PSS process product detection method and system

Abstract

The invention provides a PSS process product detection method and system, and the method comprises the following steps: S1, collecting. FLT files of a PSS process product, and obtaining a product data set of the PSS process product; s2, constructing a classification and judgment model, training the classification and judgment model through the product data set, and judging a quality inspection result of the PSS process product by using the classification and judgment model; s3, obtaining a numerical calculation result of the product parameters according to the product data set; s4, summarizing numerical calculation results according to preset parameters, outputting judgment information, and jointly judging output parameters and quality inspection results according to the output judgment information and judgment results of the classification judgment model. According to the PSS process product detection method based on the inverted residual module and the numerical constraint provided by the invention, the quality and specification of the product are detected through model classification and numerical calculation constraint.

Description

Technical Field

[0001] The present invention relates to the technical field of PSS process product detection, and specifically, to a PSS process product detection method and system, and in particular to a PSS process product detection method and system based on self-attention deep separable convolution. Background Art

[0002] PSS (Patterned Sapphire Substrate) is a technology for making periodic patterns on sapphire substrates. The PSS process mainly includes using photolithography and etching technology to form patterns of different shapes on the sapphire surface, such as stripes, hemispheres, cones, etc., and then growing GaN materials on them to change the longitudinal epitaxy of GaN materials into lateral epitaxy. PSS technology can improve the brightness and efficiency of LEDs. PSS patterns can be from micron to nanometer, and etching technology can be from wet to dry.

[0003] The existing PSS process product detection method mainly uses target detection to identify defects by marking them. However, for small and fuzzy defects, target detection requires the extraction of feature maps. For small and fuzzy defects, feature maps are limited, resulting in poor recognition accuracy and inability to guarantee recognition accuracy. The present invention uses classification and numerical constraints for defect recognition. It classifies data, and uses self-attention deep separable convolution self-classification learning to discover the corresponding defect content, and combines numerical calculations to further constrain the recognition results. Summary of the invention

[0004] In view of the defects in the prior art, an object of the present invention is to provide a PSS process product detection method and system.

[0005] A PSS process product detection method provided by the present invention comprises the following steps:

[0006] Step S1: Collect the .FLT file of the PSS process product and obtain the product data set of the PSS process product;

[0007] Step S2: construct a classification and discrimination model, train the classification and discrimination model through the product data set, and use the classification and discrimination model to discriminate the quality inspection results of the PSS process products;

[0008] Step S3: obtaining numerical calculation results of product parameters according to the product data set;

[0009] Step S4: According to the preset parameters, the numerical calculation results are summarized and the discrimination information is output. According to the output discrimination information and the discrimination result of the classification discrimination model, the output parameters and the quality inspection results are jointly determined.

[0010] Preferably, the step S1 specifically includes the following steps:

[0011] Steps for obtaining the original file: The disk produced by the PSS process is inspected by atomic force microscopy to obtain the .FLT file;

[0012] Heat map acquisition steps: read the HEIGHT_SENSOR.FLT file in the .FLT file, decode it using ISO-8859-1 encoding, obtain the binary file inside, use its header file parameters to restore the data, and use data interpolation to resize the data to a 480*240 browsing data mode to obtain the heat map data;

[0013] RGB image acquisition steps: read the TM_AMPLITUDE.FLT file in the .FLT file, decode it using ISO-8859-1 encoding, obtain the binary file inside, use its header file parameters to restore the data, and use data interpolation to resize the data to a 480*240 browsing data mode to obtain the RGB image data;

[0014] 3D contour map acquisition steps: the obtained heat map data is mapped to the 3D field of view map according to the relationship between the corresponding values ​​of the pixel points, and the 3D contour map data is obtained;

[0015] Steps for obtaining the curve data graph: convert the obtained heat map data into grayscale values, use Hough transform to identify the circle, take the horizontal line formed by connecting the center of the circle and 160 pixel points in the horizontal plane where the circle is located, and display it visually to obtain the curve data graph data.

[0016] Preferably, the step S2 specifically includes the following steps:

[0017] Dataset preparation steps: Convert the .FLT data file into RGB image data, and divide the RGB image data into training data set, verification data set, and test data set according to the preset ratio;

[0018] Dataset enhancement steps: For RGB image data, the input image is resized to 480x240 pixels, data tensorization is performed, the diversity of training data is increased by flipping the image horizontally and vertically, and the image data is standardized so that the pixel values ​​of each channel have zero mean and unit variance;

[0019] Model construction steps: Combine the depthwise separable convolution module, the inverted residual module, the Swish activation function and the self-attention mechanism to build an object detection model;

[0020] Model training step: Iterate the object detection model through training dataset.

[0021] Preferably, the step S3 specifically includes the following steps:

[0022] Inflection point calculation steps: find the center of the circle through the heat map, set the center of the circle as the center, map it to the 3D contour map, and obtain the curve data. According to the curve data, use the first-order derivative of the curve to calculate and obtain the first-order derivative curve. According to the first-order derivative curve, obtain a top inflection point at the highest position, as well as the left inflection point and the right inflection point on the left and right sides of the top inflection point;

[0023] Angle calculation steps: According to the curve data, customize the upper and lower horizontal lines that intersect with the curve data. The two intersection points of the upper horizontal line and the curve data are connected to the top inflection point, and the angle formed is the vertex angle. The two intersection points of the lower horizontal line and the curve data are connected to the vertex to form the two lines and the lower horizontal line, and the angles are the left low angle and the right low angle. The angle values ​​of the vertex angle, left low angle, and right low angle are calculated through trigonometric function relationships.

