, most of the existing additive manufacturing defect detection technologies only focus on the monitoring of surface defects or conduct unified detection of the interior after the finished product is completed, and cannot achieve real-time monitoring, resulting in the inability to handle defects in a timely manner after they occur; in addition, in a complex environment, the detection accuracy of the existing additive manufacturing defect detection technologies is affected by various factors, which is prone to false alarms and missed detections.

[0004] In summary, the existing additive manufacturing defect detection technologies have low real-time performance, accuracy, and reliability. This makes it difficult to timely detect and correct workpiece defect problems during the additive manufacturing process, increases production losses, and reduces production efficiency. Summary of the Invention

[0005] The technical problem to be solved by the present disclosure is to overcome the defects in the existing additive manufacturing defect detection technologies, such as only focusing on the monitoring of surface defects or conducting unified detection of the interior after the finished product is completed, resulting in low real-time performance, accuracy, reliability, production efficiency, and large production losses. The present disclosure provides a method for training a defect detection model, a defect detection method, a system and a device.

[0006] The present disclosure solves the above technical problems through the following technical solutions:

[0007] The present disclosure provides a method for training a defect detection model, and the training method includes:

[0008] Obtaining a sample powder bed spreading image and corresponding sample defect data during the additive manufacturing process;

[0009] Obtain the sample time series data corresponding to the powder spreading image of the sample during the laser melting process;

[0010] Using the sample powder spreading image and the sample time series data as inputs, and the corresponding sample defect data as outputs, train to obtain the defect detection model.

[0011] Optionally, the step of using the sample powder spreading image and the sample time series data as inputs, and the corresponding sample defect data as outputs, to train to obtain the defect detection model includes:

[0012] Respectively obtain the first feature data corresponding to the sample powder spreading image and the sample time series data through a convolution module;

[0013] Perform feature fusion processing on the first feature data to obtain second feature data;

[0014] Using the second feature data as an input and the corresponding sample defect data as an output, train to obtain the defect detection model.

[0015] Optionally, the step of respectively obtaining the first feature data corresponding to the sample powder spreading image and the sample time series data through a convolution module includes:

[0016] Respectively obtain the third feature data corresponding to the sample powder spreading image and the sample time series data through the convolution module;

[0017] Adopt a feature interaction mechanism to perform feature interaction processing on the third feature data to obtain the first feature data.

[0018] Optionally, the step of performing feature fusion processing on the first feature data to obtain second feature data includes:

[0019] Respectively obtain the weights corresponding to each of the first feature data by using a channel attention mechanism;

[0020] Based on the weights, perform feature fusion processing on the first feature data to obtain the second feature data;

[0021] And / or,

[0022] The sample time series data includes at least one of laser power data and molten pool intensity voltage data.

[0023] The present disclosure also provides a defect detection method, and the defect detection method includes:

[0024] Obtain the actual powder spreading image of the powder bed during the additive manufacturing process;

[0025] Obtain the actual time series data corresponding to the actual powder spreading image during the laser melting process;

[0026] Input the actual powder spreading image and the actual time series data into the defect detection model to output actual defect data;

[0027] Wherein, the defect detection model is obtained based on the training method of the above-mentioned defect detection model.

[0028] The present disclosure also provides a training system for a defect detection model, and the training system includes:

[0029] A first data acquisition module, configured to acquire a sample powder spreading image of a powder bed and corresponding sample defect data during the additive manufacturing process;

[0030] A second data acquisition module, configured to acquire sample time series data corresponding to the sample powder spreading image during the laser melting process;

[0031] A model training module, configured to use the sample powder spreading image and the sample time series data as inputs, and the corresponding sample defect data as outputs, to train and obtain the defect detection model.

[0032] Optionally, the model training module includes:

[0033] A first data acquisition unit, configured to respectively acquire first feature data corresponding to the sample powder spreading image and the sample time series data through a convolution module;

[0034] A second data acquisition unit, configured to perform feature fusion processing on the first feature data to obtain second feature data;

[0035] A model training unit, configured to use the second feature data as an input, and the corresponding sample defect data as an output, to train and obtain the defect detection model.

[0036] Optionally, the first data acquisition unit includes:

[0037] A third data acquisition subunit, configured to respectively acquire third feature data corresponding to the sample powder spreading image and the sample time series data through the convolution module;

[0038] A first data acquisition subunit, configured to perform feature interaction processing on the third feature data by using a feature interaction mechanism to obtain the first feature data.

