Product quality detection method and system for extruder

Abstract

The invention provides a product quality detection method and system for an extruder, and the method comprises the steps: collecting image data of the surface and interior of a product in real time, and installing an intelligent sensor at the same time to collect multi-modal sensing data in the production process; processing the surface image data by using a convolutional neural network, identifying surface defects and internal defects in real time, and integrating the surface defects and the internal defects to obtain a final defect graph; feature extraction is carried out from the stored data and defect graphs, and real-time monitoring and anomaly detection are carried out; constructing a self-adaptive control model, and adjusting operation parameters of the extruder in real time according to an analysis result; and multiple machine learning algorithms are selected for integration, and prediction results of multiple models are fused by using a weighted average method to obtain an optimal product quality detection result. According to the invention, high-precision detection of surface and internal defects of an extruded product can be realized, and process parameters in the extrusion process can be monitored and optimized in real time, so that the consistency of product quality and the production efficiency are greatly improved.

Description

Technical Field

[0001] The invention belongs to the technical field of product quality detection, and in particular relates to a product quality detection method and system for an extruder. Background Art

[0002] Extruders are widely used in industries such as plastics, rubber, and food to extrude raw materials into various products. Traditional methods for testing the quality of extruders mainly rely on manual inspection and simple optical inspection equipment. Although these methods can identify surface defects of products to a certain extent, they have many shortcomings. First, manual inspection relies on the experience of operators and is easily affected by subjective factors, resulting in unstable and inefficient inspection results. Secondly, traditional optical inspection equipment mostly uses a single imaging technology, which can only detect surface defects of products, and cannot effectively identify internal defects such as bubbles and uneven density inside the product. In addition, most of the existing inspection equipment lacks intelligent means and cannot achieve real-time monitoring and adaptive adjustment of the extrusion process. This leads to inconsistency in the quality of extruded products, increases scrap rates and production costs, and affects the competitiveness of enterprises.

[0003] In order to overcome the above problems, some researchers have tried to introduce computer vision and machine learning technologies to improve detection accuracy and efficiency in recent years. However, these methods still have some limitations. For example, relying solely on surface images for defect detection cannot fully understand the internal quality of the product; traditional machine learning algorithms are highly dependent on data, and the generalization ability of the model is limited; in actual production environments, data noise and environmental changes have a greater impact on the detection results. In addition, the existing quality inspection systems are mostly independent inspection modules, lacking linkage control with the overall process of the extruder, and cannot achieve real-time optimization and intelligent adjustment of the extrusion process. Summary of the invention

[0004] The purpose of the present invention is to design a product quality detection method and system for an extruder. By introducing advanced artificial intelligence and sensor technology, the present invention can not only achieve high-precision detection of surface and internal defects of extruded products, but also monitor and optimize process parameters in the extrusion process in real time, thereby greatly improving product quality consistency and production efficiency.

[0005] In order to achieve the above object, the present invention provides a product quality detection method for an extruder in a first aspect, the method comprising:

[0006] S1. Install high-resolution industrial cameras and thermal imagers at key locations of the extruder to collect real-time image data of the product surface and interior respectively. Install intelligent sensors to collect multimodal sensor data during the production process, pre-process and normalize the image data and multimodal sensor data, synchronize the processed data to a unified timeline, and use a distributed database to store the synchronized data;

[0007] S2. Use a convolutional neural network to process the surface image data, identify surface defects in real time based on the processing results of the surface image data, use a deep learning algorithm to analyze the thermal imaging data, identify internal defects based on the analysis results of the thermal imaging data, and integrate the surface defects and internal defects to obtain the final defect map;

[0008] S3, extracting features from the stored data and defect graph, building a multi-layer perception model to generate a monitoring state vector, and using a long short-term memory network model to perform real-time monitoring and anomaly detection on the monitoring state vector. When an anomaly is detected, the anomaly information and specific feature vectors are fed back to the extruder control system, and the control system adjusts the operating parameters of the extruder according to the feedback information;

