A method for determining the pass / failability of tensile force-stretch ratio curves based on image recognition

Abstract

This invention discloses a method for determining the pass / failability of a tensile force-stretch ratio curve based on image recognition, comprising the following steps: S1, collecting tensile force-time series data under multiple sets of preset process parameters; S2, extracting the stretching segment data, converting the tensile force-time data into tensile force-stretch ratio data, and plotting a curve based on the time series data; S3, constructing a curve classification model based on a dual-path feature extraction network; S4, iteratively training the curve classification model, and using the trained curve classification model for determining the pass / failability of the tensile force-stretch ratio curve. This invention utilizes the tensile force-stretch ratio curve generated during the stretching process of a biaxial stretching apparatus to establish a deep mapping from the production dynamic curve to the final quality index of the film, thereby determining the process pass / failability.

Description

Technical Field

[0001] This invention belongs to the field of thin film quality testing technology, specifically relating to a method for determining the pass / failability of tensile force-stretch ratio curves based on image recognition. Background Technology

[0002] The tensile force-draw ratio curve of a thin film reveals the rheological characteristics of the material at specific temperatures, draw ratios, and drawing speeds. Experienced technicians can preliminarily determine the rationality of current process parameters and predict risks such as uneven film thickness, unbalanced stress distribution, or film breakage and delamination by observing the yield point location, the slope of the reinforcing section, and the smoothness of the curve. However, in actual production, the following difficulties exist: firstly, expert experience cannot be replicated; secondly, offline testing is time-consuming, quality feedback is delayed, and efficiency is low.

[0003] Currently, most technical solutions for automated classification of such curves use conventional convolutional neural networks to directly classify curve images. Gao Zhenguo et al. proposed a network traffic anomaly detection method and system based on curve filling graph classification. Its core logic is to convert one-dimensional time-series signals into two-dimensional curve filling graphs, and then use convolutional neural networks for classification. The dataset is constructed by collecting one-dimensional time-series traffic data under preset states (normal / abnormal), converting the one-dimensional time-series data into two-dimensional images. The image conversion not only involves drawing curves, but also selectively filling time-series curves. The anomaly detection model is obtained by iterative training using a CNN network, and finally the trained model is applied to online detection of real-time traffic. While the aforementioned technology achieves the conversion and classification from one-dimensional signals to two-dimensional images, it still has some shortcomings when applied to industrial monitoring and inspection scenarios for biaxially stretched thin films: the feature extraction scale is singular, making it difficult to consider both overall trends and local abrupt changes. This is because the model architecture involved in this technology is a single-path linear structure, and its feature extraction depth is relatively limited. However, in the biaxially stretched thin film process, it is necessary to consider both the overall trend of the curve and local changes. The trend of the curve shape can reflect the stretching quality of the film, and local fluctuations in the curve, such as abrupt changes in tensile force, also reflect the rationality of the process parameters. In addition, this technology lacks the ability to focus on the physical curve region and is easily affected by background noise. Although it enhances features through selective filling, the network still assigns equal computational weights to non-critical pixel regions in the image during model computation. Summary of the Invention

[0004] The main objective of this invention is to overcome the shortcomings and deficiencies of the prior art and to propose a method for determining the pass / failability of tensile force-tensile ratio curves based on image recognition.

[0005] To achieve the above objectives, the present invention adopts the following technical solution:

[0006] A method for determining the pass / failability of tensile force-tensile ratio curves based on image recognition includes the following steps:

[0007] S1. Collect tensile force-time series data under multiple sets of preset process parameters;

[0008] S2. Extract the data of the tensile segment, convert the tensile force-time data into tensile force-tensile ratio data, and plot a curve based on the time series data;

[0009] S3. Construct a curve graph classification model based on a dual-path feature extraction network;

[0010] S4. Iteratively train the curve classification model and use the trained curve classification model to determine the pass / failability of the tensile force-tensile ratio curve.

