Dangerous Goods Detection Method, 3D Printer and Detection Device

Abstract

The present invention relates to a method for detecting dangerous goods, a 3D printer and a detection device. The detection method encodes the sliced images of the currently printed product of the 3D printing device by using a deep learning model to obtain feature vectors with three-dimensional information, and then compares the feature vectors of the model to be printed with those of the prohibited printing models in the dangerous goods library to identify the prohibited printing dangerous goods, and can intervene in the printing process in a timely manner during the product printing process to prevent the dangerous goods from being printed, so as to technically assist in realizing the control of dangerous goods.

Description

Technical Field

[0001] This application relates to the field of 3D printing technology, and in particular to a method for detecting dangerous goods and a 3D printer. Background Art

[0002] With 3D printing technology, people can conveniently and quickly obtain various complex three-dimensional structural parts and components. However, in recent years, there have been frequent reports overseas of using 3D printing technology to manufacture guns or gun parts. The problem to be solved by this application is to technically detect the products currently printed by a 3D printer to determine whether they are controlled dangerous goods. Summary of the Invention

[0003] The purpose of this application is to provide a method for a 3D printer to detect the currently printed item, which can timely detect dangerous goods and take appropriate measures.

[0004] To achieve the above purpose, this application adopts the following technical solution: A method for detecting dangerous goods, including the following steps,

[0005] a. Obtain the printed slice image of the model to be printed;

[0006] b. Use a deep learning model to encode the slice image to obtain a feature vector with three-dimensional information of the model to be printed;

[0007] c. Provide a dangerous goods library storing prohibited printed models, compare the feature vector of the model to be printed with the feature vectors of the prohibited printed models in the dangerous goods library, and determine whether the model to be printed belongs to the prohibited printed items;

[0008] d. If it is determined that it does not belong to the prohibited printed items, continue printing, and simultaneously loop through steps a - c until printing is completed;

[0009] e. If it is determined that it belongs to the prohibited printed items, execute an interference printing program to prevent the remaining part from being printed.

[0010] The detection method of this application is implanted in the 3D printing program. Before or during 3D printing, the slice image of the model to be printed is obtained, and the deep learning model is used to convert the slice image into a high-dimensional feature vector. Then, by comparing the similarity of the feature vector with the feature vectors of the prohibited printed items in the dangerous goods library, it is possible to timely detect whether the currently printed model belongs to the prohibited printed items and timely intervene in the printing program.

[0011] In a possible implementation manner, step a obtains the slice image during the printing process.

[0012] In a possible implementation, the deep learning model is an encoding-decoding model, which can encode the sliced images of the model to be printed to obtain a high-dimensional feature vector with a fixed length.

[0013] In a possible implementation, the deep learning model is trained through the following method:

[0014] (C11) Select a 3D model, slice the 3D model to obtain a number of sliced images for training, encode the number of sliced images to obtain a high-dimensional vector, and then perform multi-layer decoding on this high-dimensional vector to output a point cloud data model;

[0015] (C12) Randomly place the above 3D model in n poses. Each time it is randomly placed, re-slice it to obtain the sliced images for training once, and repeat step (C11) to obtain n identical or approximate point cloud data models. n is a natural number greater than 0. Then, randomly pair the sliced images obtained from the n transformations with the n point cloud data models, and train the neural network to learn the structural relationship between the sliced images and the point cloud models;

[0016] (C13) Select m 3D models, and repeat the training in the above steps (C11) and (C12). m is a natural number greater than 0. Ensure that the number of training samples m×n≥10000 to obtain the deep learning model.

[0017] In a possible implementation, in step (C11), each sliced image is scanned in the X and Y directions with mutually perpendicular scan lines to obtain the intersection points of the scan lines and the edges of the sliced images. These intersection points carry xy coordinate information, and then add the height position information of the current sliced image to obtain discrete points with spatial xyz coordinates, and generate the point cloud data model.

[0018] In a possible implementation, in step c, calculate the similarity between the feature vector of the model to be printed and the feature vectors of each prohibited printing model in the dangerous goods library; if the similarity exceeds the set threshold, execute step e, and if the similarity does not exceed the set threshold, execute step d.