[0039] Optionally, the second data acquisition unit includes:

[0040] A weight acquisition subunit, configured to respectively acquire weights corresponding to each of the first feature data by using a channel attention mechanism;

[0041] A second data acquisition subunit, configured to perform feature fusion processing on the first feature data based on the weights to obtain the second feature data;

[0042] and / or

[0043] The sample time series data includes at least one of laser power data and molten pool intensity voltage data.

[0044] The present disclosure also provides a defect detection system, which includes:

[0045] An actual image acquisition module, configured to acquire an actual powder spreading image of a powder bed during an additive manufacturing process;

[0046] A time data acquisition module, configured to acquire actual time series data corresponding to the actual powder spreading image during a laser melting process;

[0047] A model output module, configured to input the actual powder spreading image and the actual time series data into a defect detection model to output actual defect data;

[0048] Wherein, the defect detection model is obtained based on the training system of the defect detection model described above.

[0049] The present disclosure also provides an electronic device, including a memory, a processor, and a computer program stored on the memory and configured to run on the processor. When the processor executes the computer program, it implements the training method of the defect detection model described above, or implements the defect detection method described above.

[0050] The present disclosure also provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, it implements the training method of the defect detection model described above, or implements the defect detection method described above.

[0051] The present disclosure also provides a computer program product, including a computer program. When the computer program is executed by a processor, it implements the training method of the defect detection model as described above, or implements the defect detection method as described above.

[0052] On the basis of conforming to common knowledge in the art, the above preferred conditions can be combined arbitrarily to obtain various preferred examples of the present disclosure.

[0053] The positive and progressive effects of the present disclosure are as follows:

[0054] The present disclosure combines the sample powder spreading images and the corresponding sample time series data to train a defect detection model, ensuring the effectiveness and reliability of the defect detection model. By inputting the actual powder spreading images and the actual time series data into the defect detection model to output the actual defect data, real-time anomaly detection in the additive manufacturing process is achieved, enabling timely discovery and handling of defects, thereby reducing production losses, lowering the false alarm rate of defect detection, making full use of the complementarity between different modalities, significantly improving the accuracy and reliability of defect detection, reducing the subsequent detection cost, increasing production efficiency, and promoting the sustainable development of the additive manufacturing industry. Description of the Drawings

[0055] Figure 1 It is a flowchart of the training method of the defect detection model according to Embodiment 1 of the present disclosure;

[0056] Figure 2 It is a specific example diagram of the sample powder spreading image according to Embodiment 1 of the present disclosure;

[0057] Figure 3 It is a flowchart of step S13 in the training method of the defect detection model according to Embodiment 2 of the present disclosure;

[0058] Figure 4 It is a flowchart of step S131 in the training method of the defect detection model according to Embodiment 2 of the present disclosure;

[0059] Figure 5 It is a flowchart of step S132 in the training method of the defect detection model according to Embodiment 2 of the present disclosure;

[0060] Figure 6 It is a time-laser power value curve diagram according to Embodiment 2 of the present disclosure;

[0061] Figure 7 It is a time-molten pool intensity voltage value curve diagram according to Embodiment 2 of the present disclosure;

[0062] Figure 8 It is a flowchart of the defect detection method according to Embodiment 3 of the present disclosure;

[0063] Figure 9 It is a specific example diagram of the defect detection image according to Embodiment 3 of the present disclosure;

[0064] Figure 10 It is an architecture example diagram of the multi-modal convolutional neural network according to Embodiment 3 of the present disclosure;

[0065] Figure 11 It is a module schematic diagram of the training system of the defect detection model according to Embodiment 4 of the present disclosure;

[0066] Figure 12 It is a module schematic diagram of the training system of the defect detection model according to Embodiment 5 of the present disclosure;

[0067] Figure 13 It is a schematic diagram of the modules of the defect detection system according to Embodiment 6 of the present disclosure;

[0068] Figure 14 It is a schematic diagram of the structure of the electronic device according to Embodiment 7 of the present disclosure. Specific embodiments

[0069] The present disclosure will be further described below by way of embodiments, but the present disclosure is not limited to the scope of the described embodiments.

[0070] In the embodiments of the present disclosure, prefix words such as "first" and "second" are only used to distinguish different described objects, and have no limiting effect on the position, order, priority, quantity or content of the described objects, etc. The use of ordinal numbers and other prefix words for distinguishing described objects in the embodiments of the present disclosure does not constitute a limitation on the described objects. For the description of the described objects, refer to the description in the context of the embodiments, and there should be no redundant limitation due to the use of such prefix words. In addition, in the description of this embodiment, unless otherwise specified, the meaning of "a plurality" is two or more.