[0024] Arc margin calculation steps: connect the top inflection point with the left inflection point and the right inflection point to obtain two straight lines. Use the point-to-straight line distance formula to traverse the distance from the curve data point to the straight line within the range of the two straight lines to calculate the maximum arc margin;

[0025] Calculation of alumina height: Screen the derivatives of the first-order derivative curve points between the two horizontal lines, and calculate the value of alumina according to the rule that the screened derivative points are smaller than the first two values ​​and larger than the last two values, and the height value of alumina is calculated.

[0026] Preferably, in step S4, the output data visualization results, model prediction results and numerical calculation results are combined, and the data are aligned and constrained under the output results of the classification and discrimination model to obtain the final detection results.

[0027] The present invention also provides a PSS process product detection system, comprising the following modules:

[0028] Module M1: Collect the .FLT files of PSS process products and obtain the product data set of PSS process products;

[0029] Module M2: Construct a classification and discrimination model, train the classification and discrimination model through the product data set, and use the classification and discrimination model to discriminate the quality inspection results of PSS process products;

[0030] Module M3: Obtain the numerical calculation results of product parameters based on the product data set;

[0031] Module M4: According to the pre-set parameters, summarize the numerical calculation results, output the discrimination information, and jointly determine the output parameters and quality inspection results based on the output discrimination information and the discrimination results of the classification discrimination model.

[0032] Preferably, the module M1 specifically includes the following modules:

[0033] Original file acquisition module: The disk produced by the PSS process is inspected by atomic force microscopy to obtain the .FLT file;

[0034] Heat map acquisition module: read the HEIGHT_SENSOR.FLT file in the .FLT file, decode it using ISO-8859-1 encoding, obtain the binary file inside, use its header file parameters to restore the data, and use data interpolation to resize the data to a 480*240 browsing data mode to obtain the heat map data;

[0035] RGB image acquisition module: reads the TM_AMPLITUDE.FLT file in the .FLT file, decodes it using ISO-8859-1 encoding, obtains the binary file inside, uses its header file parameters to restore the data, and uses data interpolation to resize the data to a 480*240 browsing data mode to obtain the RGB image data;

[0036] 3D contour map acquisition module: the obtained heat map data is mapped to the 3D field of view map according to the relationship between the corresponding values ​​of the pixel points, and the 3D contour map data is obtained;

[0037] Curve data graph acquisition module: convert the obtained heat map data into grayscale values, use Hough transform to identify circles, take the horizontal line formed by connecting the center of the circle and 160 pixel points in the horizontal plane where the circle is located, and display it visually to obtain the curve data graph data.

[0038] Preferably, the module M2 specifically includes the following modules:

[0039] Dataset production module: convert the .FLT data file into RGB image data, and divide the RGB image data into training data set, verification data set, and test data set according to the preset ratio;

[0040] Dataset enhancement module: For RGB image data, the input image is resized to 480x240 pixels, data tensorization is performed, the diversity of training data is increased by flipping the image horizontally and vertically, and the image data is standardized so that the pixel values ​​of each channel have zero mean and unit variance;

[0041] Model building module: Combine the depthwise separable convolution module, the inverted residual module, the Swish activation function and the self-attention mechanism to build the target detection model;

[0042] Model training module: Iterates the training of the target detection model through the training data set.

[0043] Preferably, the module M3 specifically includes the following modules:

[0044] Inflection point calculation module: find the center of the circle through the heat map, set the center of the circle as the center, map it to the 3D contour map, and obtain the curve data. According to the curve data, use the first-order derivative of the curve to calculate and obtain the first-order derivative curve. According to the first-order derivative curve, get a top inflection point at the highest position, as well as the left and right inflection points on the left and right sides of the top inflection point;

[0045] Angle calculation module: According to the curve data, customize the upper and lower horizontal lines that intersect with the curve data. The two intersection points of the upper horizontal line and the curve data are connected to the top inflection point, and the angle formed is the vertex angle. The two intersection points of the lower horizontal line and the curve data are connected to the vertex to form the two lines and the lower horizontal line, which are the left low angle and the right low angle. The angle values ​​of the vertex angle, left low angle, and right low angle are calculated through trigonometric function relationships.

[0046] Arc margin calculation module: connect the top inflection point with the left inflection point and the right inflection point to obtain two straight lines. Use the point-to-straight line distance formula to traverse the distance from the curve data point to the straight line within the range of the two straight lines to calculate the maximum arc margin distance;

[0047] Alumina height calculation module: filter the derivatives of the first-order derivative curve points between the two horizontal lines, calculate the value of alumina according to the rule that the filtered derivative points are smaller than the first two values ​​and larger than the last two values, and calculate the height value of alumina.

[0048] Preferably, in the module M4, the output data visualization results, model prediction results and numerical calculation results are combined, and the data are aligned and constrained under the output results of the classification and discrimination model to obtain the final detection results.