[0011] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0012] 1. This invention optimizes the process system of a biaxial stretching apparatus for thin films, establishing a deep mapping from the production dynamic curve to the final quality index of the film using the tensile force-stretch ratio curve generated during the stretching process, thereby determining the process qualification. To make reasonable use of the tensile force-stretch ratio curve, based on the characteristics of actual data acquisition from laboratory-level biaxial stretching apparatus equipment, the original tensile force-time series data obtained from the data acquisition system is processed through a series of software methods. This includes firstly extracting stretching segment data from the original tensile force-time series data, mainly including determining the starting point of the stretching segment, determining the number of data points N in the stretching segment, and resetting the sequence data of the segment. Then, the extracted tensile force-time series data is converted into tensile force-stretch ratio data according to the formula derived from the working conditions.

[0013] 2. Based on the actual characteristics of the tensile force curve in the biaxial stretching process of thin films, this invention constructs an asymmetric cascaded dual-path feature extraction and fusion model architecture based on the ResNet network. The model's discrimination accuracy is significantly improved. Unlike the linear single-path convolutional neural network structure of existing technologies, this invention can simultaneously and accurately identify the overall topological orientation and local abrupt changes of the tensile force curve of thin films. The classification accuracy and F1 score are much higher than those of traditional single-path models, effectively solving the problem of the difficulty in quantifying and inheriting human experience in the process system.

[0014] 3. This invention couples a spatial attention module (SAM) into the feature path. This module automatically learns and generates pixel-level weight masks by performing average pooling and max pooling in parallel along the channel dimension. This achieves saliency enhancement of curved regions and adaptive suppression of background regions on the feature layer. In contrast, existing technologies mainly rely on static image preprocessing methods such as "selective filling" to enhance features, but they still perform equal weight calculations on all pixels in the image during network computation, making it impossible to distinguish interference regions at the feature level. Attached Figure Description

[0015] Figure 1 This is a flowchart of the method of the present invention.

[0016] Figure 2 This is a graph of the original tensile force-time series data in the embodiment.

[0017] Figure 3 This is a tensile force-tensile ratio curve in the embodiment.

[0018] Figure 4a These are images of some of the qualified samples in the embodiments.

[0019] Figure 4b These are images of some of the non-compliant samples in the embodiments.

[0020] Figure 5 This is an architecture diagram of the curve graph classification model based on a dual-path feature extraction network in the embodiment.

[0021] Figure 6 This is the classification result of the model in the embodiment. Detailed Implementation

[0022] The present invention will be further described in detail below with reference to the embodiments and accompanying drawings, but the embodiments of the present invention are not limited thereto.

[0023] Examples; such as Figure 1 As shown, a method for determining the pass / failability of a tensile force-tensile ratio curve based on image recognition includes the following steps:

[0024] S1. Collect tensile force-time series data under multiple sets of preset process parameters, specifically:

[0025] The film biaxial stretching test is carried out according to the preset process parameter group. The film biaxial stretching instrument senses the magnitude of the tensile force through the strain gauge force sensor arranged at the clamp position on the fixture. After the signal is converted by the industrial grade signal amplifier, it enters the A / D acquisition module, collects and saves multiple sets of tensile force-time series data, and uses the upper computer monitoring software to monitor the tensile force-time series data. At the same time, the experimental results of the corresponding process parameter group are recorded, that is, whether the film quality is qualified or not.

[0026] S2. Extract the data of the tensile segment, convert the tensile force-time data into tensile force-tensile ratio data, and plot a curve based on the time series data;

[0027] Because the data acquisition module of the biaxial stretching machine collects data from the entire stretching test process, such as... Figure 2 As shown, the graph is a plot of the original tensile force-time series data. The entire tensile test process may include data from the subsequent shaping process and some data from the preheating stage (during the heat preservation time) before the formal tensile test begins. The tensile segment data generated by the formal and effective tensile process is the data of interest in this embodiment. Therefore, the tensile segment data is extracted by software method.