[0071] Embodiment 1

[0072] This embodiment provides a training method for a defect detection model, as Figure 1 shown, the training method includes:

[0073] S11. Obtain sample powder bed laying images and corresponding sample defect data during the additive manufacturing process;

[0074] S12. Obtain sample time series data corresponding to the sample powder bed laying images during the laser melting process;

[0075] S13. Use the sample powder bed laying images and the sample time series data as inputs, and the corresponding sample defect data as outputs to train and obtain a defect detection model.

[0076] Specifically, the forming process of the selective laser melting technology mainly includes the following steps:

[0077] (1) Obtain processing data. Design a 3D model of the formed part through CAD (a kind of drawing software) software, and use slicing software to perform layer slicing on the 3D model to generate the processing path data of each layer.

[0078] (2) Lay powder. Use a powder laying component (such as a scraper or a roller) to evenly lay metal powder on the working plane to ensure that the surface of the powder bed is flat and free of air bubbles.

[0079] (3) Laser scanning and melting. The laser beam scans according to the processing path data generated in step (1). The laser energy melts the powder material to form interlayer bonding and constructs the contour of the formed part.

[0080] (4) Lowering the working platform. After one layer of processing is completed, the working platform automatically descends to the set thickness of the formed layer to prepare for powder spreading and laser melting of the next layer.

[0081] (5) Repeat steps (2), (3), and (4) until all the processing path data is processed, and finally a complete formed part is formed.

[0082] (6) Post-processing. Remove the excess unmelted powder, which can be done by blowing air, vibrating, or cleaning. Depending on the need, some formed parts also need to be heat-treated to improve their mechanical properties and density.

[0083] During the processing of preparing samples by selective laser melting technology, an industrial camera is used to collect the sample powder spreading images of the powder bed for each layer during the processing, and the sample time series data of the same layer is obtained through the laser melt pool intensity monitoring system. The sample time series data can be multiple curves regarding time. The laser melt pool intensity monitoring system is a monitoring system in the prior art and will not be elaborated here. As Figure 2 is a specific example diagram of the sample powder spreading image.

[0084] Defect annotation is performed on the sample powder spreading images of each layer of the powder bed to obtain the annotated sample powder spreading images, that is, the corresponding sample defect data is obtained.

[0085] Taking the single-layer sample powder spreading image and the sample time series data of the same layer as inputs, and the corresponding sample defect data as outputs, they are respectively input into the multi-modal convolutional neural network (MMFCNN) for model training to obtain a trained network model, that is, a defect detection model.

[0086] In this embodiment, by combining the sample powder spreading image and the corresponding sample time series data, a defect detection model is trained, ensuring the effectiveness and reliability of the defect detection model.

[0087] Embodiment 2

[0088] This embodiment provides a training method for a defect detection model, which is a further improvement of Embodiment 1.

[0089] In an implementable solution, as Figure 3 shown, step S13 includes:

[0090] S131. Respectively obtain the first feature data corresponding to the sample powder spreading image and the sample time series data through the convolutional module;

[0091] S132. Perform feature fusion processing on the first feature data to obtain second feature data;

[0092] S133. Use the second feature data as input and the corresponding sample defect data as output to train a defect detection model.

[0093] Specifically, there are several sample time series data, for example, there can be two. The sample powder spreading image and the two sample time series data are three modal data. The sample powder spreading image is a two-dimensional spatial signal, and the two sample time series data are both one-dimensional time series signals. For each modal data, the first feature data is extracted through an independent convolution module, and finally, the first feature data is fused through a feature fusion module to obtain the second feature data for training the defect detection model. The first feature data can be a feature map.

[0094] The architecture of the convolution module is established through a series of convolutional neural network (CNN) layers. Each modal data sequentially undergoes convolution operation of the convolution layer, batch normalization operation, and application operation of the activation function through the convolution module. The convolution layer uses a learnable convolution kernel to perform convolution operation on the modal data to extract local features. The batch normalization (BN) operation is used to alleviate the internal covariate shift, standardize the feature distribution, accelerate the training convergence, and suppress overfitting. The activation function is used to introduce non-linear expression ability.

[0095] The one-dimensional time series signal, that is, the sample time series data, is processed through discrete one-dimensional convolution. The corresponding mathematical expression is as follows:

[0096]

[0097] Among them, x[n] represents the discrete input sequence (such as the time-power curve), with a length of N, n represents the index of the discrete input sequence, ω[k] represents the one-dimensional convolution kernel (filter), with a length of K, m represents the index of the output sequence, and * represents the discrete convolution operator.

