DETAILED ACTION
This action is responsive to Applicant’s reply filed May 8th 2026. This action is made final.
Status of the Claims
Claims 1, 8 and 15 are amended.
Claims 4 and 11 are canceled.
Claims 16-21 are added.
Claim status is currently pending and under examination for claims 1-3, 8-10 and 15-21 of which independent claims are 1, 8 and 15.
Notice of Pre-AIA or AIA Status
The present application, filed on or after March 16, 2013, is being examined under the first inventor to file provisions of the AIA .
Response to Amendment
Applicant’s arguments regarding the art rejections are moot in view of the new grounds of rejection necessitated by applicant’s amendment.
Claim Rejections - 35 USC § 103
In the event the determination of the status of the application as subject to AIA 35 U.S.C. 102 and 103 (or as subject to pre-AIA 35 U.S.C. 102 and 103) is incorrect, any correction of the statutory basis (i.e., changing from AIA to pre-AIA ) for the rejection will not be considered a new ground of rejection if the prior art relied upon, and the rationale supporting the rejection, would be the same under either status.
The following is a quotation of 35 U.S.C. 103 which forms the basis for all obviousness rejections set forth in this Office action:
A patent for a claimed invention may not be obtained, notwithstanding that the claimed invention is not identically disclosed as set forth in section 102, if the differences between the claimed invention and the prior art are such that the claimed invention as a whole would have been obvious before the effective filing date of the claimed invention to a person having ordinary skill in the art to which the claimed invention pertains. Patentability shall not be negated by the manner in which the invention was made.
The following are the references relied upon in the rejections below:
Lin, Jinhua, Lin Ma, and Yu Yao. "Segmentation of casting defect regions for the extraction of microstructural properties." Engineering applications of artificial intelligence 85 (2019): 150-163.
Hattori (JP 2015090535 A)
Yu (CA 3149727 A1)
Claims 1-3, 8-10, 15-16 and 18-20 are rejected under 35 U.S.C. 103 as being unpatentable over Lin in view of Hattori, further in view of Yu.
Regarding Claim 1, Lin teaches:
A prediction system implemented on a computing apparatus and configured to train a model for predicting a defect of a target product ((P. 161, Sec. 7, ¶1) “A 3D CNN is proposed for segmenting casting defect regions with high precision”
(P. 155, Sec. 6, ¶1) “We provide 200 labels for each metallographic image manually, for a total of 1.6 million labels for training and inference of crdCNN”
(P. 152, Sec. 3.2, ¶1) “three-dimensional ISN is a conceptually useful and extensible framework for 3D CDR segmentation. … Three-dimensional ISN is a fully convolutional neural network … 3D ISN can also predict the class label and perform 3D bounding box regression simultaneously”
A computer (computing apparatus) is implied by training a CNN for segmenting casting defect regions (CDRs) and predicting a class label (predicting a defect of a casting).),
the prediction system comprising: processing circuitry configured to (A computer is implied by training a CNN, which further implies processing circuitry.):
receive, from a storage device, information on an existing product including: a defect characteristic value indicating a defect associated with a location in the existing product ((P. 150, Sec. 1) “2D metallographic images of castings are sampled via scanning electron microscope and the classes of CDRs are labeled on 2D metallographic images manually.”
(P. 155, Sec. 6, ¶1) “We obtain 400 specimens from each sampled casting, for a total of 400 ∗ 20 specimens. By scanning specimens with a scanning electron microscope, 8000 pieces of 2D metallographic images were obtained for training and testing of cdr CNN. The thicknesses of the specimens range from 0.05 cm to 0.15 cm. We provide 200 labels for each metallographic image manually, for a total of 1.6 million labels for training and inference of crdCNN. … The EM-30plus SEM is used to observe the microstructures of the castings, the observation range is ±0.5cm, and the 2D metallographic images are sampled simultaneously”
(P. 154, Sec. 5, ¶1) “Microstructural properties of CDR are important for evaluating a casting’s ability to resist external force. In this paper, four important microstructural properties are extracted to measure the micromechanical performance of CDR: porosity, hardness, tensile strength and elongation. (1) Porosity is the percentage of the pore volume of CDR relative to the total volume of CDR, which reflects the compactness of the casting material. The porosity of CDR is used to measure the shrinkage performance of castings. A higher porosity value of CDR not only affects the shrinkage size of castings but also causes casting defects such as internal stress, deformation and cracking.”