[0009] S4, based on the real-time monitoring data and comparing and analyzing with the historical data, construct an adaptive control model to adjust the operating parameters of the extruder in real time according to the analysis results;

[0010] S5. Extract key features from the stored data and real-time monitoring data, select multiple machine learning algorithms for integration and use the weighted average method to fuse the prediction results of multiple models, use the optimization algorithm to predict and optimize the extrusion parameters, and combine the operating parameters adjusted in real time by the adaptive control model to obtain the optimal extruder operating parameter combination and the optimal product quality test results;

[0011] Wherein, the S5 includes:

[0012] S501, extract key features from the stored data and real-time monitoring data, and construct a feature vector X(t); wherein the key features include temperature T prod 、Pressure P extr and speed V extr ; The feature vector X(t) is expressed as follows:

[0013] X(t)=[T prod (t), P extr (t), v extr (t), ΔT prod (t), ΔP extr (t), Δv extr (t),…]

[0014] Among them, Tprod (t) represents the extruder production temperature at time t, P extr (t) represents the extrusion pressure at time t, V extr (t) represents the extrusion speed at time t, ΔT prod (t) represents the rate of change of temperature, expressed as the temperature change from time t to the previous time divided by the time interval, ΔP extr (t) represents the rate of change of pressure, expressed as the pressure change between time t and the previous time divided by the time interval, Δv extr (t) represents the rate of change of speed, which is expressed as the change in speed from time t to the previous time divided by the time interval;

[0015] The product quality test results are used as labels Y(t):

[0016] Y(t)=[D surf (t), D int (t)]

[0017] Among them, D surf (t) represents the surface defect detection result at time t, D int (t) represents the internal defect detection result at time t;

[0018] S502, selecting multiple machine learning models for integration to obtain an integrated model, wherein the multiple machine learning models include a support vector machine, a random forest, and a gradient boosting decision tree;

[0019] S503, using a weighted average method to fuse the prediction results of multiple models of the integrated model;

[0020] S504, using incremental SVM to perform online learning and real-time optimization on the integrated model, as shown below:

[0021]

[0022] Among them, w t+1 represents the updated weight vector at time t+1; w t represents the weight vector at time t; η represents the learning rate; y t represents the actual label at time t; σ represents the activation function; x t Represents the feature vector at time t;

[0023] S505. According to the prediction result of the optimized integrated model, the operating parameters of the extruder are adjusted in real time to achieve the best product quality; wherein the real-time adjustment of the operating parameters of the extruder is expressed as follows:

[0024]

[0025] Among them, α T , α P and α V are the adjustment coefficients for temperature, pressure and speed, respectively. represents the prediction result of the integrated model at time t; Y(t) represents the actual detection result at time t; T prod (t+Δt) represents the adjusted temperature at time t+Δt; P extr (t+Δt) represents the adjusted pressure at time t+Δt; v extr (t+Δt) represents the adjusted speed at time t+Δt;

[0026] S506. Regularly evaluate the performance of the integrated model using mean square error and coefficient of determination as evaluation indicators.

[0027] Furthermore, the intelligent sensor includes a temperature sensor, a pressure sensor and a displacement sensor;

[0028] The preprocessing includes: denoising the collected surface image and smoothing it with a Gaussian filter; calibrating and denoising the thermal imaging data; using temperature sensors, pressure sensors and displacement sensors to monitor the parameters of the extrusion process in real time, collecting data at a fixed frequency and transmitting it to the edge computing device, denoising the sensor data, and smoothing it with a Kalman filter.

