Real-time quality monitoring method in extruded and pulled product production process

Abstract

The invention relates to the technical field of product detection, in particular to a real-time quality monitoring method in a pultruded product production process, which comprises the following steps: arranging an image acquisition device in a production line, acquiring an image of a pultruded product in real time through the image acquisition device, and processing the acquired image to obtain outer surface information of the pultruded product; a neural network model is constructed, the neural network model is trained by using historical outer surface information of the extruded and pulled product, current outer surface information of the extruded and pulled product is input into the trained neural network model to obtain feature data of the extruded and pulled product, and the neural network model is optimized through a quick connection strategy and a random discarding strategy; analyzing the feature data, and performing quality judgment by using a classifier to obtain a quality judgment report of the extruded and pulled product; according to the invention, through synchronous detection in the production process, online continuous detection is realized, the labor cost is greatly saved, and meanwhile, the influence of human factors in detection of inspectors is avoided.

Description

Technical Field

[0001] The invention relates to the technical field of product detection, and in particular to a real-time quality monitoring method in a production process of an extruded product. Background Art

[0002] Compared with traditional materials, fiber-reinforced resin-based composite materials have many advantages, such as high specific strength, light weight, high specific modulus, good fatigue resistance and good vibration reduction performance. The various components of the composite material play a synergistic role in performance, and have superior comprehensive performance that a single material cannot match.

[0003] There are many ways to form fiber-reinforced resin-based composite materials, among which pultrusion is an efficient production method that can achieve large-scale, continuous production. At present, compared with traditional processing methods, pultrusion has higher production efficiency, and the pultrusion speed can reach 0.4-1.2m / min, which may meet the market demand for a large number of high-quality composite products. At the same time, it also brings huge pressure to the inspection and testing of the appearance, size, straightness, etc. of pultruded products. Summary of the invention

[0004] In view of this, the present invention aims to propose a real-time quality monitoring method in the production process of extruded products, in order to solve at least one of the above-mentioned technical problems.

[0005] To achieve the above object, the technical solution of the present invention is achieved as follows: A real-time quality monitoring method in a pultruded product production process, the method comprising: An image acquisition device is arranged in the production line, and images of the pultruded product are acquired in real time by the image acquisition device, and the acquired images are processed to obtain the outer surface information of the pultruded product; Constructing a neural network model, using historical outer surface information of the pultruded product to train the neural network model, inputting current outer surface information of the pultruded product into the trained neural network model to obtain characteristic data of the pultruded product, wherein the neural network model is optimized by a shortcut connection strategy and a random discard strategy; The characteristic data is analyzed, and a classifier is used to make a quality judgment, thereby obtaining a quality judgment report of the pultruded product.

[0006] Furthermore, the quick connection strategy specifically includes: The neural network model includes multiple convolutional layers, which are divided into multiple blocks. Each block extracts features from input data. By establishing a shortcut connection channel between the head and tail of the block, the input of the block is transmitted to the output of the block.

[0007] Furthermore, the random discard strategy specifically includes: A discard rate is preset, and a corresponding mask vector is generated for each number of neurons in the neural network model. The output of each neuron is multiplied element by element by the corresponding mask vector to complete random discarding, and the gradient of the neurons that are not discarded is calculated and the weight is updated; wherein the mask vector is obtained by independent sampling of the Bernoulli distribution, and the non-zero expected value of the mask vector is 1-discarding rate.

[0008] Furthermore, the process of analyzing the characteristic data includes: A feature subset is constructed by iteratively adding or removing features, and the performance of the subset is evaluated in each iteration and the feature with the best performance is selected to obtain an optimal subset of feature data; wherein the feature with the best performance is selected through a forward strategy or a backward strategy.

[0009] Furthermore, the forward strategy specifically includes: Starting from an empty feature set, add one feature at a time to the features outside the current feature subset to form a new feature subset. Use the support vector machine algorithm to evaluate the performance of the feature subset, and use the best-performing feature subset as the current optimal subset. When adding any new features cannot improve the model performance, terminate the selection process.