2D metallographic images are obtained by scanning specimens from castings and capture microstructures of castings (porosity). The 2D images are labeled with classes of casting defect regions (CDRs) and therefore the labels are a defect characteristic value indicating a defect associated with a location (region) in a casting (existing product). 2D metallographic images are obtained for training a crdCNN, therefore the images must be received from a storage device to perform training.),
shape information indicating a three-dimensional shape of the existing product ((P. 151, Sec. 3, ¶1) “3D RPN is proposed for learning 3D geometric features from 3D volume input. The 3D volume input consists of 3D CAD models of castings that are encoded by the truncated signed distance function (TSDF) with a grid size of 0.01 m. … Three-dimensional RPN takes the 3D volumes of castings as input and outputs multiscale 3D proposals.”),
and conditional information indicating a manufacturing condition of the existing product ((P. 155, Sec. 6, ¶1) “twenty casting models are sampled to evaluate the performance of cdrCNN.… The processing parameters of the iron casting models that were adopted during theOL”L:k;uiop[ experiment include the following: (1) melting temperature: 1400 °C and (2) initial sand mold temperature: 20 °C.”);
input the shape information to a shape-processing model implemented using a convolutional neural network ((P. 151, Sec. 3.1, ¶1) “a multiscale 3D region proposal network is proposed for learning 3D geometric features from 3D volume input. The learned geometric features include the casting size, the physical size of the receptive region, the 3D coordinates and the object location. Three-dimensional RPN takes 3D casting models as inputs and outputs multiscale 3D proposals. Furthermore, to output 3D bounding boxes of various sizes, 3D RPN targets three scales with three sizes of receptive regions. Three-dimensional RPN consists of convolutional layers, a rectified linear unit (ReLU) and a cls + reg layer.”),
the shape-processing model configured to perform a convolution operation on the three-dimensional shape and output a feature vector representing the shape of the existing product (A three-dimensional RPN (‘shape-processing model’) is a convolutional neural network that takes a 3D casting model (‘three-dimensional shape’) and outputs a multiscale 3D proposal (feature vector representing the shape of the existing product). 3D proposals are bounding boxes that capture geometric features including casting size, 3D coordinates and object location, therefore 3D proposals represent the shape of a casting.);
train a defect prediction model using the defect characteristic value, the feature vector output from the shape-processing model, and the manufacturing condition of the existing product ((P. 151, Sec. 3, ¶1) “cdrCNN, consists of three critical subnetworks: 3D RPN, ISN and 3D RoIAlign. … (2) ISN is proposed for generating 3D semantic instances for CDR. ISN takes 3D proposals as inputs and outputs 3D semantic instances of CDR. Taking 3D RPN and ORN (a joint object detection network that was proposed by Song and Xiao, 2014) as the baseline architecture, a new 3D FCN branch is added into the ISN pipeline. Three-dimensional volumes are used as inputs of our 3D backpropagation pipeline and 2D metallographic images are used as inputs of our 2D backpropagation pipeline; the class label and 3D bounding box regression are learned together in this joint pipeline. (3) 3D RoIAlign is proposed for eliminating the misalignment between the 3D proposals and the extracted features. Furthermore, cdrCNN learns the geometric features from 2D metallographic images and 3D CAD volumes jointly and yields three outputs for each CDR: a class label, a bounding box and a 3D semantic instance. This joint learning architecture can segment CDRs with complex and irregular microstructures.”
(P. 152, Sec. 3.2, ¶1) “3D ISN can also predict the class label and perform 3D bounding box regression simultaneously”
See Figure 9 on P. 157 depicting how a mold produce process is used to create castings used for training and inference of cdrCNN. See (P. 155, Sec. 6, ¶1) describing processing parameters (‘manufacturing conditions’) of iron casting models are melting temperature and initial sand mold temperature.
A cdrCNN (‘defect prediction model’) consists of subnetworks 3D ISN and 3D RoIAlign. cdrCNN takes 3D proposals (feature vector output from the shape-processing model) and 2D metallographic images (defect characteristic value) to jointly learn to predict a class label for a casting defect region.),
the defect prediction model including a machine learning model configured to associate spatial shape features and manufacturing parameters with location-specific defect predictions ((P. 159-160, Sec. 6.1, Last Paragraph) “the ISN subnetwork can extract features in each layer and visualize 16-level porosity values for each CDR.”