[0049] Compared with the prior art, the present invention has the following beneficial effects:

[0050] 1. The present invention uses the idea of ​​classification, based on the technical means of self-attention deep separable convolution and numerical constraints, and achieves a classification accuracy of 99% combined with numerical constraints to achieve a classification accuracy of 99.8%; the present invention uses the idea of ​​classification and introduces self-attention deep separable convolution to allow the model to learn the differences in data by itself. The deep separable convolution is the main body, which reduces the number of model parameters, runs faster and is more streamlined. It is combined with the inverted residual module to improve feature extraction, introduces nonlinear transformation, and self-attention can pay attention to the differences in classification. Coupled with numerical constraints, the accuracy of recognition is improved.

[0051] 2. The present invention proposes a method for detecting PSS process products based on an inverted residual module and numerical constraints, which detects the quality and specifications of the products through model classification and numerical calculation constraints. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Other features, objects and advantages of the present invention will become more apparent from the detailed description of non-limiting embodiments made with reference to the following drawings:

[0053] Figure 1 It is a flow chart of the steps of the PSS process product testing method;

[0054] Figure 2 Schematic diagram of the workflow of depth-wise separable convolution;

[0055] Figure 3 Schematic diagram of the workflow of the inverted residual module;

[0056] Figure 4 Schematic diagram of the work flow of the attention mechanism;

[0057] Figure 5 This is the heat map obtained from the HEIGHT_SENSOR.FLT file;

[0058] Figure 6 To obtain the RGB image from the TM_AMPLITUDE.FLT file;

[0059] Figure 7 For the general Figure 6 The effect of mapping the center of the circle to the horizontal line on the 3D contour map;

[0060] Figure 8 To obtain the center effect, the diagram and mapping are shown in Figure 7 The curve data graph obtained above;

[0061] Fig. 9 Schematic diagram of pictures corresponding to three data sets;

[0062] Fig.10 for Figure 8The derivative curve calculated from the data in and the schematic diagram showing the mapping of the inflection point to the curve;

[0063] Fig.11 Schematic diagram for corner calculation;

[0064] Fig.12 This is the effect diagram of arc margin calculation;

[0065] Fig.13 It is a display diagram of the aluminum oxide picture and the corresponding calculated aluminum oxide height points;

[0066] Fig.14 For visual display and numerical results. DETAILED DESCRIPTION

[0067] The present invention is described in detail below in conjunction with specific embodiments. The following embodiments will help those skilled in the art to further understand the present invention, but are not intended to limit the present invention in any form. It should be noted that, for those of ordinary skill in the art, several changes and improvements can also be made without departing from the concept of the present invention. These all belong to the protection scope of the present invention.

[0068] Embodiment 1:

[0069] like Figure 1 As shown, this embodiment provides a PSS process product detection method, comprising the following steps:

[0070] Step S1: Collect the .FLT file of the PSS process product and obtain the product data set of the PSS process product;

[0071] Step S1 specifically includes the following steps:

[0072] Steps for obtaining the original file: The disk produced by the PSS process is inspected by atomic force microscopy to obtain the .FLT file;

[0073] Heat map acquisition steps: read the HEIGHT_SENSOR.FLT file in the .FLT file, decode it using ISO-8859-1 encoding, obtain the binary file inside, use its header file parameters to restore the data, and use data interpolation to resize the data to a 480*240 browsing data mode to obtain the heat map data;

[0074] RGB image acquisition steps: read the TM_AMPLITUDE.FLT file in the .FLT file, decode it using ISO-8859-1 encoding, obtain the binary file inside, use its header file parameters to restore the data, and use data interpolation to resize the data to a 480*240 browsing data mode to obtain the RGB image data;

[0075] 3D contour map acquisition steps: the obtained heat map data is mapped to the 3D field of view map according to the relationship between the corresponding values ​​of the pixel points, and the 3D contour map data is obtained;

[0076] Steps for obtaining the curve data graph: convert the obtained heat map data into grayscale values, use Hough transform to identify the circle, take the horizontal line formed by connecting the center of the circle and 160 pixel points in the horizontal plane where the circle is located, and display it visually to obtain the curve data graph data.

[0077] Step S2: construct a classification and discrimination model, train the classification and discrimination model through the product data set, and use the classification and discrimination model to discriminate the quality inspection results of the PSS process products;

[0078] Step S2 specifically includes the following steps:

[0079] Dataset preparation steps: Convert the .FLT data file into RGB image data, and divide the RGB image data into training data set, verification data set, and test data set according to the preset ratio;

[0080] Dataset enhancement steps: For RGB image data, the input image is resized to 480x240 pixels, data tensorization is performed, the diversity of training data is increased by flipping the image horizontally and vertically, and the image data is standardized so that the pixel values ​​of each channel have zero mean and unit variance;

[0081] Model construction steps: Combine the depthwise separable convolution module, the inverted residual module, the Swish activation function and the self-attention mechanism to build an object detection model;

[0082] Model training step: Iterate the object detection model through training dataset.