[0010] Furthermore, the backward strategy specifically includes: Starting from the full feature set, remove one feature at a time from the current feature subset to form a new feature subset. Use the support vector machine algorithm to evaluate the performance of the feature subset, and use the best-performing feature subset as the current optimal subset. When removing any feature cannot improve the model performance, terminate the selection process.

[0011] Furthermore, the process of using a classifier to make a quality judgment includes: analyzing the extracted feature data, using a classifier to make a quality judgment, and evaluating the quality status of each feature data.

[0012] Furthermore, the process of using the classifier to perform quality determination includes: A linear model is established, and the score is calculated through the linear combination of features. The sigmoid function is used to convert the score into a probability value to indicate the probability that the sample belongs to a certain category. The logarithmic loss function is used to evaluate the accuracy of the model prediction.

[0013] Furthermore, the process of evaluating the quality status of each sample includes: Construct a confusion matrix, calculate multiple performance indicators based on the data in the confusion matrix, evaluate the performance of the classification model, perform in-depth analysis based on the ROC curve and AUC value, and complete the quality status assessment of the feature data.

[0014] Compared with the prior art, the present invention has the following beneficial effects: Synchronous testing during the production process realizes online continuous testing, which greatly saves labor costs and avoids the influence of human factors in the inspection by inspectors. The test data is stored in the computer and can be directly converted into the product factory inspection report, avoiding the occasional input errors that occur when inspectors make manual reports, greatly improving the factory inspection efficiency, result accuracy and traceability of pultruded products. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 Schematic diagram of the workflow of the real-time quality monitoring method in the production process of pultruded products described in an embodiment of the present invention. DETAILED DESCRIPTION

[0016] The following will be combined with the embodiments of the present invention to clearly and completely describe the technical solution of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0017] like Figure 1 As shown, this embodiment provides a real-time quality monitoring method in the production process of pultruded products, the method comprising: S1. An image acquisition device is arranged in the production line, and images of the pultruded product are acquired in real time by the image acquisition device, and the acquired images are processed to obtain the outer surface information of the pultruded product; S2. Constructing a neural network model, using historical outer surface information of the pultruded product to train the neural network model, inputting current outer surface information of the pultruded product into the trained neural network model to obtain characteristic data of the pultruded product, wherein the neural network model is optimized by a shortcut connection strategy and a random discard strategy; S3. Analyze the characteristic data and use a classifier to make quality judgments to obtain a quality judgment report of the extruded product.

[0018] The shortcut connection strategy specifically includes: the neural network model includes multiple convolutional layers, the multiple convolutional layers are divided into multiple blocks, and each block extracts features from the input data; by establishing a shortcut connection channel between the head and the tail of the block, the input of the block is passed to the output of the block.

[0019] The core idea of ​​the shortcut connection strategy is to pass the input signal directly to the subsequent layers through shortcut connections, so that the network can learn residuals instead of global features. Specifically, it introduces a cross-layer connection (also called skip connection) in each block to add the input signal directly to the output of the block. This design makes it easier for the network to transfer gradients during backpropagation, thereby solving the gradient vanishing problem in deep network training.

[0020] In some embodiments, the specific workflow of the quick connection strategy is as follows: 1) Construct the input layer of the neural network: The input is an image of an extruded product with a size of H×W×C, where H and W are the height and width of the image, and C is the number of channels (usually 1, representing a grayscale image). The image is standardized so that each pixel value is between 0 and 1. It is usually normalized by subtracting the mean of the dataset and dividing by the standard deviation.

[0021] 2) Construct the convolutional layer of the neural network: The first convolutional layer uses a larger convolution kernel (7*7) and a larger stride (2) to quickly reduce the size of the feature map. The formula is: Y=f(X*W+b); where X is the input image, W is the convolution kernel, b is the bias, f is the activation function, and * represents the convolution operation.