(P. 153, Figure 3 Caption) “Architecture of 3D ISN. Numbers denote the sizes of sliding windows, spatial resolution and channels. All convolutional kernels are of size 3 ∗ 3 ∗ 3, except in the output layer. The transposed convolutional kernels are of size 2 ∗ 2 ∗ 2 with stride 2. The multi-sized sliding window can ensure the multiscale segmentation of CDR. A three-dimensional CDR instance is colored by 16-level RGB values, where each color level represents a porosity value of CDR. The porosity is an important microstructural property”
(P. 154, Sec. 5, ¶1) “Porosity is the percentage of the pore volume of CDR relative to the total volume of CDR, which reflects the compactness of the casting material. The porosity of CDR is used to measure the shrinkage performance of castings. A higher porosity value of CDR not only affects the shrinkage size of castings but also causes casting defects such as internal stress, deformation and cracking.”);
input a feature vector representing a three-dimensional shape of the target product to the trained defect prediction model ((P. 160, Sec. 6.2, ¶1) “We sample CDRs of various shapes and extract their microstructural properties to evaluate the robustness of cdrCNN. The sampled CDRs are bounded by red squares in Fig. 12 and labeled as C1–C5. These CDRs are detected and segmented by cdrCNN successfully”
See Figure 12 on P. 159 depicting sampled casting defect regions (C1-C5) are three-dimensional shapes of a casting used to evaluate the cdrCNN (defect prediction model).);
output, via the defect prediction model, a defect characteristic value indicating a defect associated with a location in the target product ((P. 158, Sec. 6.1, Last Paragraph) “cdrCNN can extract microstructural properties of CDR in 16 levels.”
(P. 155, Sec. 5, ¶1) “To obtain the porosity of each point inside of CDR, we construct the contour map of CDR, as shown in Fig. 7. Using a 2D plane that is parallel to the horizontal coordinate system to divide the 3D instance of CDR, the intersection plane between the 2D plane and the 3D instance is the contour map of CDR. That is, the contour map is the two-dimensional horizontal profile of CDR and the pixel values of the contour map correspond to the porosity values of the points of CDR instances. Taking sixteen RGB values to represent the 16-level porosity values of CDR, the contour map is visualized as a 2D RGB image. As the contour map of CDR shows, the central point of CDR has a higher porosity value and the external point of CDR has a lower porosity value.”
Lin discloses Figure 7 (reproduced below) on P. 156 depicting a contour map with 16 different colors to represent porosity values (‘defect characteristic value’) and their locations (points) on a three-dimensional casting (‘target product’).
PNG
media_image1.png
728
1021
media_image1.png
Greyscale
);
and display, via a display device, the defect characteristic value indicating the defect associated with the location in the target product (A contour map is generated with 16 RGB values to represent the 16-level porosity values of each casting defect region. A display device is implied by visualizing a contour map with 16 RGB values as a 2D RGB image.),
wherein, when the target product is a casting, … and wherein the feature vector includes … a casting volume, … and a thickness of the casting ((P. 151, Sec. 3, ¶1) “3D RPN is proposed for learning 3D geometric features from 3D volume input. The 3D volume input consists of 3D CAD models of castings”
(P. 155, Sec. 6, ¶1) “twenty casting models are sampled to evaluate the performance of cdrCNN. The sampling process is illustrated in Fig. 9. … We obtain 400 specimens from each sampled casting, for a total of 400 ∗ 20 specimens. By scanning specimens with a scanning electron microscope, 8000 pieces of 2D metallographic images were obtained for training and testing of cdr CNN. The thicknesses of the specimens range from 0.05 cm to 0.15 cm.”).
However, Lin does not teach manufacturing conditions that include all of internal cooling temperature, die opening sequence, and air blow sequence, which is taught by Hattori:
wherein, when the target product is a casting, the manufacturing condition includes all of internal cooling temperature, die opening sequence, and air blow sequence (The Examiner interprets “internal cooling temperature” according to its broadest reasonable interpretation (BRI) in view of the applicant’s specification as encompassing temperature of circulating water. This interpretation is consistent with the descriptions in the Applicant’s specification at [0024], (see excerpt below).
Applicant’s written description at [0024] “The internal cooling temperature is the temperature of water that passes through the inside of a die for cooling a casting.”