[0019] In a possible implementation, the prohibited printing models in the dangerous goods library include product models and part models, and step c includes determining whether the model to be printed belongs to a product model or a part model.

[0020] In a possible implementation, the interference printing program includes: adjusting the printing parameters to cause printing failure or deformation of the printed model.

[0021] In a possible implementation, the interference printing program includes: ending the current printing model, turning off the printer power supply, or providing a fault signal to the 3D printer.

[0022] In a possible implementation, the interference printing program includes: sending an alarm message to the outside world.

[0023] The second technical solution provided by the present invention is: a 3D printer, including,

[0024] A memory storing executable program code; and,

[0025] A processor coupled to the memory;

[0026] The processor is capable of calling the executable program code stored in the memory and executing the hazardous substance detection method.

[0027] The present invention also provides a third technical solution: a detection device communicatively connected to the 3D printer, which includes,

[0028] A memory storing executable program code; and,

[0029] A processor coupled to the memory; the processor is capable of calling the executable program code stored in the memory and executing the hazardous substance detection method.

[0030] It can be understood that the 3D printer in the second solution provided above and the detection device in the third solution can both execute the detection method provided above. Therefore, the beneficial effects that can be achieved can refer to the beneficial effects in the hazardous substance detection method provided above, and will not be elaborated here. Description of the Drawings

[0031] Figure 1 It is a flowchart of a hazardous substance detection method provided by an embodiment of the present application. Detailed Embodiments

[0032] To describe in detail the technical content, structural features, achieved objectives and effects of the invention, the technical solutions in the embodiments of the present application will be described below in conjunction with the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all of the embodiments. In the following description, for the purpose of explanation, many specific details are set forth to provide a detailed description of various exemplary embodiments or implementations of the invention. However, various exemplary embodiments may also be implemented without these specific details or in the case of one or more equivalent arrangements. In addition, various exemplary embodiments may be different, but not necessarily exclusive. For example, without departing from the inventive concept, the specific shapes, structures and characteristics of the exemplary embodiments may be used or implemented in another exemplary embodiment. In the description of the present application, unless otherwise specified, the meaning of "a plurality" is two or more.

[0033] Figure 1 The flowchart structural diagram of the dangerous goods detection method provided by the embodiment of the present application is shown. This method can be implanted in a 3D printing device and a printing program, or can be burned into a separate detection device, and the dangerous goods detection and intervention printing work are realized by connecting the detection device with the 3D printing device.

[0034] Please refer to Figure 1 As shown, the dangerous goods detection method provided by the present application includes the following steps:

[0035] Step a: Obtain the printing slice image of the model to be printed.

[0036] In step a, the slice image can be obtained at the execution component (such as the optical machine) of the printer, or directly obtained in the computer control system of the printer. The purpose is to exclude the disguise and interference of the model to be printed and re-model the model to be printed. In a preferred embodiment, step a is to obtain the slice image during the printing process, grab discrete points in the slice image, and form a point cloud data model.

[0037] Step b: Use a deep learning model to encode the slice image to obtain a feature vector with three-dimensional information of the model to be printed. Specifically, the deep learning model is an encoding and decoding model, which encodes the slice image of the model to be printed to obtain a high-dimensional 3D feature vector with a fixed length. The high-dimensional vector can contain hundreds to thousands of dimension information, and "fixed length" means that all point cloud models of any size and scale are unified into a vector with equal length. This vector must be able to represent the morphological information of this model, so as to facilitate subsequent vector comparison.

[0038] Step c: Provide a dangerous goods library storing models prohibited from printing, compare the feature vector of the model to be printed with the feature vectors of the models prohibited from printing in the dangerous goods library, and determine whether the model to be printed belongs to prohibited printing items. The dangerous goods library includes a dangerous product library and a dangerous parts library.

[0039] Dangerous products are usually controlled implements and special articles, such as guns, controlled knives, police equipment, cartridges, etc. Dangerous parts are usually key parts of controlled implements. In step c, the feature vector of the model to be printed is extracted. First, the feature vector is brought into the dangerous product library for comparison to analyze whether it is a dangerous product. If it is, it is determined as a prohibited printing item and the corresponding interference printing program is executed; if it is not a dangerous product, the feature vector is brought into the dangerous parts library for comparison to analyze whether it is a dangerous part. If it is determined as a dangerous part, it is determined as a prohibited printing item and the corresponding interference printing program is executed. If it is not a dangerous part, the printing program is continued.