[0098] The two-dimensional spatial signal, that is, the sample powder spreading image, is processed through two-dimensional convolution. The corresponding mathematical expression is as follows:

[0099]

[0100] Among them, X represents the input image matrix, with a size of H in ×W in ,W represents the two-dimensional convolution kernel, with a size of H k ×W k ,u, v represent the indices of the two-dimensional convolution kernel, and Y[i, j] represents the value at the position (i, j) in the output feature map.

[0101] The calculation of the batch normalization operation is divided into two steps, namely standardization and affine transformation in sequence.

[0102] The corresponding mathematical expression for standardization is as follows:

[0103]

[0104] The corresponding mathematical expression for affine transformation is as follows:

[0105]

[0106] Among them, μ and σ 2 respectively represent the mean and variance of the input data of the current batch. ∈ is a very small constant to prevent division by zero. γ and β represent the learnable scaling and translation parameters, x represents the input data, represents the input data after standardization, and y represents the input data after affine transformation.

[0107] The application operation of the activation function can apply the ReLU (Rectified Linear Unit) activation function, and its mathematical expression is as follows:

[0108] f(x) = max(0, x);

[0109] Among them, x represents the eigenvalue output by the convolutional layer, and f(x) represents the eigenvalue after applying the activation function.

[0110] The feature fusion module performs four fusion strategies on the three first feature data F x , F y , F z They are divided into maximum fusion operation, average fusion operation, concatenation fusion operation and summation fusion operation.

[0111] The maximum fusion operation is used to calculate the maximum value of the three first feature data in the c-th channel at the same spatial position a, and the corresponding mathematical expression is as follows:

[0112]

[0113] Among them, F x , F y , F z respectively represent the three first feature data, a represents the spatial position, c represents the channel ordinal number, represents the output of the maximum fusion operation of the three first feature data.

[0114] The average fusion operation is used to calculate the average value of the three first feature data in the c-th channel at the same spatial position a, and the corresponding mathematical expression is as follows:

[0115]

[0116] Among them, F x 、F y 、F z respectively represent three first feature data, a represents the spatial position, c represents the channel ordinal number, represents the output of the average fusion operation of the three first feature data.

[0117] The concatenation fusion operation is used to stack the first feature data along the channel direction, and the corresponding mathematical expression is as follows:

[0118]

[0119] Among them, represents the output of the concatenation fusion operation of the three first feature data, F x 、F y 、F z respectively represent three first feature data, and C represents the number of channels.

[0120] The summation fusion operation is used to calculate the sum of the first feature data at the same spatial position. Since the importance of each first feature data is not clear, learnable parameters α, β, γ are assigned to the three first feature data, and these parameters represent the weights of each first feature data. The corresponding mathematical expression is as follows:

[0121]

[0122] Among them, a represents the spatial position, c represents the channel ordinal number, C represents the number of channels, k takes integer values from 1 to C, F x 、F y 、F z respectively represent three first feature data, represents the output of the summation fusion operation of the three first feature data.

[0123] In this solution, the first feature data is obtained by feature extraction of the sample powder spreading image and the sample time series data through the convolution module, and then the first feature data is subjected to feature fusion processing to obtain the second feature data, so as to train a defect detection model, which can comprehensively obtain the defect area information of the additive manufacturing workpiece and improve the accuracy and reliability of the defect detection model.

[0124] In an implementable solution, as Figure 4 shown, step S131 includes:

[0125] S1311. Obtain the corresponding third feature data of the sample powder spreading image and the sample time series data through the convolution module respectively;

[0126] S1312. Use the feature interaction mechanism to perform feature interaction processing on the third feature data to obtain the first feature data.

[0127] Specifically, the global features of a single modality are crucial for classification. The feature interaction mechanism is introduced to utilize this global information so that each branch contains important information from other branches. To promote information complementarity among multi-modal data, asymmetric interactions are performed on the three modality data (one-dimensional time series signals) or (two-dimensional spatial signals). Feature interactions are performed between any two of the three modality data. Since the operation of feature interaction is asymmetric, selecting features from three domains for feature interaction ultimately results in redundant information in the fused features. Among them, represents the set of real numbers, B represents the batch size, C represents the number of channels, L represents the length of the one-dimensional time series signal, and H, W represent the dimensions of the two-dimensional spatial signal. The feature interaction mechanism includes two operations: significant feature extraction and feature concatenation. Taking F x and F y as an example, the operation process is as follows.