The Examiner interprets “die opening sequence” according to its BRI in view of the applicant’s specification as encompassing acceleration when opening a mold. This interpretation is consistent with the descriptions in the Applicant’s specification at [0026], (see excerpt below).
Applicant’s written description at [0026] “The die opening sequence is the sequence of opening a die made up of a plurality of die components.”
(P. 2, Last Paragraph) “When managing the resin molding process, manufacturing condition variables related to the quality of the molded product, such as the temperature of the mold and cooling water, the air pressure that drives the mold, the acceleration when opening and closing the mold, the flow rate of the molten resin and the resin”
(P. 2, ¶5) “Further, among the variables for managing the resin injection molding machine, the variables directly related to the quality of the product are shown in FIG. 5 of Non-Patent Document 1. The pressure, for example, the air pressure for moving the mold, the acceleration, for example, when the mold moves It is also necessary to manage the acceleration of the water and the temperature of the circulating water, for example, the temperature of the circulating water that cools the mold”
The temperature of circulating water is the temperature of water that cools a mold, therefore the temperature of circulating water is ‘internal cooling temperature’. The acceleration when opening a mold represents how fast a mold is opened after a product is molded, therefore acceleration when opening a mold is a die opening sequence.),
Before the effective filing date of the claimed invention, it would have been obvious to a person of ordinary skill in the art to modify the method of Lin with the manufacturing condition variables disclosed by Hattori to train a machine learning model with manufacturing condition variables. By training a machine learning model with manufacturing condition variables, the model can learn to associate casting quality with manufacturing condition variables, thereby allowing manufacturers to predict defects and reduce the likelihood of producing defective castings.
Furthermore, the combined method of Lin/Hattori does not teach a feature vector includes a die volume of a casting and a casting surface area, which is taught by Yu:
and wherein the feature vector includes a die volume of the casting, a casting volume, a casting surface area, and a thickness of the casting ((P. 8, Lines 5-8) “the shape information may include at least one of a total volume of the product manufactured through the mold, the volume of cavities of the mold, the number of cavities, the number of gates of the mold, a surface area of the product … a maximum thickness of the product”
(Abstract) “a standard data extraction unit for extracting target standard data of a product produced by a mold from mold information about the mold to which a first molding material in a molten state is supplied”
(P. 8, Lines 17-22) “the standard data extraction unit 210 may generate mold information by scanning the mold for manufacturing the product and extract shape information of the product from the mold information. … the standard data extraction unit 210 may generate mold information by receiving a mold drawing of the product and extract shape information of the product from the mold information.”
(P. 11, Lines 6-12) “the determination unit 250 compares the manufacturing standard data of the product, which is manufactured under the molding condition output by the molding condition output unit 220, with the target standard data, which is extracted by the standard data extraction unit 210, and determines whether the molding condition is appropriate. Specifically, when the manufacturing standard data deviates from a predetermined reference range from the target standard data, the determination unit 250 determines that the molding condition is inappropriate”
(P. 11, Lines 20-22) “When the molding condition is determined to be inappropriate, the determination unit 250 transmits a stop command to the injection molding machine 100. Accordingly, the injection molding machine 100 stops producing the product.”
Total volume of a product manufactured (‘a casting volume’), volume of cavities of a mold (‘a die volume of the casting’), a surface area of the product, and maximum thickness of the product are shape information extracted from a product (casting). The shape information is compared to predetermined reference ranges to determine if the product should stop being produced.).
Before the effective filing date of the claimed invention, it would have been obvious to a person of ordinary skill in the art to modify the combined method of Lin/Hattori with the shape information extraction method disclosed by Yu to extract die volume and surface area of a casting. By extracting die volume and surface area of a casting, the extracted shape information can be used to determine if current molding conditions are not standard, thereby casting production can be stopped early to prevent defects and save materials.
Regarding Claims 2 and 9, the combined method of Lin/Hattori/Yu teaches:
the prediction system according to claim 1, wherein: the target product is a casting ((P. 160, Sec. 6.2, ¶1) “We sample CDRs of various shapes and extract their microstructural properties to evaluate the robustness of cdrCNN. The sampled CDRs are bounded by red squares in Fig. 12 and labeled as C1–C5. These CDRs are detected and segmented by cdrCNN successfully”
See Figure 12 on P. 159 depicting sampled casting defect regions (C1-C5) are three-dimensional shapes of a casting (‘target product’) used to evaluate the cdrCNN (defect prediction model).);
and the defect of the target product, indicated by the defect characteristic value, includes at least one of seizure, shrinkage, flow line, galling, raw material deformation, die cracking, or entrapment of the target product ((P. 158, Sec. 6.1, Last Paragraph) “cdrCNN can extract microstructural properties of CDR in 16 levels.”