[0040] The dangerous parts library contains two types of part information:

[0041] (1) Dangerous part information

[0042] The dangerous part information and the dangerous product library are judged by safety experts and dynamically maintained through networking.

[0043] (2) Part information that looks dangerous but is safe

[0044] When the printer initially works, the number of such parts is 0. When a suspected dangerous part appears, the printer reports the 3D information through networking. It is decided by experts whether the part is a safe part. If it is a safe part, the part is recorded in the part information that looks dangerous but is safe, and the printer is allowed to print normally; if it is a dangerous part, it is added to the dangerous part information.

[0045] Taking a specific embodiment as an example, the first step of this method is to obtain the sliced image of the model to be printed. The second step is to use a deep learning model to extract the feature vector for the sliced image of the model to be printed, that is, to represent the model to be printed with a high-dimensional vector of a fixed length. The third step is to calculate the similarity between this high-dimensional vector and the feature vectors of all models prohibited from printing to determine whether it is a dangerous item.

[0046] In a specific embodiment given by the present invention, the implementation process of this method specifically includes the following steps. (1) Obtain the sliced image of the item to be printed.

[0047] (1) If the sliced image of the item to be printed can be directly read, then directly capture the sliced image.

[0048] (2) If the printer outputs a 3D model, slice it first and then obtain the sliced images. (II) Extract the model feature vectors of the item to be printed.

[0049] It should be noted here that even for the same 3D model, different placement methods and postures will result in completely different forms, projections, etc. of the sliced images. Therefore, directly comparing the original sliced images cannot obtain the correct comparison results.

[0050] Therefore, this application constructs an encoder-decoder model based on deep learning. The input of this deep learning model is the sliced image. By encoding several sliced images of a printing model through the encoder, an intermediate high-dimensional vector can be obtained. Then, this high-dimensional vector goes through multiple layers of decoding and can output a point cloud data model. Through multiple trainings, the neural network learns the structural relationship between the sliced image and the point cloud data model, and thus can obtain the feature vector with 3D information according to the sliced image.

[0051] Furthermore, the training method of this deep learning model is as follows:

[0052] First, prepare the training data.

[0053] We first establish a point cloud data model from the sliced images. Horizontally and vertically scan each sliced image to obtain the intersection points of the scan lines and the edges of the sliced image. These intersection points carry xy information, and then add the height position information of the current sliced image to become the spatial xyz points. Until all sliced images are scanned, a point cloud data model is obtained.

[0054] The training data of the deep learning model is very easy to obtain. Given a 3D training model, then place this training model arbitrarily, slice the training model, and obtain the sliced images. Each time it is randomly placed, the data from the training image to the point cloud model is obtained once. As long as we require that no matter how the deep learning network is placed and sliced, the finally obtained point cloud data model is the same or approximate, we can extract 3D information from the sliced images. Suppose we have 100 3D training models, and each training model is randomly placed 100 times, which is 10,000 training samples.

[0055] In the point cloud data model established from different sliced images, the data of each point may be different. Therefore, it is impossible to match by directly comparing the point cloud models established from the original point cloud data. However, when the same object is placed 100 times, the shape of the point cloud is approximately the same. Here, a simple alignment step is required: (a) Find the two points with the farthest distance as the first axis; (b) Then find the points on the plane orthogonal to this axis and find the two points with the farthest distance as the second axis; (c) These two axes determine a plane. Take this as the bottom surface, and then find the point with the highest distance from the bottom surface as the third axis. In this way, the alignment is completed. Theoretically, for the point cloud data models scanned from the same 3D model, they can all be aligned approximately in this way.

[0056] Then, randomly pair 100 sliced images of the same object with 100 aligned point cloud data models so that the neural network can learn the structural relationship between the sliced images and the point cloud models.

[0057] After the deep learning model is successfully trained, the intermediate high-dimensional vector theoretically can comprehensively contain the information of the training model, which is called the feature vector. In actual use, the encoder of the deep learning model can be used to convert the sliced images of the model to be printed into feature vectors.