[0128] Significant feature extraction is used to perform max pooling on F y along the channel dimension to obtain the global significant feature The corresponding mathematical expression is as follows:

[0129] G y = MaxPool(F y );

[0130] Feature concatenation is used to concatenate G y and F x along the channel dimension. The corresponding mathematical expression is as follows:

[0131]

[0132] Among them, represents the enhanced feature, and Concat(·) represents the concatenation operation along the channel dimension.

[0133] In this solution, cross-modal information fusion is achieved through the feature interaction mechanism, making full use of the complementarity between different modality data, improving the accuracy and reliability of the first feature data, and further enhancing the comprehensiveness, accuracy, and reliability of the defect detection model.

[0134] In an implementable solution, as Figure 5 shown, step S132 includes:

[0135] S1321. Use the channel attention mechanism to obtain the weight corresponding to each first feature data respectively;

[0136] S1322. Based on weights, perform feature fusion processing on the first feature data to obtain second feature data.

[0137] Specifically, in a convolutional neural network, each channel of a feature map corresponds to the activation response of a convolutional kernel. Introducing a channel attention mechanism into a convolutional neural network can be regarded as a process of selecting semantics. The channel attention mechanism learns the weights of each channel and improves the representation performance of convolutional features by suppressing irrelevant features. The channel attention mechanism includes two operations, namely global information compression and channel weight learning.

[0138] Global information compression is used to perform adaptive average pooling on the first feature data in the spatial dimension to obtain a channel descriptor The corresponding mathematical expression is as follows:

[0139] z = AvgPool(F);

[0140] Channel weight learning is used to generate attention weights through two fully connected layers (MLP). The corresponding mathematical expression is as follows:

[0141] Attention = σ(W2·ReLU(W1·z));

[0142] Where, and represent learnable weight matrices, r represents the compression ratio, σ represents the Sigmoid activation function, and maps the attention weights to [0, 1].

[0143] In this solution, the importance of each first feature data, that is, the weight, is dynamically adjusted through the channel attention mechanism. Through adaptive feature extraction, the most representative features are automatically selected for analysis, which can suppress irrelevant features, more effectively mine the information in different first feature data, enhance the second feature data, and thus improve the accuracy, reliability, and robustness of the defect detection model.

[0144] In an implementable solution, the sample time series data includes at least one of laser power data and molten pool intensity voltage data.

[0145] Specifically, the laser power data is a time-laser power value curve, as Figure 6 shown; the molten pool intensity voltage data is a time-molten pool intensity voltage value curve, as Figure 7 shown.

[0146] The sample time series data can be standardized. Standardization is a data preprocessing method that linearly transforms the data by feature (such as each sensor channel of the time series), making its mean 0 and standard deviation 1, thereby eliminating the dimensional difference and distribution offset. For time series data (such as the time-pool intensity voltage value curve or the time-laser power value curve), assuming the original data of a certain channel is x = {x1, x2, …, x N}, the standardization formula is:

[0147]

[0148] where μ represents the mean of the original data, σ represents the standard deviation of the original data, x standardized represents the data obtained after standardization processing, and i represents the index of the original data.

[0149] In addition, the training process of the model is optimized using the cross-entropy loss function, and the network parameters are updated through the backpropagation algorithm.

[0150] In this solution, a defect detection model is trained through sample powder spreading images, laser power data, and pool intensity voltage data, enabling the defect detection model to comprehensively analyze and predict information such as infrared thermal imaging images and clad laser data in the internal filling path during the additive manufacturing process. The defect detection model can identify powder spreading defects and defects during the laser melting process, improving the practicality and reliability of the model.

[0151] In this embodiment, a defect detection model is trained by combining sample powder spreading images and corresponding sample time series data, ensuring the effectiveness and reliability of the defect detection model.

[0152] Embodiment 3

[0153] This embodiment provides a defect detection method, as Figure 8 shown, the defect detection method includes:

[0154] S21. Obtain the actual powder spreading image of the powder bed during the additive manufacturing process;

[0155] S22. Obtain the actual time series data corresponding to the actual powder spreading image during the laser melting process;

[0156] S23. Input the actual powder spreading image and the actual time series data into the defect detection model to output actual defect data;

[0157] where the defect detection model is obtained based on the training method of the defect detection model in Embodiment 1 or Embodiment 2.

[0158] Specifically, the actual time series data can be multiple curves regarding time, such as the time-molten pool intensity voltage value curve and the time-laser power value curve. The unlabeled actual powder spreading image and the corresponding actual time series data are input into the trained defect detection model for defect detection to obtain actual defect data.