[0029] Furthermore, the step S2 specifically includes:

[0030] S201, obtaining surface image data I' from S1 surf (x, y), divide each image into blocks, calculate the gray level co-occurrence matrix of each sub-block, obtain the texture features of each sub-block, and classify each sub-block using a support vector machine according to the texture features of each sub-block to determine whether there are defects; wherein the texture features include contrast, homogeneity and energy;

[0031] S202, obtaining internal thermal imaging data T′ from S1 int (x, y), use the Canny edge detection algorithm to perform edge detection on each thermal image, use the watershed algorithm to segment the image based on the edge results, divide the thermal image into different areas, and calculate the average temperature T of each area avg and temperature standard deviation T std ,Using the logistic regression model, each area is classified to determine whether there is an internal defect;

[0032] S203, obtain the confidence map C of the surface defect through the output probability of the classifier surf (x, y) and confidence map of internal defects Cint (x, y), the surface defect confidence map C surf (x, y) and internal defect confidence map C int (x, y) is fused using the Bayesian weighted fusion method to obtain the comprehensive confidence map C comb (x, y), for the comprehensive defect map D final (x, y) is post-processed to extract the boundary and shape of the defect area, the processed defect information is displayed through a visual interface, and is fed back to the control system in real time through the edge computing device.

[0033] Furthermore, the image is divided into m×n sub-blocks, and the gray-level co-occurrence matrix C(i, j) is calculated as follows:

[0034]

[0035] Among them, 6 represents the indicator function, which takes the value of 1 when the condition is true, otherwise it takes the value of 0; d x and d y Indicates the offset of adjacent pixels; I′ surf (x, y) represents the smoothed surface image, I′ surf (x+d x ,y+d y ) represents a smoothed surface image based on the offset of adjacent pixels;

[0036] Calculate the texture features of each sub-block, including contrast, homogeneity and energy, as follows:

[0037]

[0038] Use support vector machine to classify each sub-block to determine whether there are defects:

[0039] Training set T = {(F k , yk)} is given by the eigenvector F k = [contrast, homogeneity, energy] and the corresponding label yk∈{0,1}, where 0 means no defect and 1 means defect;

[0040] Classify each sub-block to obtain the defect map D surf (x, y), a value of 1 indicates a defect, and a value of 0 indicates no defect.

[0041] Furthermore, the S202 specifically includes:

[0042] Calculate the gradient strength and direction of the internal image, expressed as follows:

[0043]

[0044] Among them, G x , G y represents the corresponding coordinates of the internal image gradient intensity, and G represents the direction of the internal image;

[0045] Non-maximum suppression, thinning edges, is expressed as follows:

[0046]

[0047] Apply double threshold detection to determine the true edge, which is expressed as follows:

[0048]

[0049] Among them, E′(x, y) represents the real edge;

[0050] Use the watershed algorithm to segment the image and divide the thermal image into different areas:

[0051] Label(x,y)=watershed(E′(x,y))

[0052] Among them, Label(x, y) represents the segmented area, and watershed represents the watershed function;

[0053] Calculate the average temperature T of each area avg and temperature standard deviation T std :

[0054]

[0055] Where N represents the number of samples;

[0056] Use the logistic regression model to classify each area to determine whether there are internal defects:

[0057] Training set R = {(G k , z k )} By the eigenvector G k =[T avg , T std ] and the corresponding label z k ∈{0, 1}, where 0 means no defect and 1 means defect;

[0058] Classify each area to get the defect map D int (x, y), a value of 1 indicates a defect, and a value of 0 indicates no defect.

[0059] Furthermore, the S203 specifically includes:

[0060] The confidence map C of the surface defect is calculated by the output probability of the classifier surf (x, y) and confidence map of internal defects C int (x, y) confidence, expressed as follows:

[0061] C surf (x, y) = SVM_prob(F k )

[0062] C int (x, y) = LogReg_prob(G k )

[0063] The surface defect confidence map C is transformed into surf (x, y) and internal defect confidence map C int (x, y) are fused to obtain the comprehensive confidence map C comb (x, y), expressed as follows:

[0064]

[0065] Among them, P surf and P int denote the prior probabilities of surface defects and internal defects respectively;

[0066] According to the comprehensive confidence map C comb (x, y) generates the final defect map D final (x, y), using threshold T comb Binarization is performed, which is expressed as follows:

[0067]

[0068] Use morphological operations to analyze the integrated defect map D final (x, y) is post-processed to extract the boundary and shape of the defect area; wherein the post-processing is to clean up the noise, and the formula is as follows:

[0069] D morp (x, y) = dilate(erode(D final (x,y),SE),SE)

[0070] Among them, erode and dilate represent morphological erosion and dilation operations respectively, and SE represents the structural element;

[0071] The processed defect information is displayed through a visual interface and fed back to the control system in real time through edge computing devices.