[0083] Step S3: obtaining numerical calculation results of product parameters according to the product data set;

[0084] Step S3 specifically includes the following steps:

[0085] Inflection point calculation steps: find the center of the circle through the heat map, set the center of the circle as the center, map it to the 3D contour map, and obtain the curve data. According to the curve data, use the first-order derivative of the curve to calculate and obtain the first-order derivative curve. According to the first-order derivative curve, obtain a top inflection point at the highest position, as well as the left inflection point and the right inflection point on the left and right sides of the top inflection point;

[0086] Angle calculation steps: According to the curve data, customize the upper and lower horizontal lines with the curve data. The two intersection points of the upper horizontal line and the curve data graph are connected to the top inflection point, and the angle formed is the vertex angle. The two intersection points of the lower horizontal line and the curve data are connected to the vertex to form the two lines and the lower horizontal line. The angles are the left low angle and the right low angle. The angle values ​​of the vertex angle, left low angle, and right low angle are calculated through trigonometric function relationships.

[0087] Arc margin calculation steps: connect the top inflection point with the left inflection point and the right inflection point to obtain two straight lines. Use the point-to-straight line distance formula to traverse the distance from the curve data point to the straight line within the range of the two straight lines to calculate the maximum arc margin;

[0088] Calculation of alumina height: Screen the derivatives of the first-order derivative curve points between the two horizontal lines, and calculate the value of alumina according to the rule that the screened derivative points are smaller than the first two values ​​and larger than the last two values, and the height value of alumina is calculated.

[0089] Step S4: According to the preset parameters, the numerical calculation results are summarized, the discrimination information is output, and the output parameters and the quality inspection results are jointly determined according to the output discrimination information and the discrimination results of the classification discrimination model;

[0090] In step S4, the output data visualization results, model prediction results and numerical calculation results are combined, and the data is aligned and constrained under the output results of the classification and discrimination model to obtain the final detection results.

[0091] Step S1: Product data set: including HEIGHT_SENSOR.FLT and TM_AMPLITUDE.FLT. HEIGHT_SENSOR.FLT file is saved in 3D data format, and TM_AMPLITUDE.FLT is saved in RGB image data format.

[0092] The RGB image data is used as a model classification data set to provide a data set for the classification model in step S2. The 3D contour image data is used to obtain curve data in step S3 to provide curve data for numerical calculation.

[0093] The training iteration is performed through the RGB image obtained from the product. The classification model outputs the quality inspection results of the PSS process products by obtaining the RGB image and outputting the classification results and numerical constraints. The classification model obtains the RGB image from TM_AMPLITUDE.FLT through step S1, and then normalizes the image and converts it into the corresponding Tensor data format to input it into the model, and the result is calculated by the model.

[0094] The main body of this embodiment uses deep separable convolution to replace the traditional convolution kernel. First, each channel of the incoming 3-channel RGB image undergoes a convolution operation to obtain 3 feature maps, and then a 1*1 convolution is performed to fuse cross-channel information to generate a new feature map. Based on the deep separable convolution content, a 1*1 convolution is first performed to increase the dimension, and then a DW convolution is used to extract features (3*3 convolution kernel). After using the self-attention mechanism, a 1*1 convolution is used to reduce the dimension. The activation function used throughout the process is the swish activation function.

[0095] This embodiment calculates the angle information, arc margin information, height information, aluminum oxide height information, etc. through step S3, and makes a judgment based on the angle value of the vertex angle combined with the classification result. For example, if it is classified as a hat, but the calculated angle is less than the value set for the hat, it is acceptable. Similarly, for the calculated curvature information, if it is greater than the set value, the output result is directly judged as NG.

[0096] like Figure 1 As shown in FIG. 1 , the workflow of the PSS process detection of this embodiment is shown. Figure 2 As shown, the workflow of the depth-separable convolution is used to obtain RGB Figure 3 Channel information, feature extraction is achieved by using depthwise convolution + point-by-point convolution.

[0097] like Figure 3 As shown in the figure, it is the workflow of the inverted residual module. The whole process is "expansion--convolution-attention-compression", which is shuttle-shaped; the convolution operation is: convolution dimension increase (1×1)-DW convolution feature extraction (3×3)-attention (3*3,5*5)-convolution dimension reduction (1×1); the swish activation function is used.

[0098] like Figure 4 The figure shows the implementation process of the attention mechanism. By learning the mutual characteristics between different numbers of channels, the classification features can be learned. The size of the convolution kernel is adjusted through two convolution kernels (3*3, 5*5) to adapt to targets of different scales. This mechanism allows the network to more effectively capture the multi-scale features of the image space while reducing the waste of computing resources. Attention includes three operations: splitting, fusion, and selection, which realize the fusion and selection of multi-scale information.

[0099] like Figure 5 As shown, this is a heat map obtained from the HEIGHT_SENSOR.FLT file, which is a top view of the 3D data mapped to an inverted 2D plane in order to find the center of the circle.

[0100] like Figure 6 As shown, the RGB image is obtained from the TM_AMPLITUDE.FLT file, which is used as a data set core display image for model classification.

[0101] like Figure 7 As shown in Figure 5 The center of the circle is found based on the 3D contour map, and then the center of the circle is mapped to the horizontal line on the 3D contour map. Figure 8 Get curve data.

[0102] like Figure 8 As shown in the figure, the center effect is shown, and there is also a map mapped in Figure 7 The curve data graph obtained above provides curve data for numerical calculation.

[0103] like Fig. 9 As shown, the three pictures in the figure correspond to the three data sets mentioned above, showing the differences in the data.

[0104] like Fig.10 As shown, through Figure 8 The derivative curve is calculated, and the inflection point is mapped to the curve and displayed to provide inflection point data for numerical calculation.