[0058] The so-called "intermediate" means that a series of input two-dimensional sliced images are processed by the encoder and converted into an intermediate vector of K dimensions, and then converted into a three-dimensional point cloud data model by the decoder. In the process of using the deep learning model for printing, only the encoder needs to be used, that is, a series of two-dimensional sliced images are input and converted into an intermediate vector of K dimensions. This process can be regarded as converting a large number of two-dimensional sliced images into a one-dimensional array in the form of (0, 0.222, 0.182, 0.38,..., 0.293). Each 3D model corresponds to a series of two-dimensional sliced images and can be converted into a one-dimensional array of the same length. Since the length of the intermediate vector of any 3D model is the same, the comparison process of different 3D models becomes very simple.

[0059] (3) Compare the similarity between the feature vector of the model to be printed and the feature vector of the prohibited printing models in the dangerous goods library.

[0060] Through the feature vector extraction in the second step, any 3D model becomes a feature vector with a unified length, such as 1000 dimensions. "1000 dimensions in length" means that the deep learning model encodes the sliced images of each 3D model into a vector with a length of 1000 (also called the feature vector). This is a high-dimensional numerical vector, and each numerical value represents the feature or attribute of the sliced image in a specific aspect.

[0061] The deep learning model learns to extract useful features from images by learning a large amount of sliced image data and stores these features in feature vectors. Each sliced image is represented as a feature vector of length 1000, where the numerical value of each dimension represents the intensity or amplitude of the image on a specific feature. Each of these dimensions is positive, so the similarity between any two feature vectors can be calculated. There are many methods for calculating similarity, such as cosine distance, Euclidean distance, Mahalanobis distance, etc. After calculating a similarity number, then compare the numbers corresponding to the model to be printed and the prohibited printing model to determine whether they are similar.

[0062] When judging similarity, a threshold is used for judgment. Calculate the similarity between the feature vector of the model to be printed and the feature vectors of each prohibited printing model in the dangerous goods library; if the similarity exceeds the set threshold, execute step e, if the similarity does not exceed the set threshold, execute step d. This threshold is a given value, set artificially by the control unit or the software provider.

[0063] Step d: If it is determined that the current printing model does not belong to the prohibited printing items, continue printing, and loop steps a - c until the printing is completed.

[0064] Step e: If it is determined that the current printing model belongs to the prohibited printing items, execute the interference printing program to prevent the remaining part from being printed.

[0065] In this embodiment, the interference printing program achieves the goal of preventing the printing of dangerous goods by adjusting the printing parameters to cause printing failure or deformation of the printed model. Adjusting the printing parameters means outputting the normal printing parameters after distortion or offset, such as changing the pattern projected by the optical machine or changing the projection intensity, changing the layer thickness, changing the exposure time, etc. By changing the parameters, the printing material cannot meet the fixed forming conditions or the printed item is distorted and deformed, unable to meet the processing and use requirements of dangerous goods.

[0066] In another embodiment of the present application, the interference printing program prevents the printing of dangerous goods from being completed by forcibly ending the current printing model or turning off the printer power supply, or providing a fault signal to the 3D printer.

[0067] In yet another embodiment of the present application, the interference printing program includes: sending an alarm message to the outside world through the network and uploading the information of the model to be printed of the current printed product. This information can be reported to the management department and law enforcement department of dangerous goods, and can also issue an alarm in the form of sound, light, electricity, etc.

[0068] Common point cloud feature extraction algorithms include PFH (Point Feature Histogram), FPFH (Fast Point Feature Histogram), and SHOT (Signature of Histograms of Orientations) algorithms.

[0069] Common algorithms in deep learning include PointNet (Point Cloud Network), PointNet++, PointCNN (convolution on X-transformed points), and DGCNN (Dynamic Graph Convolutional Neural Network), etc.

[0070] In this application, by reprocessing and modeling the sliced images of the currently printed product of the 3D printing device, and by performing a similarity comparison between the model to be printed and the dangerous product library, dangerous items prohibited from printing are identified, and the printing process can be intervened in a timely manner during the product printing process to prevent the completion of printing of dangerous items, thereby technically assisting in the control of dangerous items.