(P. 154, Sec. 5, ¶1) “Porosity is the percentage of the pore volume of CDR relative to the total volume of CDR, which reflects the compactness of the casting material. The porosity of CDR is used to measure the shrinkage performance of castings. A higher porosity value of CDR not only affects the shrinkage size of castings but also causes casting defects such as internal stress, deformation and cracking.”).
Regarding Claims 3 and 10, the combined method of Lin/Hattori/Yu teaches:
the prediction system according to claim 1, wherein the defect characteristic value includes a value indicating a degree of the defect of the target product ((P. 155, Sec. 5, ¶1) “To obtain the porosity of each point inside of CDR, we construct the contour map of CDR, as shown in Fig. 7. … the contour map is the two-dimensional horizontal profile of CDR and the pixel values of the contour map correspond to the porosity values of the points of CDR instances. Taking sixteen RGB values to represent the 16-level porosity values of CDR, the contour map is visualized as a 2D RGB image. As the contour map of CDR shows, the central point of CDR has a higher porosity value and the external point of CDR has a lower porosity value.”
Lin discloses Figure 7 (reproduced above) on P. 156 depicting a contour map with 16 different colors to represent porosity values (‘defect characteristic value’) and their locations (points) on a three-dimensional casting (‘target product’). Each level indicates a different porosity value (degree of defect), with higher porosity values indicating shrinkage and other casting defects (see P. 154, Sec. 5, ¶1).).
Regarding Claim 8, the rejection of claim 1 is incorporated. The difference in scope being:
A prediction method implemented by a computing apparatus ((P. 155, Sec. 6, ¶1) “We provide 200 labels for each metallographic image manually, for a total of 1.6 million labels for training and inference of crdCNN”
A computer (computing apparatus) is implied by training a model to perform inference.).
Regarding Claim 15, the rejection of claim 1 is incorporated. The difference in scope being:
a non-transitory computer-readable storage medium storing instructions that, when executed by a processor, cause the processor to ((P. 155, Sec. 6, ¶1) “We provide 200 labels for each metallographic image manually, for a total of 1.6 million labels for training and inference of crdCNN”
A computer is implied by training a model, which further implies a non-transitory computer-readable storage medium storing instructions to be executed by a processor.).
Regarding Claim 16, the combined method of Lin/Hattori/Yu teaches:
The prediction system according to claim 1, wherein the processing circuitry is further configured to separate the information on the existing product into the defect characteristic value of the existing product and the shape information of the existing product ((P. 153, Sec. 3.3, ¶1) “cdrCNN employs 3D RPN as a 3D pipeline for extracting 3D geometric features from casting models; however, the pipeline cannot extract 2D features. To address this, taking the ORN pipeline as the baseline framework, cdrCNN constructs a joint pipeline for extracting 2D and 3D features simultaneously (see Fig. 1). cdrCNN employs the residual network (ResNet) (He et al., 2016) and feature pyramid network (FPN) (Lin et al., 2016) as a 2D baseline framework for extracting 2D features from metallographic images”
See (P. 150, Sec. 1) describing classes of casting defect regions (defect characteristic values) are labeled on 2D metallographic images of castings.
(P. 151, Sec. 3, ¶1) “cdrCNN, consists of three critical subnetworks: 3D RPN, ISN and 3D RoIAlign. … Three-dimensional RPN takes the 3D volumes of castings as input and outputs multiscale 3D proposals. … ISN takes 3D proposals as inputs and outputs 3D semantic instances of CDR. … cdrCNN learns the geometric features from 2D metallographic images and 3D CAD volumes jointly”
A cdrCNN (defect prediction model) consists of subnetworks 3D ISN and 3D RPN. 3D RPN takes 3D volumes of castings (‘shape information’) as input and outputs 3D proposals used by 3D ISN. 2D metallographic images (defect characteristic value) are input to a residual network ResNet for learning. Geometric features of the 2D images and the 3D volumes are jointly learned on their respective models, therefore separating existing product information into defect characteristic value of the existing product (2D image classes) and 3D volumes of castings (‘shape information of the existing product’).),
and the machine learning model is trained based on the separated information (See Figure 1 on P. 152 depicting the architecture of cdrCNN (defect prediction model) includes a ResNet model to learn 2D metallographic images and a 3D RPN model to learn a 3D volume of casting.).