[0105] like Fig.11 As shown, two horizontal lines are defined to connect the inflection points and provide for the corner calculation.

[0106] like Fig.12 As shown in the figure, it is the effect diagram of arc margin calculation. Fig.13 As shown in the figure, it is a picture of aluminum oxide and a picture showing the corresponding calculated aluminum oxide height points. Fig.14 As shown, there are visual display diagrams and numerical result pictures.

[0107] The present invention also provides a PSS process product detection system, which can be implemented by executing the process steps of the PSS process product detection method, that is, those skilled in the art can understand the PSS process product detection method as a preferred implementation of the PSS process product detection system.

[0108] Embodiment 2:

[0109] This embodiment provides a PSS process product detection system, including the following modules:

[0110] Module M1: Collect the .FLT files of PSS process products and obtain the product data set of PSS process products;

[0111] Module M1 specifically includes the following modules:

[0112] Original file acquisition module: The disk produced by the PSS process is inspected by atomic force microscopy to obtain the .FLT file;

[0113] Heat map acquisition module: read the HEIGHT_SENSOR.FLT file in the .FLT file, decode it using ISO-8859-1 encoding, obtain the binary file inside, use its header file parameters to restore the data, and use data interpolation to resize the data to a 480*240 browsing data mode to obtain the heat map data;

[0114] RGB image acquisition module: reads the TM_AMPLITUDE.FLT file in the .FLT file, decodes it using ISO-8859-1 encoding, obtains the binary file inside, uses its header file parameters to restore the data, and uses data interpolation to resize the data to a 480*240 browsing data mode to obtain the RGB image data;

[0115] 3D contour map acquisition module: the obtained heat map data is mapped to the 3D field of view map according to the relationship between the corresponding values ​​of the pixel points, and the 3D contour map data is obtained;

[0116] Curve data graph acquisition module: convert the obtained heat map data into grayscale values, use Hough transform to identify circles, take the horizontal line formed by connecting the center of the circle and 160 pixel points in the horizontal plane where the circle is located, and display it visually to obtain the curve data graph data.

[0117] Module M2: Construct a classification and discrimination model, train the classification and discrimination model through the product data set, and use the classification and discrimination model to discriminate the quality inspection results of PSS process products;

[0118] Module M2 specifically includes the following modules:

[0119] Dataset production module: convert the .FLT data file into RGB image data, and divide the RGB image data into training data set, verification data set, and test data set according to the preset ratio;

[0120] Dataset enhancement module: For RGB image data, the input image is resized to 480x240 pixels, data tensorization is performed, the diversity of training data is increased by flipping the image horizontally and vertically, and the image data is standardized so that the pixel values ​​of each channel have zero mean and unit variance;

[0121] Model building module: Combine the depthwise separable convolution module, the inverted residual module, the Swish activation function and the self-attention mechanism to build the target detection model;

[0122] Module M3 specifically includes the following modules:

[0123] Inflection point calculation module: find the center of the circle through the heat map, set the center of the circle as the center, map it to the 3D contour map, and obtain the curve data. According to the curve data, use the first-order derivative of the curve to calculate and obtain the first-order derivative curve. According to the first-order derivative curve, get a top inflection point at the highest position, as well as the left and right inflection points on the left and right sides of the top inflection point;

[0124] Angle calculation module: According to the curve data, customize the upper and lower horizontal lines that intersect with the curve data. The two intersection points of the upper horizontal line and the curve data are connected to the top inflection point, and the angle formed is the vertex angle. The two intersection points of the lower horizontal line and the curve data are connected to the vertex to form the two lines and the lower horizontal line, which are the left low angle and the right low angle. The angle values ​​of the vertex angle, left low angle, and right low angle are calculated through trigonometric function relationships.

[0125] Arc margin calculation module: connect the top inflection point with the left inflection point and the right inflection point to obtain two straight lines. Use the point-to-straight line distance formula to traverse the distance from the curve data point to the straight line within the range of the two straight lines to calculate the maximum arc margin distance;

[0126] Alumina height calculation module: filter the derivatives of the first-order derivative curve points between the two horizontal lines, calculate the value of alumina according to the rule that the filtered derivative points are smaller than the first two values ​​and larger than the last two values, and calculate the height value of alumina.

[0127] Model training module: Iterates the training of the target detection model through the training data set.

[0128] Module M3: Obtain the numerical calculation results of product parameters based on the product data set;

[0129] Module M4: summarizes the numerical calculation results according to the preset parameters, outputs the discrimination information, and jointly determines the output parameters and quality inspection results according to the output discrimination information and the discrimination results of the classification discrimination model;

[0130] In module M4, the output data visualization results, model prediction results and numerical calculation results are combined, and the data is aligned and constrained under the output results of the classification and discrimination model to obtain the final detection results.

[0131] Embodiment 3:

[0132] Those skilled in the art may understand this embodiment as a more specific description of Embodiment 1 and Embodiment 2.

[0133] The present invention provides a detection method and system for PSS process products based on self-attention deep separable convolution, including: step S1: reading the .FLT file generated by an atomic force microscope to obtain relevant parameters and their basic data; step S2: using a model to discriminate quality inspection results from the basic data; step S3: obtaining numerical calculation results and display diagrams from relevant parameters and basic data; step S4: according to pre-set parameters, summarizing the numerical calculation results to output discrimination information and jointly determining the output parameters and quality inspection results with the output results of the classification model.