[0072] Furthermore, in S3, the real-time monitoring and anomaly detection specifically include:

[0073] A multi-layer perception model is established to fuse the extracted feature data and generate the overall monitoring state vector S(t), which is expressed as follows:

[0074] S(t)=[F surf , F int , ΔD temp (t), ΔD press (t), ΔD disp (t)]

[0075] Among them, F surf The feature vector representing the surface image, F int The characteristic vector representing the thermal imaging data, ΔD temp (t), ΔD press (t) and ΔD disp (t) represents the rate of change of temperature, pressure and displacement data respectively.

[0076] The long short-term memory network model is used for real-time monitoring and anomaly detection. The model input is the current and historical state vector S(t), and the output is the anomaly score A(t), which is expressed as follows:

[0077] A(t)=LSTM(S(t-τ), S(t-τ+1),…,S(t))

[0078] Where T represents the time window length.

[0079] According to the anomaly score A(t), it is judged whether the system has an anomaly. When A(t) exceeds the threshold T A When an alarm is triggered, the abnormal information is recorded, as shown below:

[0080] if A(t)>T A then trigger alarm

[0081] Among them, T A Threshold for scoring anomalies;

[0082] Among them, when an abnormality is detected, the abnormal information and specific feature vectors are fed back to the extruder control system, and the control system adjusts the operating parameters of the extruder according to the feedback information to correct the abnormality, as shown below:

[0083] ΔP extr =K p ΔD press (t)

[0084] ΔT prod =K t ΔD temp (t)

[0085] Δv extr=K v ΔD disp (t)

[0086] Among them, K p , K t and K v Represents the control gain coefficients for pressure, temperature and speed adjustment, ΔP extr , ΔT prod and ΔV extr Respectively represent the adjustment amount of extruder pressure, temperature and speed.

[0087] Further, in S4, the adaptive control model is constructed as follows:

[0088] According to the state error E(t), the control target C(t) is defined as follows:

[0089] C(t)=K c ·E(t)

[0090] Among them, K c represents the control gain matrix, which is used to adjust the effect of the error on the control target;

[0091] Decompose the control target C(t) into the temperature control target C T (t), pressure control target C P (t) and speed control target C V (t):

[0092] C T (t) = [c T1 (t), c T2 (t),…,c Ti (t), c Tn (t)]

[0093] C P (t) = [c P1 (t), c P2 (t),…,c Pi (t), c Pn (t)]

[0094] C V (t) = [c V1 (t), c V2 (t), ..., c Vi (t), c Vn (t)]

[0095] Among them, c Ti (t), c Pi (t) and c Vi (t) are the target values ​​of each control variable at time t.

[0096] Furthermore, a PID controller is designed for each control target, and the operating parameters of the extruder are adjusted to achieve the control target, wherein the temperature PID controller formula is expressed as follows:

[0097]

[0098] e T (t) = C T (t)-T prod (t)

[0099] Among them, u T (t) represents the temperature control input, e T (t) represents the temperature error, K pT , K iT and K dT They represent the proportional, integral and derivative gains of temperature control respectively;

[0100] The pressure PID controller formula is as follows:

[0101]

[0102] e P (t) = C P (t)-P extr (t)

[0103] Among them, u P (t) represents the pressure control input, e P (t) represents the pressure error, K pP , K iP and K dP They represent the proportional, integral and differential gains of pressure control respectively;

[0104] The speed PID controller formula is as follows:

[0105]

[0106] e V (t) = C V (t)-v extr (t)

[0107] Among them, u V (t) represents the speed control input, e V (t) represents the speed error, K pv , K iV and K dV Respectively represent the proportional, integral and differential gains of speed control;

[0108] Control input u T (t),u P (t) and uV (t) is converted into actual control instructions to adjust the operating parameters of the extruder, as shown below:

[0109] P extr (t+Δt)=P extr (t)+u P (t)

[0110] T prod (t+Δt)=T prod (t)+u T (t)

[0111] v extr (t+Δt)=v extr (t)+u V (t)

[0112] Among them, P extr (t+Δt), T prod (t+Δt) and v extr (t+Δt) represent the adjusted pressure, temperature and speed respectively.