[0028] Extract the data from the stretched segment, including:

[0029] S21. The sequence data involved in this embodiment has the following characteristics: before stretching begins, the tensile force of the fixture is 0, and the data is relatively flat or has slight fluctuations; after stretching begins, the data starts to rise due to the tensile force. Based on these characteristics, the starting point of the stretching segment is determined by the positive and negative values ​​of the first derivative of the time series data and a threshold value for positive values, specifically:

[0030] A data window is set according to the density of the data points collected. The size of the data window is adjusted according to the actual situation (the size of the data window used in this embodiment is 10). The stretching start point is determined according to the proportion of the probability that the first derivative of the data in the window is positive and whether the value of the derivative reaches the threshold when it is positive. The required data points are recorded as needed during the process.

[0031] S22. Based on the working principle and process flow of the laboratory biaxial stretching apparatus, the preset process parameters are obtained. The number of data points N in the stretching section is determined according to the preset stretching speed v, the stretching ratio n of the film, the data acquisition frequency f, and the original length l of the film. The formula is:

[0032] ;

[0033] Then, N data points starting from the beginning of the stretching are extracted, and the tensile force-time series data of the stretching segment is reset according to the acquisition frequency.

[0034] S23. After determining the number of data points N for the stretching segment, the tensile force data for the corresponding stretching segment is determined. However, the corresponding time data is not data starting from 0. Therefore, the time series data for this segment needs to be reset. Specifically, based on the sampling frequency f, the sampling time interval is determined by 1 / f. With 1 / f as the tolerance and 0 as the starting point, the N time data corresponding to the N tensile force data are filled with an arithmetic sequence of time data to finally obtain the tensile force-time series data of the stretching segment.

[0035] In step S2, the tensile force-time data is converted into tensile force-stretch ratio data, specifically as follows:

[0036] Let the stretch ratio be The length of the sample after stretching is L. Since synchronous constant speed biaxial stretching is used, the stretching speed v is constant, so L = vt. Therefore, the stretch ratio is... Based on this formula, the tensile force-time data is processed by software and converted into a sequence of tensile force-stretch ratio data. The converted tensile force-stretch ratio curve is shown below. Figure 3 As shown.

[0037] In step S2, a curve is plotted based on the time series data, specifically as follows:

[0038] Draw a curve graph and scale it uniformly to 224×224 pixels;

[0039] To avoid interference from coordinate axes, grid lines, titles, etc., in the dataset images, we uniformly do not save coordinate axes, grid lines, and other irrelevant information.

[0040] Based on the experimental results in step S1, each curve is labeled and categorized into two classes, corresponding to qualified and unqualified, respectively. For example... Figure 4a As shown, these are images from a portion of the qualified datasets. Figure 4b The image shown is an image of a partially invalid dataset.

[0041] To improve the model's generalization and robustness, the input image to the curve classification model based on a dual-path feature extraction network is standardized and reconstructed. First, a pixel linear scaling operator is used. Discrete grayscale data is mapped to continuous floating-point tensors using channel normalization operators. The data is zero-mean normalized, where the mean vector is... and standard deviation vector The statistical distribution characteristics are taken from the image datasets respectively.

[0042] S3. Construct a curve graph classification model based on a dual-path feature extraction network; such as... Figure 5 The diagram shown is the overall architecture of the model.

[0043] The curve graph classification model based on the dual-path feature extraction network includes a shared basic feature extraction layer, an overall trend feature extraction path, and a local mutation feature extraction path.

[0044] Before entering the curve classification model based on a dual-path feature extraction network, the input image also undergoes data partitioning based on hierarchical five-fold cross-validation, specifically as follows:

[0045] To fully utilize the limited experimental data and verify the model's stability, a 5K-Fold cross-validation mechanism is used to make the final average accuracy and F1 score more objective. The 5K-Fold method is as follows:

[0046] The total sample set is divided into 5 subsets, and four of them are selected as the training set and one as the validation set each time.

[0047] In actual production, qualified samples are usually more numerous than abnormal samples. Therefore, weighted sampling is used to address class imbalance. By calculating the sampling weight W=1 / {number of samples in each class} for each sample, the model automatically increases the frequency of rare samples based on the weights during subsequent training iterations. This effectively solves the problem of the model easily shifting to the majority class in industrial scenarios with small sample sizes.