[0134] Step S1: Get data display

[0135] Step S1.1: Get the original file:

[0136] The disk produced by the PSS process was inspected by atomic force microscopy to obtain the .FLT file.

[0137] Step S1.2: Get the heat map:

[0138] By reading the HEIGHT_SENSOR.FLT file, decoding it using ISO-8859-1 encoding, obtaining the binary file inside, and then using its header file parameters to restore the data, and then resizing the data to 480*240 browsing data mode using data interpolation. Figure 5 shown.

[0139] Step S1.3: Get the RGB image:

[0140] By reading the TM_AMPLITUDE.FLT file, decoding it using ISO-8859-1 encoding, obtaining the binary file inside, and then using its header file parameters to restore the data, and then resizing the data to 480*240 browsing data mode using data interpolation. Figure 6 shown.

[0141] Step S1.4: Obtain 3D contour map:

[0142] The heat map data is mapped to the 3D view map according to the relationship between the pixel points and the corresponding values, and then displayed. Figure 7 shown.

[0143] Step S1.5: Get curve data graph:

[0144] Based on the heat map obtained above, we convert it to grayscale and use Hough transform to identify the circle, then take a horizontal line of 160 pixels from the center of the circle and visualize it. Figure 8 shown.

[0145] Step S2: Model classification and discrimination

[0146] Step S2.1: Dataset:

[0147] Step S2.1.1: Dataset creation:

[0148] From step S1, we can get the RGB image converted from the .FLT data file, such as Fig. 9 As shown, the visualization shows that the description is that due to different production stability, different types of defects will occur, such as platforms, caps, etc.

[0149] Therefore, it is easier to divide the data. The response products are put into corresponding folders through manual classification, which are divided into training data, validation data and test data. Due to the needs of the production line, it is impossible to collect all the data. Therefore, the training is carried out according to the ratio of 70% of training data and 30% of validation data, and then tested on the product line. After feedback and manual observation, the data collection is completed on the premise of confirming that the accuracy rate is achieved.

[0150] Step S2.1.2: Dataset enhancement:

[0151] Rational data augmentation is used to improve the generalization of the model. For the data images, the input image size is adjusted to 480x240 pixels, and the data is tensored. By flipping the image horizontally, the diversity of the training data is increased to prevent the model from overfitting to a specific direction of the image. By flipping the image vertically, the diversity of the training data is further increased to prevent the model from overfitting to a specific direction of the image. The image data is standardized so that the pixel values ​​of each channel have zero mean and unit variance, thereby accelerating the training convergence of the model.

[0152] Step S2.2: Model architecture:

[0153] Regarding the choice of model architecture, the application scenario is the daily office computer host Windows system (usually CPUi5 type), and the computing resources are limited. Therefore, the model architecture requires a lightweight network. A lightweight network focusing on mobile and embedded devices is selected. Therefore, the model architecture is constructed by combining the depthwise separable convolution module, the inverted residual module, the Swish activation function and the self-attention mechanism.

[0154] Step S2.3: Model training:

[0155] Step S2.3.1: Data preprocessing and loading:

[0156] First, we define the data preprocessing steps, including image resizing, random horizontal and vertical flipping, and normalization. These steps ensure that the input images have consistent shape and distribution during training.

[0157] Step S2.3.2: Training model:

[0158] Since the recognition result is a three-classification problem, the cross entropy loss function is designed to calculate the difference between the model prediction output and the true label, and the Adam optimizer is used to update the model parameters according to the calculated loss. The learning rate is set to 0.0001. The training process includes the following steps: setting the model to training mode, traversing the training data for forward and backward propagation, evaluating the model performance on the validation set, and saving the best model parameter file as .pth.

[0159] Table 1 Configuration parameters

[0160] Parameter name Parameter Value Remark Training round (epoch) 600 Batch size(batch) 64 Learning rate (lr) 0.0001 Image size (ingsize) 480*240 Optimizer Adam

[0161] Training result output:

[0162] Table 2 Algorithm results

[0163]

[0164] Step S2.3.3: Model inference:

[0165] Since the test data is continuously produced by the PSS process, we selected 2,000 data for testing, of which 1,000 are OK data, 500 are hat data, and 500 are platform data. The recognition accuracy results are as follows.

[0166] Table 3 Recognition accuracy results

[0167]

[0168]

[0169] It can be seen that the output recognition results show that the accuracy of OK recognition is 99.8%, the recognition effect of hat and platform is 100%, and the comprehensive output is 99.9%.

[0170] Step S3: Numerical calculation and display:

[0171] Step S3.1: Inflection point calculation:

[0172] The inflection point is calculated as the point and center of the circle corresponding to the arc in the image. According to the curve data graph, the first-order derivative of the curve can be calculated to obtain the first-order derivative curve, and then the point closest to 0 is obtained. The leftmost point is the first inflection point, the highest corresponding value is the cusp of the second inflection point, and the rightmost point is the third inflection point. Fig.10 shown.