[0159] In this embodiment, the actual powder spreading image and the actual time series data are input into the defect detection model to output actual defect data, realizing real-time anomaly detection in the additive manufacturing process, being able to timely discover and handle defects, thereby reducing production losses, reducing the false alarm rate of defect detection, making full use of the complementarity between different modalities, significantly improving the accuracy and reliability of defect detection, reducing the later detection cost, improving production efficiency, and promoting the sustainable development of the additive manufacturing industry.

[0160] The working principle of the defect detection method in this embodiment is illustrated below with specific examples:

[0161] During the selective laser melting process, an industrial camera is used to collect the sample powder spreading images of the powder bed during the forming process of each layer in real time. At the same time, the time-molten pool intensity voltage value curve and the time-laser power value curve, that is, the sample time series data, are obtained through the laser molten pool intensity monitoring system.

[0162] Defect labeling is performed on the collected sample powder spreading images to generate a labeled training dataset. At the same time, the time-molten pool intensity voltage value curve and the time-laser power value curve are subjected to standardization processing to obtain the standardized sample time series data.

[0163] The labeled sample powder spreading images and their corresponding standardized sample time series data are used as inputs and fed into the multi-modal convolutional neural network in batches for training. The training process of the model is optimized using the cross-entropy loss function, and the network parameters are updated through the backpropagation algorithm.

[0164] After the model training is completed, the unlabeled actual powder spreading image and the corresponding actual time series data are input into the trained MMFCNN model for defect detection. The model outputs the defect detection result, that is, the actual defect data. The actual defect data includes the defect detection image and the defect detection text. The defect detection image is as Figure 9 shown, Figure 9 The positions of multiple defects are marked in the actual powder spreading image. The defect detection text is shown in the following table. The first row in the table represents the radius of the defect, the diameter of the defect, the x coordinate of the center point of the defect, the y coordinate of the center point of the defect, the z coordinate of the center point of the defect, the volume of the defect, and the surface area of the defect. The second to sixth rows in the table represent the information corresponding to each defect.

[0165]

[0166] For the detected defects, corresponding reports are generated and can be fed back to the operator through the human-machine interaction interface for further processing and adjustment.

[0167] Such as Figure 10 As shown in the architecture example diagram of the multi-modal convolutional neural network, the multi-modal convolutional neural network includes three convolutional layers, two of which are one-dimensional convolutional layers (1-D convolution) and one is a two-dimensional convolutional layer (2-D convolution). The one-dimensional convolutional layer is used to extract features from the input actual time series data, and the two-dimensional convolutional layer is used to extract features from the input actual powder spreading image. After feature extraction, the third feature data is obtained. The feature interaction mechanism (Crossattention) is used to perform feature interaction processing on the third feature data to obtain the first feature data, which is represented as a feature map (feature map). Feature fusion processing is performed on the first feature data to obtain the second feature data. The second feature data is input to the fully connected layer (FC) to output the actual defect data, completing defect detection.

[0168] Embodiment 4

[0169] This embodiment provides a training system for a defect detection model. As Figure 11 shown, the training system includes:

[0170] The first data acquisition module 11 is used to acquire the sample powder spreading images of the powder bed and the corresponding sample defect data during the additive manufacturing process;

[0171] The second data acquisition module 12 is used to acquire the sample time series data corresponding to the sample powder spreading images during the laser melting process;

[0172] The model training module 13 is used to train a defect detection model by taking the sample powder spreading images and the sample time series data as inputs and the corresponding sample defect data as outputs.

[0173] In this embodiment, by combining the sample powder spreading images and the corresponding sample time series data, a defect detection model is trained, ensuring the effectiveness and reliability of the defect detection model.

[0174] Embodiment 5

[0175] This embodiment provides a training system for a defect detection model, which is a further improvement of Embodiment 4.

[0176] In an implementable solution, as Figure 12 shown, the model training module 13 includes:

[0177] The first data acquisition unit 131 is configured to respectively obtain first feature data corresponding to the sample powder spreading image and the sample time series data through the convolution module;

[0178] The second data acquisition unit 132 is configured to perform feature fusion processing on the first feature data to obtain second feature data;

[0179] The model training unit 133 is configured to use the second feature data as input and the corresponding sample defect data as output to train a defect detection model.

[0180] In an implementable solution, the first data acquisition unit 131 includes:

[0181] The third data acquisition subunit 1311 is configured to respectively obtain third feature data corresponding to the sample powder spreading image and the sample time series data through the convolution module;

[0182] The first data acquisition subunit 1312 is configured to perform feature interaction processing on the third feature data by using a feature interaction mechanism to obtain first feature data.