[0022] 3) Constructing the pooling layer of the neural network The maximum pooling (3*3 pooling kernel, stride 2) is used to further reduce the size of the feature map while retaining important features. The formula is expressed as: ; Where m and n are the sizes of the pooling windows, i represents the row index of the feature map, and j represents the column index of the feature map.

[0023] 4) Divide multiple convolutional layers into multiple blocks: Each block consists of two or three convolutional layers, and the input is directly added to the output to form a shortcut connection. For a two-layer block, the calculation is as follows: ; ; Among them, W 1 and W 2 is the convolution kernel, b 1 and b 2 is bias.

[0024] For a block containing three convolutional layers, the calculation is as follows: ; Among them, W 1 and W 3The convolution kernel is usually 1×1, which is used for dimensionality reduction and dimensionality increase. 2 The convolution kernel is 3×3.

[0025] 5) Perform global average pooling: After the last block, a global average pooling layer is used to reduce the size of each feature map to 1 and obtain a feature vector. The formula is: ; Among them, Y c is the output of channel c, X ijc is the pixel value of the feature map.

[0026] The random discard strategy specifically includes: A discard rate is preset, and a corresponding mask vector is generated for each number of neurons in the neural network model. The output of each neuron is multiplied element by element by the corresponding mask vector to complete random discarding, and the gradient of the neurons that are not discarded is calculated and the weight is updated; wherein the mask vector is obtained by independent sampling of the Bernoulli distribution, and the non-zero expected value of the mask vector is 1-discarding rate.

[0027] In some embodiments, the specific execution process of the above random discard strategy is as follows: 1) Generate mask vector: For the output of each layer, a mask vector with the same shape as the output is generated. Each element in the mask vector is a binary value independently sampled from a Bernoulli distribution with a non-zero expected value of 1-p, where p is the dropout rate.

[0028] 2) Apply the mask vector: Multiply the mask vector by the output of the neuron element by element. If a position in the mask vector is 0, the corresponding neuron output will also be set to 0 (that is, the neuron is discarded). If a position in the mask vector is 1, the corresponding neuron output remains unchanged.

[0029] 3) Scaling output (training phase): During the training phase, the output is not scaled because the dropped neurons are already implemented by the mask. To simplify the calculations in the test phase, the output of the training phase may be scaled by multiplying it by 1 / (1-p) after applying the mask to keep the overall output expectation unchanged.

[0030] 4) Adjustments during the testing phase: In the test phase, no neurons are dropped. In order to maintain the same expected activation value as in the training phase, the outputs of all neurons are multiplied by 1-p. This can be seen as compensation for the lack of scaling in the training phase, making the network performance consistent in the training and test phases.

[0031] The random dropout strategy effectively prevents the neural network from over-relying on specific neurons or features, improving the robustness and generalization ability of the model. Each iteration trains a different sub-network, and the final effect is equivalent to integrating the feature learning results of multiple different networks.

[0032] The process of analyzing the characteristic data includes: A feature subset is constructed by iteratively adding or removing features, and the performance of the subset is evaluated in each iteration and the feature with the best performance is selected to obtain an optimal subset of feature data; wherein the feature with the best performance is selected through a forward strategy or a backward strategy.

[0033] The forward strategy specifically includes: Starting from an empty feature set, add one feature at a time to the features outside the current feature subset to form a new feature subset. Use the support vector machine algorithm to evaluate the performance of the feature subset, and use the best-performing feature subset as the current optimal subset. When adding any new features cannot improve the model performance, terminate the selection process.