Regarding Claim 18, the combined method of Lin/Hattori/Yu teaches:
The prediction system according to claim 1, wherein the processing circuitry is further configured to cause the defect prediction model to construct a regression expression based on an association among the defect characteristic value, the feature vector, and the manufacturing condition of the existing product ((P. 153, Sec. 4, ¶1) “cdrCNN is designed to segment CDR, while a nonlinear topological dimension (ntd) is proposed in this section for characterizing the geometric features of the segmented CDR. Taking ntd as the key feature parameter, we can extract the microstructural properties from the segmented CDR precisely”
(P. 154, Sec. 4, Last Paragraph) “the nonlinear topological dimension is determined by the logarithmic ratio of B to 𝛾. When cubes with different values of 𝛾 are used to cover the segmented CDR (𝐷), the number of cubes (B) will also change. Therefore, the least-square method is used to establish the empirical formula between B and 𝛾 in a logarithmic coordinate system. … With respect to the shrinkage defects with complex geometric features, Fig. 6 shows the approximation curves of the nonlinear topological dimension.”
(P. 154, Sec. 5, ¶2) “porosity refers to the percentage of the pore volume of CDR relative to the total volume of CDR. It reflects the compactness of the casting material. … However, the geometric features of CDR are complex and irregular and it is almost impossible to determine the quality in advance via simulation computation. To address this, we employ ntd as a critical feature parameter to compute the solid volume of CDR. Then, the porosity is computed as the percentage of the pore volume relative to the total volume of CDR. Formula (6) is used to obtain ntd from each segmented CDR.”
See Figure 9 on P. 157 depicting how a mold produce process is used to create castings used for training and inference of cdrCNN. See (P. 155, Sec. 6, ¶1) describing processing parameters (‘manufacturing conditions’) of iron casting models are melting temperature and initial sand mold temperature.
Nonlinear topological dimension (ntd) is used to extract microstructural properties from CDRs segmented by cdrCNN (defect prediction model). To calculate ntd, a least-square method is used. To use least-squares a regression expression must be used, therefore, ntd is a regression expression. To calculate porosity of a casting, ntd is used.
Porosity is tested using type C sand castings (see P. 155, Sec. 5, ¶2) and manufacturing conditions are initial sand model temperature, therefore porosity and sand mold temperature (‘manufacturing condition of the existing product’) are associated. 2D metallographic images are labeled with classes of casting defect regions and depict microstructures (porosity) of castings, (see P. 155, Sec. 6, ¶1), therefore labels of 2D images (defect characteristic value) are associated with porosity. cdrCNN takes 3D proposals (feature vector output) and 2D metallographic images to jointly learn to predict a class label (porosity) for a casting defect region, (see P. 151, Sec. 3, ¶1), therefore 3d proposals are associated with porosity.
ntd is used to calculate porosity, therefore cdrCNN constructs a regression expression based on an association among the defect characteristic value, the feature vector, and the manufacturing condition of the existing product.).
Regarding Claim 19, the combined method of Lin/Hattori/Yu teaches:
The prediction system according to claim 1, wherein the defect characteristic value indicating the defect associated with the location in the target product is displayed location by location using color indication, and a degree of the defect is expressed by a type of color ((P. 155, Sec. 5, ¶1) “To obtain the porosity of each point inside of CDR, we construct the contour map of CDR, as shown in Fig. 7. … the contour map is the two-dimensional horizontal profile of CDR and the pixel values of the contour map correspond to the porosity values of the points of CDR instances. Taking sixteen RGB values to represent the 16-level porosity values of CDR, the contour map is visualized as a 2D RGB image. As the contour map of CDR shows, the central point of CDR has a higher porosity value and the external point of CDR has a lower porosity value.”
Lin discloses Figure 7 (reproduced above) on P. 156 depicting a contour map with 16 different colors to represent porosity values (‘defect characteristic value’) and their locations (points) on a three-dimensional casting (‘target product’). Each level (color) indicates a different porosity value (degree of defect), with higher porosity values indicating shrinkage and other casting defects (see P. 154, Sec. 5, ¶1).).