[0173] Step S3.2: Angle calculation:

[0174] The angle calculation is the angle between the intersection and the cusp of the two horizontal lines and the curve. The point is calculated based on the inflection point, and the two points are connected by a straight line. Then two horizontal lines are customized. The upper horizontal line connects the two intersection points of the curve to the vertex, and the angle formed is the vertex angle. The angle between the lower horizontal line and the vertex at the intersection of the curve and the horizontal line is the left and right low angles, and then the angle value is calculated through the trigonometric function relationship. Fig.11 shown.

[0175] Step S3.3: Arc margin calculation:

[0176] When calculating the angle, we get two straight lines. The principle of calculating the arc margin is to use the distance formula from the point to the straight line. We traverse the distance from the curve point to the straight line inside the two points and calculate the maximum arc margin distance on the left and right. Fig.12 shown.

[0177] Step S3.4: Alumina height calculation:

[0178] The height of alumina is the value of the heat map corresponding to the middle ring. The height of alumina is calculated by screening with rule constraints. The derivative of the curve point between the second straight line is used for screening. The derivative point is smaller than the first two values ​​and larger than the last two values. The value of alumina is calculated artificially, thus obtaining the height of alumina. Fig.13 shown.

[0179] Step S4: Comprehensive output results:

[0180] The combined output data can be viewed as the result, the model prediction result and the numerical calculation result, and the final detection result is obtained by aligning the data and constraining it under the model classification result. Fig.14 shown.

[0181] The PSS process product detection method based on the inverted residual module and numerical constraints proposed in the present invention detects the quality and specifications of the product through model classification and numerical calculation constraints.

[0182] Those skilled in the art know that, in addition to realizing the system and its various devices, modules, and units provided by the present invention in a purely computer-readable program code, the system and its various devices, modules, and units provided by the present invention can be completely realized in the form of logic gates, switches, application-specific integrated circuits, programmable logic controllers, and embedded microcontrollers by logically programming the method steps to realize the same functions. Therefore, the system and its various devices, modules, and units provided by the present invention can be considered as a hardware component, and the devices, modules, and units included therein for realizing various functions can also be regarded as structures within the hardware component; the devices, modules, and units for realizing various functions can also be regarded as both software modules for realizing the method and structures within the hardware component.

[0183] The above describes the specific embodiments of the present invention. It should be understood that the present invention is not limited to the above specific embodiments, and those skilled in the art can make various changes or modifications within the scope of the claims, which does not affect the essence of the present invention. In the absence of conflict, the embodiments of the present application and the features in the embodiments can be combined with each other arbitrarily.

Claims

1. A PSS process product detection method, characterized in that: The steps include: Step S1: Collect the .FLT file of the PSS process product and obtain the product data set of the PSS process product; Step S2: construct a classification and discrimination model, train the classification and discrimination model through the product data set, and use the classification and discrimination model to discriminate the quality inspection results of the PSS process products; Step S3: obtaining numerical calculation results of product parameters according to the product data set; Step S4: According to the preset parameters, the numerical calculation results are summarized and the discrimination information is output. According to the output discrimination information and the discrimination result of the classification discrimination model, the output parameters and the quality inspection results are jointly determined.

2. The PSS process product detection method according to claim 1, characterized in that: The step S1 specifically includes the following steps: Steps for obtaining the original file: The disk produced by the PSS process is inspected by atomic force microscopy to obtain the .FLT file; Heat map acquisition steps: read the HEIGHT_SENSOR.FLT file in the .FLT file, decode it using ISO-8859-1 encoding, obtain the binary file inside, use its header file parameters to restore the data, and use data interpolation to resize the data to a 480*240 browsing data mode to obtain the heat map data; RGB image acquisition steps: read the TM_AMPLITUDE.FLT file in the .FLT file, decode it using ISO-8859-1 encoding, obtain the binary file inside, use its header file parameters to restore the data, and use data interpolation to resize the data to a 480*240 browsing data mode to obtain the RGB image data; 3D contour map acquisition steps: the obtained heat map data is mapped to the 3D field of view map according to the relationship between the corresponding values ​​of the pixel points, and the 3D contour map data is obtained; Steps for obtaining the curve data graph: convert the obtained heat map data into grayscale values, use Hough transform to identify the circle, take the horizontal line formed by connecting the center of the circle and 160 pixel points in the horizontal plane where the circle is located, and display it visually to obtain the curve data graph data.

3. The PSS process product detection method according to claim 1, characterized in that: The step S2 specifically includes the following steps: Dataset preparation steps: Convert the .FLT data file into RGB image data, and divide the RGB image data into training data set, verification data set, and test data set according to the preset ratio; Dataset enhancement steps: For RGB image data, the input image is resized to 480x240 pixels, data tensorization is performed, the diversity of training data is increased by flipping the image horizontally and vertically, and the image data is standardized so that the pixel values ​​of each channel have zero mean and unit variance; Model construction steps: Combine the depthwise separable convolution module, the inverted residual module, the Swish activation function and the self-attention mechanism to build an object detection model; Model training step: Iterate the object detection model through training dataset.