[0034] In some embodiments, the execution process of the above forward strategy is as follows: Extracting a set of initial features from the surface image of the pultruded product, such as texture features, color features, shape features, etc.; Starting from an empty feature set, for each unselected feature, temporarily add it to the current feature subset, use the support vector machine algorithm to train on the training set, evaluate the accuracy of the current feature subset, select the feature with the highest accuracy to add to the current feature subset, compare the performance of the new feature subset with the previous best subset, if the performance is improved, update the current best feature subset; otherwise, end the selection process.

[0035] The backward strategy specifically includes: Starting from the full feature set, remove one feature at a time from the current feature subset to form a new feature subset. Use the support vector machine algorithm to evaluate the performance of the feature subset, and use the best-performing feature subset as the current optimal subset. When removing any feature cannot improve the model performance, terminate the selection process.

[0036] In some embodiments, the execution process of the above backward strategy is as follows: Extracting a set of initial features from the surface image of the pultruded product, such as texture features, color features, shape features, etc.; Starting from the full feature set, temporarily remove each feature in the current feature subset, use the support vector machine algorithm to train on the training set, evaluate the performance of the current feature subset on the validation set, select the feature subset that makes the model perform best after removal as the current subset, compare the performance of the new feature subset with the previous best subset, if the performance is improved, update the current best feature subset; otherwise, end the elimination process.

[0037] The process of using the classifier to perform quality judgment includes: The extracted feature data is analyzed, the quality is judged using a classifier, and the quality status of each sample is evaluated by the probability value output by the classifier.

[0038] The process of using the classifier to perform quality determination includes: A linear model is established, and the score is calculated through the linear combination of features. The sigmoid function is used to convert the score into a probability value to indicate the probability that the sample belongs to a certain category. The logarithmic loss function is used to evaluate the accuracy of the model prediction.

[0039] In some embodiments, the steps of using a classifier to perform quality determination are as follows: Assume that the input feature vector is , the linear combination of the models is , where w is the weight vector and b is the bias.

[0040] Use the sigmoid function to convert the linear combination results into probability values: ;in, Indicates the probability that the sample belongs to the positive class (such as qualified).

[0041] Log Loss is used to measure the accuracy of model predictions: ; in, is the actual label, is the predicted probability and m is the number of samples.

[0042] The process of evaluating the quality status of each sample includes: Construct a confusion matrix, calculate multiple performance indicators based on the data in the confusion matrix, evaluate the performance of the classification model, perform in-depth analysis based on the ROC curve and AUC value, and complete the quality status assessment of the feature data.

[0043] In some embodiments, the steps of evaluating the quality status of each sample are performed as follows: Construct a confusion matrix, which is a tool used to evaluate the performance of classification models, especially for binary and multi-classification problems. It can clearly show the correct and incorrect predictions of the model; For a binary classification problem, the confusion matrix consists of four elements: True Positives (TP): The number of samples correctly predicted by the model as positive.

[0044] False Positives (FP): The number of samples that the model incorrectly predicts as positive.

[0045] True Negatives (TN): The number of samples that the model correctly predicts as negative.

[0046] False Negatives (FN): The number of samples that the model incorrectly predicts as negative.

[0047] Based on the data in the confusion matrix, a variety of performance indicators can be calculated to evaluate the effectiveness of the classification model: Accuracy: , which indicates the proportion of samples predicted correctly by the model.

[0048] Accuracy , which represents the proportion of samples predicted to be positive that are actually positive.

[0049] Recall , which represents the proportion of samples that are actually positive that are correctly predicted to be positive.

[0050] F1 score , which is the harmonic mean of precision and recall, and is applicable to data with imbalanced categories.

[0051] The ROC curve describes the performance of the model at different thresholds. The horizontal axis is the false positive rate (FPR) and the vertical axis is the true positive rate (TPR). By adjusting the threshold, the trade-off between the sensitivity and specificity of the model can be observed.

[0052] Calculate the area under the ROC curve. The closer the AUC value is to 1, the stronger the classification ability of the model.