Regarding Claim 20, the combined method of Lin/Hattori/Yu teaches:
The prediction system according to claim 1, wherein the shape-processing model is configured to express the feature vector representing the shape of the existing product as a feature vector generated by convolution of the three-dimensional shape of the existing product ((P. 151, Sec. 3.1, ¶1) “a multiscale 3D region proposal network is proposed for learning 3D geometric features from 3D volume input. The learned geometric features include the casting size, the physical size of the receptive region, the 3D coordinates and the object location. Three-dimensional RPN takes 3D casting models as inputs and outputs multiscale 3D proposals. Furthermore, to output 3D bounding boxes of various sizes, 3D RPN targets three scales with three sizes of receptive regions. Three-dimensional RPN consists of convolutional layers, a rectified linear unit (ReLU) and a cls + reg layer.”
A three-dimensional RPN (‘shape-processing model’) is a convolutional neural network that takes a 3D casting model (‘three-dimensional shape of the existing product’) and outputs a multiscale 3D proposal (‘feature vector representing the shape of the existing product’). 3D proposals are bounding boxes that capture geometric features including casting size, 3D coordinates and object location, therefore 3D proposals represent the shape of a casting.).
The following are the references relied upon in the rejections below:
Wang (CN 107577850 A)
Claims 17 and 21 are rejected under 35 U.S.C. 103 as being unpatentable over Lin in view of Hattori, further in view of Yu and Wang.
Regarding Claim 17, the combined method of Lin/Hattori/Yu teaches The prediction system according to claim 1, however, the combination does not teach comparing defect characteristic values of a target product with analysis results obtained from a computer-aided engineering analysis, which is taught by Wang:
compare the defect characteristic values of the target product with the analysis results obtained from a computer-aided engineering (CAE) analysis to evaluate the predictive accuracy of the defect prediction model ((P. 2, Bullet 2) “ProCAST the software to simulate TC4 titanium alloy investment casting process, the casting simulation results and the X-ray detection result of collecting the related input on the basis of each position”
(P. 3, ¶4) “(a) collect simulation result data of the casting at each position and the corresponding actual shrinkage data set as the training sample of the BP network”
(P. 4, ¶5) “result can be seen from the simulation result error of hole shrinking position of TC4 titanium alloy casting obtained by the constructed BP neural network model predicting and the actual observed within 10%, which indicates that the invention successfully builds the TC4 titanium alloy casting shrinkage cavity defect prediction model by the BP neural network method, and can well predict the casting shrinkage hole position”
(P. 3, ¶7) “the difference between actual output and expected output of the input layer and the hidden layer BP network according to the error function”
Actual observed values are obtained from performing a simulation using ProCAST software. A hole shrinking position (‘defect characteristic value’) is obtained from the simulation, therefore hole shrinking positions from the simulation are ‘analysis results obtained from a CAE analysis’. A BP neural network (‘defect prediction model’) is trained to predict a hole shrinking position of a TC4 titanium alloy casting. An error is the difference between an actual observed value (analysis result) and a predicted hole shrinking position (‘defect characteristic values of the target product’). A BP neural network is considered to be trained when the error (difference) is within 10%, therefore evaluating the predictive accuracy of the defect prediction model.).
Before the effective filing date of the claimed invention, it would have been obvious to a person of ordinary skill in the art to modify the combined method of Lin/Hattori/Yu with the technique disclosed by Wang to compare predictions of a defect prediction model against CAE analysis results. By comparing predictions of a defect prediction model against CAE analysis results, CAE can be used to simulate rare defects and a defect prediction model can learn from these simulations, thereby improving defect prediction model accuracy.