[0071] In a feasible embodiment of this application, a 3D printer is provided, including:

[0072] A memory storing executable program code;

[0073] A processor coupled to the memory; the processor calls the executable program code stored in the memory to execute the dangerous item detection method.

[0074] In a feasible embodiment of this application, a detection device is provided, communicatively connected to the 3D printer, including:

[0075] A memory storing executable program code;

[0076] A processor coupled to the memory; the processor calls the executable program code stored in the memory to execute the dangerous item detection method.

[0077] The basic principles, main features and advantages of the present invention have been shown and described above. Those skilled in the art should understand that the present invention is not limited by the above embodiments, and what is described in the above embodiments and the specification only illustrates the principles of the present invention. Without departing from the spirit and scope of the present invention, the present invention will have various changes and improvements. The scope of protection claimed by the present invention is defined by the appended claims, the specification and their equivalents.

Claims

1. A dangerous goods detection method, characterized in that, It includes the following steps: Obtain the printing slice images of the model to be printed; Use a deep learning model to encode the slice images to obtain a feature vector with three-dimensional information of the model to be printed. The deep learning model is an encoding and decoding model that can encode the slice images of the model to be printed to obtain a high-dimensional feature vector with a fixed length; Provide a dangerous goods library storing models prohibited from printing, compare the feature vector of the model to be printed with the feature vectors of the models prohibited from printing in the dangerous goods library, and determine whether the model to be printed belongs to the items prohibited from printing; If it is determined that it does not belong to the items prohibited from printing, continue printing, and simultaneously loop through steps a - c until printing is completed; If it is determined that it belongs to the items prohibited from printing, execute an interference printing program to prevent the remaining part from being printed; Among them, the deep learning model is obtained through the following method: (C11), Select a three-dimensional model, slice the three-dimensional model to obtain several slice images for training, encode the several slice images to obtain a high-dimensional vector, and then perform multi-layer decoding on this high-dimensional vector to output a point cloud data model; (C12), Randomly place the above three-dimensional model in n different poses. Each time it is randomly placed, re-slice it to obtain the slice images for training once, and repeat step (C11) to obtain n identical or approximate point cloud data models. n is a natural number greater than 0. Then, randomly pair the slice images obtained from the n transformations with the n point cloud data models, and train the neural network to learn the structural relationship between the slice images and the point cloud models; (C13), Select m three-dimensional models, repeat the training in the above steps (C11) and (C12). m is a natural number greater than 0. Ensure that the number of training samples m×n≥10000 to obtain the deep learning model.

2. The method according to claim 1, characterized in that: Step a is to obtain the slice images during the printing process.

3. The method according to claim 1, wherein: In step (C11), scan each slice image in the X and Y directions with mutually perpendicular scan lines to obtain the intersection points of the scan lines and the edges of the slice images. These intersection points carry xy coordinate information, and then add the height position information of the current slice image to obtain discrete points with spatial xyz coordinates, and generate the point cloud data model.

4. The method according to claim 1, wherein In step c, calculate the similarity between the feature vector of the model to be printed and the feature vectors of each model prohibited from printing in the dangerous goods library; if the similarity exceeds the set threshold, execute step e, and if the similarity does not exceed the set threshold, execute step d.

5. The method according to claim 1, wherein The models prohibited from printing in the dangerous goods library include product models and part models. Step c includes determining whether the model to be printed belongs to a product model or a part model.

6. The method according to claim 1, characterized in that, The interference printing program includes: adjusting the printing parameters to cause printing failure or deformation of the printed model.

7. The method according to claim 1, characterized in that The interference printing program includes: ending the current printed model, turning off the printer power supply, or sending a fault signal to the 3D printer.

8. The method according to claim 1, wherein The interference printing program includes: sending an alarm message to the outside.

9. A 3D printer, characterized in that, It includes: A memory storing executable program code; And, A processor coupled to the memory; The processor is capable of invoking the executable program code stored in the memory and executing the method according to any one of claims 1-8.

10. A detection device is communicatively connected to a 3D printer, characterized in that, Comprising: A memory storing executable program code; And, A processor coupled to the memory; The processor is capable of invoking the executable program code stored in the memory and executing the method according to any one of claims 1-8.

Images