[0183] In an implementable solution, the second data acquisition unit 132 includes:

[0184] The weight acquisition subunit 1321 is configured to respectively obtain weights corresponding to each first feature data by using a channel attention mechanism;

[0185] The second data acquisition subunit 1322 is configured to perform feature fusion processing on the first feature data based on the weights to obtain second feature data.

[0186] In an implementable solution, the sample time series data includes at least one of laser power data and molten pool intensity voltage data.

[0187] In this embodiment, by combining the sample powder spreading image and the corresponding sample time series data, a defect detection model is trained, which ensures the effectiveness and reliability of the defect detection model.

[0188] Embodiment 6

[0189] This embodiment provides a defect detection system, as Figure 13 shown, the defect detection system includes:

[0190] The actual image acquisition module 21 is configured to acquire the actual powder spreading image of the powder bed during the additive manufacturing process;

[0191] The time data acquisition module 22 is configured to acquire the actual time series data corresponding to the actual powder spreading image during the laser melting process;

[0192] The model output module 23 is configured to input the actual powder spreading image and the actual time series data into the defect detection model to output actual defect data;

[0193] Among them, the defect detection model is obtained based on the defect detection model training system of Embodiment 4 or Embodiment 5.

[0194] In this embodiment, by inputting the actual powder spreading image and the actual time series data into the defect detection model to output actual defect data, real-time anomaly detection in the additive manufacturing process is realized, defects can be discovered and processed in a timely manner, thereby reducing production losses, reducing the false alarm rate of defect detection, making full use of the complementarity between different modalities, significantly improving the accuracy and reliability of defect detection, reducing the later detection cost, improving production efficiency, and promoting the sustainable development of the additive manufacturing industry.

[0195] For the system embodiment, since it basically corresponds to the method embodiment, the relevant parts can refer to the partial description of the method embodiment. The system embodiments described above are merely illustrative. The units described as separate components may or may not be physically separated, and the components as units may or may not be physical units, that is, they may be located in one place or distributed to multiple network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the present disclosure solution.

[0196] Embodiment 7

[0197] Figure 14 As shown in the structural schematic diagram of an electronic device shown in an exemplary embodiment of the present disclosure, the electronic device includes a memory, a processor, and a computer program stored in the memory and configured to run on the processor. When the processor executes the computer program, it implements the defect detection model training method or the defect detection method described in any of the above embodiments. Figure 14 The displayed electronic device 90 is merely an example and should not impose any limitation on the functions and usage scope of the embodiments of the present disclosure.

[0198] As Figure 14 shown, the electronic device 90 can be presented in the form of a general computing device. For example, it can be a server device. The components of the electronic device 90 may include, but are not limited to: at least one of the above processors 91, at least one of the above memories 92, and a bus 93 connecting different system components (including the memory 92 and the processor 91).

[0199] The bus 93 includes a data bus, an address bus, and a control bus.

[0200] The memory 92 may include volatile memory, such as random access memory (RAM) 921 and / or cache memory 922, and may further include read-only memory (ROM) 923.

[0201] The memory 92 may also include program utilities 925 (or utilities) having a set (at least one) of program modules 924. Such program modules 924 include, but are not limited to: an operating system, one or more application programs, other program modules, and program data. Each or some combination of these examples may include an implementation of a network environment.

[0202] The processor 91 executes various functional applications and data processing by running computer programs stored in the memory 92, such as the training method or defect detection method of the defect detection model provided in any of the above embodiments.

[0203] The electronic device 90 may also communicate with one or more external devices 94 (such as a keyboard, pointing device, etc.). Such communication may be carried out through the input / output (I / O) interface 95. Moreover, the electronic device 90 may also communicate with one or more networks (such as a local area network (LAN), a wide area network (WAN), and / or a public network, such as the Internet) through the network adapter 96. As shown in the figure, the network adapter 96 communicates with other modules of the electronic device 90 through the bus 93. It should be understood that although not shown in the figure, other hardware and / or software modules may be used in conjunction with the electronic device 90, including but not limited to: microcode, device drivers, redundant processors, external disk drive arrays, RAID (disk array) systems, tape drives, and data backup storage systems, etc.

[0204] It should be noted that although several units / modules or sub-units / modules of the electronic device are mentioned in the above detailed description, this division is merely exemplary and not mandatory. In fact, according to the embodiments of the present disclosure, the features and functions of two or more of the above-described units / modules may be embodied in one unit / module. Conversely, the features and functions of one unit / module described above may be further divided and embodied by multiple units / modules.