Regarding Claim 21, the combined method of Lin/Hattori/Yu teaches:
The prediction system according to claim 1, wherein the processing circuitry is further configured to, however the combination does not teach repeatedly training a defect prediction model and shape-processing model, which is taught by Wang:
repeatedly train the defect prediction model and the shape-processing model until a difference between the defect characteristic value output from the defect prediction model and a defect characteristic value obtained by simulation is less than or equal to a prescribed value ((P. 2, ¶4) “shrinkage easily occurs at the thick part of the casting and the transfer position of the casting, generally irregular shape, and also has a great difference in size, generally spherical or more spherical shape. the presence of shrinkage greatly reduces the comprehensive mechanical property of the cast and influence the using quality of the casting. Therefore, the location of the position of the shrinkage hole in the casting is particularly important”
(P. 3, ¶4) “(a) collect simulation result data of the casting at each position and the corresponding actual shrinkage data set as the training sample of the BP network”
(P. 4, ¶5) “result can be seen from the simulation result error of hole shrinking position of TC4 titanium alloy casting obtained by the constructed BP neural network model predicting and the actual observed within 10%, which indicates that the invention successfully builds the TC4 titanium alloy casting shrinkage cavity defect prediction model by the BP neural network method, and can well predict the casting shrinkage hole position”
(P. 4, ¶1) “the training process of the BP neural network is to continuously modify the weight value and the threshold value, the error E is as small as possible so as to reach the required process of BP neural network training process is the process of modifying the weight coefficient according to the negative gradient direction of the error function”
(P. 3, ¶7) “the difference between actual output and expected output of the input layer and the hidden layer BP network according to the error function”
Actual observed values are obtained from performing a simulation. A hole shrinking position (‘defect characteristic value’) is obtained from simulation. A BP neural network (‘defect prediction model’) is trained to predict a hole shrinking position of a TC4 titanium alloy casting. Shrinkage holes occur on a thick part of a casting and their positions are irregularly shaped, therefore a BP neural network that predicts a position of a shrinkage hole must process the irregular shape of the position (and therefore the BP neural network is also a shape-processing model).
A BP neural network is trained until an error function is minimized. An error is the difference between an actual observed value (defect characteristic obtained by simulation) and a predicted hole shrinking position (defect characteristic value output from the defect prediction model). A BP neural network is considered to be trained when the error (difference) is within 10%.).
Before the effective filing date of the claimed invention, it would have been obvious to a person of ordinary skill in the art to modify the combined method of Lin/Hattori/Yu with the training technique disclosed by Wang to compare predictions of a defect prediction model against a simulation model. By comparing predictions of a defect prediction model against a simulation model, a simulation model can be used to simulate rare defects and a defect prediction model can learn from these simulations, thereby improving defect prediction model accuracy.
Conclusion
The prior art made of record and not relied upon is considered pertinent to applicant's disclosure.
Jia et al. (CN 111223088 A) teaches using a deep convolutional neural network to detect the location, type, and size of casting surface defects.
Applicant's amendment necessitated the new ground(s) of rejection presented in this Office action. Accordingly, THIS ACTION IS MADE FINAL. See MPEP § 706.07(a). Applicant is reminded of the extension of time policy as set forth in 37 CFR 1.136(a).
A shortened statutory period for reply to this final action is set to expire THREE MONTHS from the mailing date of this action. In the event a first reply is filed within TWO MONTHS of the mailing date of this final action and the advisory action is not mailed until after the end of the THREE-MONTH shortened statutory period, then the shortened statutory period will expire on the date the advisory action is mailed, and any nonprovisional extension fee (37 CFR 1.17(a)) pursuant to 37 CFR 1.136(a) will be calculated from the mailing date of the advisory action. In no event, however, will the statutory period for reply expire later than SIX MONTHS from the mailing date of this final action.
Any inquiry concerning this communication or earlier communications from the examiner should be directed to PEDRO J MORALES whose telephone number is (571)272-6106. The examiner can normally be reached 8:30 AM - 6:00 PM.
Examiner interviews are available via telephone, in-person, and video conferencing using a USPTO supplied web-based collaboration tool. To schedule an interview, applicant is encouraged to use the USPTO Automated Interview Request (AIR) at http://www.uspto.gov/interviewpractice.
If attempts to reach the examiner by telephone are unsuccessful, the examiner’s supervisor, MIRANDA M HUANG can be reached at (571)270-7092. The fax phone number for the organization where this application or proceeding is assigned is 571-273-8300.
Information regarding the status of published or unpublished applications may be obtained from Patent Center. Unpublished application information in Patent Center is available to registered users. To file and manage patent submissions in Patent Center, visit: https://patentcenter.uspto.gov. Visit https://www.uspto.gov/patents/apply/patent-center for more information about Patent Center and https://www.uspto.gov/patents/docx for information about filing in DOCX format. For additional questions, contact the Electronic Business Center (EBC) at 866-217-9197 (toll-free). If you would like assistance from a USPTO Customer Service Representative, call 800-786-9199 (IN USA OR CANADA) or 571-272-1000.
/PEDRO J MORALES/Examiner, Art Unit 2124
/MIRANDA M HUANG/Supervisory Patent Examiner, Art Unit 2124
Prosecution Insights