[0048] The shared basic feature extraction layer is specifically as follows:

[0049] After data partitioning based on the 5K-Fold method, the input images (training set images) first enter the shared basic feature extraction layer of the ResNet network. The main function of this layer is to capture the basic visual features of the image (such as the edges, thickness, and basic direction of curves) as common input for subsequent paths. In the shared basic feature extraction layer, the image sequentially passes through a 7x7 convolutional layer (conv1), a batch normalization layer (bn1), a rectified linear unit (ReLU) activation function, a max pooling layer, and the residual block groups Layer1 and Layer2 of the ResNet network.

[0050] Layer1 is used to extract primary textures, and Layer2 is used to extract complex geometric line features;

[0051] The input image with a size of 224×224 enters the shared basic feature extraction layer. After processing from conv1 to Layer2, the spatial size of the feature map is gradually downsampled from 224×224 to a size of 28×28, and the low-level feature map containing basic geometric information is extracted.

[0052] The specific path for extracting overall trend features is as follows:

[0053] The overall trend feature extraction path takes the 28×28 bottom feature map obtained after processing by the shared basic feature extraction layer as input; this path focuses on capturing the overall shape of the film stretching curve and is responsible for extracting the overall topological morphology, slope trend and other mid-level features of the curve.

[0054] In the overall trend feature extraction path, the 28×28 image after the features are received by Layer 2 sequentially enters the third residual block group (Layer 3), the first spatial attention module (Spatial Attention 1), and the global average pooling layer (AvgPool). After the third residual block group, the spatial scale of the image becomes 14×14, and the resulting feat3 feature map is enhanced after passing through the first spatial attention module. The first spatial attention module focuses attention on the pixel region where the "line" is located by superimposing average pooling and max pooling, as well as 7x7 convolution and sigmoid activation, suppressing useless background filling. The feat3 data stream is multiplied by the attention weight mask to obtain the Weighted feat3, which represents the enhancement of the overall topological shape of the curve.

[0055] Finally, a global adaptive average pooling layer is used to compress the feature map to a 1×1 spatial size, outputting a 256-dimensional overall trend feature vector.

[0056] The specific path for extracting local mutation features is as follows:

[0057] The local mutation feature extraction path is responsible for capturing high-level features such as minute local mutations and subtle fluctuations at strain hardening points in the curve. The local mutation feature extraction path takes the feature map obtained from the third residual block group of the overall trend feature extraction path as input, and then passes through the fourth residual block group (Layer 4), the second spatial attention module (SpatialAttention2), and the global average pooling layer (AvgPool).

[0058] After the fourth residual block group, the spatial scale of the image is further refined to 7×7. The convolutional kernel at this stage has a large effective receptive field, which can identify abstract features in the curve, including small nonlinear changes in the strain hardening slope and subtle stress oscillation signals generated during the material yielding stage. The resulting feat4 feature map is enhanced into Weighted feat4 after passing through the second spatial attention module. Finally, the feature map is compressed to a spatial size of 1×1 through a global adaptive average pooling layer, while doubling the number of feature channels from 256 to 512, and outputting a 512-dimensional local mutation feature vector.

[0059] S4. Iteratively train the curve classification model and use the trained curve classification model for the qualification determination of the tensile force-tensile ratio curve.

[0060] The feature fusion and classification head layer concatenates the multi-scale features of the two paths, that is, it physically concatenates the 256-dimensional vector of the overall trend feature extraction path and the 512-dimensional vector of the local mutation feature extraction path to form a 768-dimensional full-scale feature vector, thus realizing a "panoramic" feature fusion from the whole to the part.

[0061] The model then sequentially passes through a random dropout layer, which discards neurons with a probability of 0.5, reducing complex dependencies between adjacent layers and thus significantly lowering the risk of overfitting and improving generalization ability. Following this, it passes through a first fully connected layer (Linear / Dense) for feature dimensionality reduction and integration, a ReLU activation layer, and a second random dropout layer. Finally, the second fully connected layer outputs the final binary classification label: qualified or unqualified. The final classification result of the model is as follows: Figure 6 As shown, from Figure 6 As can be seen, the accuracy of the model recognition is quite impressive.