[0053] Obviously, the above embodiments are merely examples for the purpose of clear explanation, and are not intended to limit the implementation methods. For those skilled in the art, other different forms of changes or modifications can be made based on the above description. It is not necessary and impossible to list all the implementation methods here. The obvious changes or modifications derived therefrom are still within the scope of protection of the invention.

Claims

1. A real-time quality monitoring method in the production process of pultruded products, characterized in that: The method comprises: An image acquisition device is arranged in the production line, and images of the pultruded product are acquired in real time by the image acquisition device, and the acquired images are processed to obtain the outer surface information of the pultruded product; Constructing a neural network model, using historical outer surface information of the pultruded product to train the neural network model, inputting current outer surface information of the pultruded product into the trained neural network model to obtain characteristic data of the pultruded product, wherein the neural network model is optimized by a shortcut connection strategy and a random discard strategy; The characteristic data is analyzed, and a classifier is used to make a quality judgment, thereby obtaining a quality judgment report of the pultruded product.

2. The real-time quality monitoring method in the production process of pultruded products according to claim 1, characterized in that: The quick connection strategy specifically includes: The neural network model includes multiple convolutional layers, which are divided into multiple blocks. Each block extracts features from input data. By establishing a shortcut connection channel between the head and tail of the block, the input of the block is transmitted to the output of the block.

3. The real-time quality monitoring method in the production process of pultruded products according to claim 1, characterized in that: The random discard strategy specifically includes: A discard rate is preset, and a corresponding mask vector is generated for each number of neurons in the neural network model. The output of each neuron is multiplied element by element by the corresponding mask vector to complete random discarding, and the gradient of the neurons that are not discarded is calculated and the weight is updated; wherein the mask vector is obtained by independent sampling of the Bernoulli distribution, and the non-zero expected value of the mask vector is 1-discarding rate.

4. The real-time quality monitoring method in the production process of pultruded products according to claim 1, characterized in that: The process of analyzing the characteristic data includes: A feature subset is constructed by iteratively adding or removing features, and the performance of the subset is evaluated in each iteration and the feature with the best performance is selected to obtain an optimal subset of feature data; wherein the feature with the best performance is selected through a forward strategy or a backward strategy.

5. The real-time quality monitoring method in the production process of pultruded products according to claim 4, characterized in that: The forward strategy specifically includes: Starting from an empty feature set, add one feature at a time to the features outside the current feature subset to form a new feature subset. Use the support vector machine algorithm to evaluate the performance of the feature subset, and use the best-performing feature subset as the current optimal subset. When adding any new features cannot improve the model performance, terminate the selection process.

6. The real-time quality monitoring method in the production process of pultruded products according to claim 4, characterized in that: The backward strategy specifically includes: Starting from the full feature set, remove one feature at a time from the current feature subset to form a new feature subset. Use the support vector machine algorithm to evaluate the performance of the feature subset, and use the best-performing feature subset as the current optimal subset. When removing any feature cannot improve the model performance, terminate the selection process.

7. The real-time quality monitoring method in the production process of pultruded products according to claim 1, characterized in that: The process of using the classifier to make quality judgments includes: analyzing the extracted feature data, using the classifier to make quality judgments, and evaluating the quality status of each feature data.

8. The real-time quality monitoring method in the production process of pultruded products according to claim 7, characterized in that: The process of using the classifier to perform quality determination includes: A linear model is established, and the score is calculated through the linear combination of features. The sigmoid function is used to convert the score into a probability value to indicate the probability that the sample belongs to a certain category. The logarithmic loss function is used to evaluate the accuracy of the model prediction.

9. The real-time quality monitoring method in the production process of pultruded products according to claim 7, characterized in that: The process of evaluating the quality status of each sample includes: Construct a confusion matrix, calculate multiple performance indicators based on the data in the confusion matrix, evaluate the performance of the classification model, perform in-depth analysis based on the ROC curve and AUC value, and complete the quality status assessment of the feature data.

Images