4. The PSS process product detection method according to claim 1, characterized in that: The step S3 specifically includes the following steps: Inflection point calculation steps: find the center of the circle through the heat map, set the center of the circle as the center, map it to the 3D contour map, and obtain the curve data. According to the curve data, use the first-order derivative of the curve to calculate and obtain the first-order derivative curve. According to the first-order derivative curve, obtain a top inflection point at the highest position, as well as the left inflection point and the right inflection point on the left and right sides of the top inflection point; Angle calculation steps: According to the curve data, customize the upper and lower horizontal lines that intersect with the curve data. The two intersection points of the upper horizontal line and the curve data are connected to the top inflection point, and the angle formed is the vertex angle. The two intersection points of the lower horizontal line and the curve data are connected to the vertex to form the two lines and the lower horizontal line, and the angles are the left low angle and the right low angle. The angle values ​​of the vertex angle, left low angle, and right low angle are calculated through trigonometric function relationships. Arc margin calculation steps: connect the top inflection point with the left inflection point and the right inflection point to obtain two straight lines. Use the point-to-straight line distance formula to traverse the distance from the curve data point to the straight line within the range of the two straight lines to calculate the maximum arc margin; Calculation of alumina height: Screen the derivatives of the first-order derivative curve points between the two horizontal lines, and calculate the value of alumina according to the rule that the screened derivative points are smaller than the first two values ​​and larger than the last two values, and the height value of alumina is calculated.

5. The PSS process product detection method according to claim 1, characterized in that: In step S4, the output data visualization results, model prediction results and numerical calculation results are combined, and the data is aligned and constrained under the output results of the classification and discrimination model to obtain the final detection results.

6. A PSS process product detection system, characterized in that: Includes the following modules: Module M1: Collect the .FLT files of PSS process products and obtain the product data set of PSS process products; Module M2: Construct a classification and discrimination model, train the classification and discrimination model through the product data set, and use the classification and discrimination model to discriminate the quality inspection results of PSS process products; Module M3: Obtain the numerical calculation results of product parameters based on the product data set; Module M4: According to the pre-set parameters, summarize the numerical calculation results, output the discrimination information, and jointly determine the output parameters and quality inspection results based on the output discrimination information and the discrimination results of the classification discrimination model.

7. The PSS process product detection system according to claim 6, characterized in that: The module M1 specifically includes the following modules: Original file acquisition module: The disk produced by the PSS process is inspected by atomic force microscopy to obtain the .FLT file; Heat map acquisition module: read the HEIGHT_SENSOR.FLT file in the .FLT file, decode it using ISO-8859-1 encoding, obtain the binary file inside, use its header file parameters to restore the data, and use data interpolation to resize the data to a 480*240 browsing data mode to obtain the heat map data; RGB image acquisition module: reads the TM_AMPLITUDE.FLT file in the .FLT file, decodes it using ISO-8859-1 encoding, obtains the binary file inside, uses its header file parameters to restore the data, and uses data interpolation to resize the data to a 480*240 browsing data mode to obtain the RGB image data; 3D contour map acquisition module: the obtained heat map data is mapped to the 3D field of view map according to the relationship between the corresponding values ​​of the pixel points, and the 3D contour map data is obtained; Curve data graph acquisition module: convert the obtained heat map data into grayscale values, use Hough transform to identify circles, take the horizontal line formed by connecting the center of the circle and 160 pixel points in the horizontal plane where the circle is located, and display it visually to obtain the curve data graph data.

8. The PSS process product detection system according to claim 6, characterized in that: The module M2 specifically includes the following modules: Dataset production module: convert the .FLT data file into RGB image data, and divide the RGB image data into training data set, verification data set, and test data set according to the preset ratio; Dataset enhancement module: For RGB image data, the input image is resized to 480x240 pixels, data tensorization is performed, the diversity of training data is increased by flipping the image horizontally and vertically, and the image data is standardized so that the pixel values ​​of each channel have zero mean and unit variance; Model building module: Combine the depthwise separable convolution module, the inverted residual module, the Swish activation function and the self-attention mechanism to build the target detection model; Model training module: Iterates the training of the target detection model through the training data set.

9. The PSS process product detection system according to claim 6, characterized in that: The module M3 specifically includes the following modules: Inflection point calculation module: find the center of the circle through the heat map, set the center of the circle as the center, map it to the 3D contour map, and obtain the curve data. According to the curve data, use the first-order derivative of the curve to calculate and obtain the first-order derivative curve. According to the first-order derivative curve, get a top inflection point at the highest position, as well as the left and right inflection points on the left and right sides of the top inflection point; Angle calculation module: According to the curve data, customize the upper and lower horizontal lines that intersect with the curve data. The two intersection points of the upper horizontal line and the curve data are connected to the top inflection point, and the angle formed is the vertex angle. The two intersection points of the lower horizontal line and the curve data are connected to the vertex to form the two lines and the lower horizontal line, which are the left low angle and the right low angle. The angle values ​​of the vertex angle, left low angle, and right low angle are calculated through trigonometric function relationships. Arc margin calculation module: connect the top inflection point with the left inflection point and the right inflection point to obtain two straight lines. Use the point-to-straight line distance formula to traverse the distance from the curve data point to the straight line within the range of the two straight lines to calculate the maximum arc margin distance; Alumina height calculation module: filter the derivatives of the first-order derivative curve points between the two horizontal lines, calculate the value of alumina according to the rule that the filtered derivative points are smaller than the first two values ​​and larger than the last two values, and calculate the height value of alumina.

10. The PSS process product detection system according to claim 6, characterized in that: In the module M4, the output data visualization results, model prediction results and numerical calculation results are combined, and the data is aligned and constrained under the output results of the classification and discrimination model to obtain the final detection results.

Images