[0205] Embodiment 8

[0206] The embodiments of the present disclosure also provide a computer-readable storage medium, on which a computer program is stored, and when the program is executed by a processor, the training method or defect detection method of the defect detection model provided in any of the above embodiments is implemented.

[0207] Among them, the readable storage medium can more specifically include, but is not limited to: portable disks, hard disks, random access memories, read-only memories, erasable programmable read-only memories, optical storage devices, magnetic storage devices, or any suitable combination of the above.

[0208] Embodiment 9

[0209] An embodiment of the present disclosure also provides a computer program product, including a computer program, where the computer program, when executed by a processor, implements the training method or defect detection method of the defect detection model described in any one of the above.

[0210] Among them, the program code for executing the computer program product of the present disclosure can be written in any combination of one or more programming languages, and the program code can be executed entirely on the user device, partially on the user device, executed as an independent software package, partially on the user device and partially on a remote device, or entirely on a remote device.

[0211] Although the specific implementation manners of the present disclosure have been described above, those skilled in the art should understand that this is only an example, and the protection scope of the present disclosure is defined by the appended claims. Without departing from the principles and essence of the present disclosure, those skilled in the art can make various changes or modifications to these implementation manners, but these changes and modifications all fall within the protection scope of the present disclosure.

Claims

1. A training method for a defect detection model, characterized in that, The training method includes: Obtaining a sample powder spreading image of a powder bed during an additive manufacturing process and corresponding sample defect data; Obtaining sample time series data corresponding to the sample powder spreading image during a laser melting process; Using the sample powder spreading image and the sample time series data as inputs, and the corresponding sample defect data as outputs, to train and obtain the defect detection model.

2. The training method of the defect detection model according to claim 1, characterized in that, The step of using the sample powder spreading image and the sample time series data as inputs, and the corresponding sample defect data as outputs, to train and obtain the defect detection model includes: Respectively obtaining first feature data corresponding to the sample powder spreading image and the sample time series data through a convolution module; Performing feature fusion processing on the first feature data to obtain second feature data; Using the second feature data as an input, and the corresponding sample defect data as an output, to train and obtain the defect detection model.

3. The training method of the defect detection model according to claim 2, characterized in that, The step of respectively obtaining first feature data corresponding to the sample powder spreading image and the sample time series data through the convolution module includes: Respectively obtaining third feature data corresponding to the sample powder spreading image and the sample time series data through the convolution module; Performing feature interaction processing on the third feature data by using a feature interaction mechanism to obtain the first feature data.

4. The training method of the defect detection model according to claim 2 or 3, characterized in that The step of performing feature fusion processing on the first feature data to obtain second feature data includes: Respectively obtaining weights corresponding to each of the first feature data by using a channel attention mechanism; Based on the weights, performing feature fusion processing on the first feature data to obtain the second feature data; And / or The sample time series data includes at least one of laser power data and molten pool intensity voltage data.

5. A defect detection method, characterized in that, The defect detection method includes: Obtaining an actual powder spreading image of a powder bed during an additive manufacturing process; Obtaining actual time series data corresponding to the actual powder spreading image during a laser melting process; Inputting the actual powder spreading image and the actual time series data into the defect detection model to output actual defect data; Wherein, the defect detection model is obtained based on the training method of the defect detection model according to any one of claims 1-4.

6. A training system for a defect detection model, characterized in that, The training system includes: A first data acquisition module for obtaining a sample powder spreading image of a powder bed during an additive manufacturing process and corresponding sample defect data; A second data acquisition module for obtaining sample time series data corresponding to the sample powder spreading image during a laser melting process; A model training module for using the sample powder spreading image and the sample time series data as inputs, and the corresponding sample defect data as outputs, to train and obtain the defect detection model.

7. A defect detection system, characterized in that, The defect detection system includes: An actual image acquisition module for obtaining an actual powder spreading image of a powder bed during an additive manufacturing process; A time data acquisition module for obtaining actual time series data corresponding to the actual powder spreading image during a laser melting process; A model output module for inputting the actual powder spreading image and the actual time series data into the defect detection model to output actual defect data; Among them, the defect detection model is obtained based on the training system of the defect detection model described in claim 6.

8. An electronic device, comprising a memory, a processor, and a computer program stored in the memory and adapted to run on the processor, characterized in that, When the processor executes the computer program, it implements the training method of the defect detection model described in any one of claims 1 to 4, or implements the defect detection method described in claim 5.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the training method of the defect detection model described in any one of claims 1 to 4, or implements the defect detection method described in claim 5.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the training method of the defect detection model described in any one of claims 1 to 4, or implements the defect detection method described in claim 5.

Images