[0062] It should also be noted that, in this specification, terms such as "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0063] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for determining the pass / failability of a tensile force-tensile ratio curve based on image recognition, characterized in that, Includes the following steps: S1. Collect tensile force-time series data under multiple sets of preset process parameters; S2. Extract the data of the tensile segment, convert the tensile force-time data into tensile force-tensile ratio data, and plot a curve based on the time series data; S3. Construct a curve graph classification model based on a dual-path feature extraction network; S4. Iteratively train the curve classification model and use the trained curve classification model to determine the pass / failability of the tensile force-tensile ratio curve.

2. The method for determining the pass / failability of a tensile force-tensile ratio curve based on image recognition according to claim 1, characterized in that, Step S1 is as follows: The film biaxial stretching test is carried out according to the preset process parameter group. The film biaxial stretching instrument senses the magnitude of the tensile force through the strain gauge force sensor arranged at the clamp position on the fixture. After the signal is converted by the industrial grade signal amplifier, it enters the A / D acquisition module, collects and saves multiple sets of tensile force-time series data, and uses the upper computer monitoring software to monitor the tensile force-time series data. At the same time, the experimental results of the corresponding process parameter group are recorded, that is, whether the film quality is qualified or not.

3. The method for determining the pass / failability of a tensile force-tensile ratio curve based on image recognition according to claim 1, characterized in that, Step S2, specifically extracting the stretching segment data, includes: S21. The starting point of the stretching segment is determined by the sign of the first derivative of the time series data and a threshold for positive values, specifically as follows: Set a data window based on the density of the data points collected. The size of the data window is adjusted according to the actual situation. Determine the stretching start point based on the probability that the first derivative of the data in the window is positive and whether the value of the derivative reaches the threshold when it is positive. Record the required data points during the process. S22. Determine the number of data points N in the stretching section; based on the preset stretching speed v, the membrane stretching ratio n, the data acquisition frequency f, and the original length l of the membrane, determine the number of data points N in the stretching section, using the following formula: ； Then, N data points starting from the beginning of the stretching are extracted, and the tensile force-time series data of the stretching segment is reset according to the acquisition frequency. S23. Based on the sampling frequency f, the sampling time interval is determined by 1 / f. With 1 / f as the tolerance and 0 as the starting point, the N time data corresponding to the N tensile force data are filled with an arithmetic sequence of time data to finally obtain the tensile force-time series data of the tensile segment.

4. The method for determining the pass / failability of a tensile force-tensile ratio curve based on image recognition according to claim 3, characterized in that, In step S2, the tensile force-time data is converted into tensile force-stretch ratio data, specifically as follows: Let the stretch ratio be The length of the sample after stretching is L. Since synchronous constant speed biaxial stretching is used, the stretching speed v is constant, so L = vt. Therefore, the stretch ratio is... According to the formula, the tensile force-time data is processed by software and converted into tensile force-stretch ratio sequence data.

5. The method for determining the pass / failability of a tensile force-tensile ratio curve based on image recognition according to claim 4, characterized in that, In step S2, a curve is plotted based on the time series data, specifically as follows: Draw a curve graph and scale it uniformly to 224×224 pixels; To avoid interference, coordinate axes and grid lines are not saved; Each curve is labeled according to the experimental results in step S1, and then divided into two categories based on the experimental results, corresponding to qualified and unqualified respectively.

6. The method for determining the pass / failability of a tensile force-tensile ratio curve based on image recognition according to claim 5, characterized in that, In step S3, the curve classification model based on the dual-path feature extraction network includes a shared basic feature extraction layer, an overall trend feature extraction path, and a local mutation feature extraction path. Before entering the curve classification model based on a dual-path feature extraction network, the input image also undergoes data partitioning based on hierarchical five-fold cross-validation, specifically as follows: To fully utilize the limited experimental data and verify the model's stability, a 5K-Fold cross-validation mechanism is used to make the final average accuracy and F1 score more objective. The 5K-Fold method is as follows: The total sample set is divided into 5 subsets, and four of them are selected as the training set and one as the validation set each time. To address class imbalance, a weighted sampling process is employed. By calculating the sampling weight W = 1 / {number of class samples} for each sample, the model automatically increases the frequency of rare samples based on the weights during subsequent training iterations.

7. The method for determining the pass / failability of a tensile force-tensile ratio curve based on image recognition according to claim 6, characterized in that, The shared basic feature extraction layer is specifically as follows: After data partitioning based on the 5K-Fold method, the input image first enters the shared basic feature extraction layer of the ResNet network. In this layer, the image passes through the 7x7 convolutional layer conv1, the batch normalization layer, the modified linear unit (ReLU) activation function, the max pooling layer, and the residual block groups Layer1 and Layer2 of the ResNet network in sequence. Layer1 is used to extract primary textures, and Layer2 is used to extract complex geometric line features; After the input image with a size of 224×224 enters the shared basic feature extraction layer, it is processed from conv1 to Layer2. The spatial size of the feature map is gradually downsampled from 224×224 to a size of 28×28, and the low-level feature map containing basic geometric information is extracted.

8. The method for determining the conformity of tensile force-tensile ratio curve based on image recognition according to claim 7, characterized in that, The specific path for extracting overall trend features is as follows: The overall trend feature extraction path takes the 28×28 low-level feature map obtained after processing by the shared basic feature extraction layer as input; In the overall trend feature extraction path, the image with a spatial size of 28×28 after the features are received by Layer 2 sequentially enters the third residual block group, the first spatial attention module, and the global average pooling layer. After the third residual block group, the spatial scale of the image becomes 14×14, and the resulting feat3 feature map is enhanced after passing through the first spatial attention module. The first spatial attention module focuses attention on the pixel region where the line is located by superimposing average pooling and max pooling, as well as 7x7 convolution and sigmoid activation, and suppresses useless background filling. The feat3 data stream is multiplied with the attention weight mask to obtain the Weighted feat3, which represents the enhancement of the overall topological shape of the curve. Finally, the feature map is compressed to a 1×1 spatial size by a global adaptive average pooling layer, outputting a 256-dimensional overall trend feature vector. The specific path for extracting local mutation features is as follows: The local mutation feature extraction path takes the feature map obtained by the overall trend feature extraction path through the third residual block group as input, and then passes through the fourth residual block group, the second spatial attention module, and the global average pooling layer in sequence. After the fourth residual block group, the spatial scale of the image is further refined to 7×7; the resulting feat4 feature map is enhanced into Weighted feat4 after passing through the second spatial attention module, and finally the feature map is compressed to a spatial size of 1×1 through a global adaptive average pooling layer, while doubling the number of feature channels from 256 to 512, outputting a 512-dimensional local mutation feature vector.

9. The method for determining the conformity of tensile force-tensile ratio curve based on image recognition according to claim 6, characterized in that, Before inputting the image into a curve classification model based on a dual-path feature extraction network, it is normalized and reconstructed. First, it is processed by a pixel linear scaling operator. Discrete grayscale data is mapped to continuous floating-point tensors using channel normalization operators. The data is zero-mean normalized, where the mean vector is... and standard deviation vector The statistical distribution characteristics are taken from the image datasets respectively.

10. The method for determining the pass / failability of a tensile force-tensile ratio curve based on image recognition according to claim 8, characterized in that, The graph classification model based on the dual-path feature extraction network also includes a feature fusion and classification head layer; The feature fusion and classification head layer concatenates the multi-scale features of the two paths, that is, it physically concatenates the 256-dimensional vector of the overall trend feature extraction path and the 512-dimensional vector of the local mutation feature extraction path to form a 768-dimensional full-scale feature vector. The neurons are then passed through a random dropout layer, which randomly discards neurons with a probability of 0.5, reducing the complex dependencies between adjacent layers and thus significantly reducing the risk of overfitting and improving generalization ability. After that, the neurons pass through a first fully connected layer for feature dimensionality reduction and integration, a linear unit ReLU activation layer, and a second random dropout layer. Finally, the second fully connected layer outputs the final binary classification label, i.e., qualified or unqualified.

Images