[0113] In a second aspect of the present invention, a product quality detection system for an extruder is provided, the system comprising:

[0114] The data acquisition module is used to install high-resolution industrial cameras and thermal imagers at key positions of the extruder to collect real-time image data of the product surface and interior respectively. At the same time, intelligent sensors are installed to collect multimodal sensor data during the production process, and the image data and multimodal sensor data are pre-processed and normalized. The processed data are synchronized to a unified timeline and the synchronized data is stored using a distributed database;

[0115] The defect analysis module is used to process the surface image data using a convolutional neural network, identify surface defects in real time based on the processing results of the surface image data, analyze the thermal imaging data using a deep learning algorithm, identify internal defects based on the analysis results of the thermal imaging data, and integrate surface defects and internal defects to obtain the final defect map;

[0116] The quality inspection module is used to extract features from the stored data and defect graphs, build a multi-layer perception model to generate a monitoring state vector, and use a long short-term memory network model to perform real-time monitoring and anomaly detection on the monitoring state vector. When an anomaly is detected, the anomaly information and specific feature vectors are fed back to the extruder control system, and the control system adjusts the operating parameters of the extruder according to the feedback information;

[0117] The parameter adjustment module is used to compare and analyze the real-time monitoring data with the historical data, and to build an adaptive control model to adjust the operating parameters of the extruder in real time according to the analysis results;

[0118] The quality inspection operation optimization module is used to extract key features from the stored data and real-time monitoring data, select multiple machine learning algorithms for integration and use the weighted average method to fuse the prediction results of multiple models, use the optimization algorithm to predict and optimize the extrusion parameters, and combine the operating parameters adjusted in real time by the adaptive control model to obtain the optimal extruder operating parameter combination and the optimal product quality inspection results;

[0119] The steps performed by the quality inspection operation optimization module include:

[0120] S501, extract key features from the stored data and real-time monitoring data, and construct a feature vector X(t); wherein the key features include temperature T prod 、Pressure P extr and speed v extr ; The feature vector X(t) is expressed as follows:

[0121] X(t)=[T prod (t), P extr (t), v extr (t), ΔT prod (t), ΔP extr (t), Δv extr (t),…]

[0122] Among them, T prod (t) represents the extruder production temperature at time t, P extr (t) represents the extrusion pressure at time t, v extr (t) represents the extrusion speed at time t, ΔT prod (t) represents the rate of change of temperature, expressed as the temperature change from time t to the previous time divided by the time interval, ΔP extr (t) represents the rate of change of pressure, expressed as the pressure change between time t and the previous time divided by the time interval, Δv extr (t) represents the rate of change of speed, which is expressed as the change in speed from time t to the previous time divided by the time interval;

[0123] The product quality test results are used as labels Y(t):

[0124] Y(t)=[D surf (t), D int (t)]

[0125] Among them, D surf (t) represents the surface defect detection result at time t, D int (t) represents the internal defect detection result at time t;

[0126] S502, selecting multiple machine learning models for integration to obtain an integrated model, wherein the multiple machine learning models include a support vector machine, a random forest, and a gradient boosting decision tree;

[0127] S503, using a weighted average method to fuse the prediction results of multiple models of the integrated model;

[0128] S504, using incremental SVM to perform online learning and real-time optimization on the integrated model, as shown below:

[0129]

[0130] Among them, w t+1 represents the updated weight vector at time t+1; w t represents the weight vector at time t; η represents the learning rate; y t represents the actual label at time t; σ represents the activation function; x t Represents the feature vector at time t;

[0131] S505. According to the prediction result of the optimized integrated model, the operating parameters of the extruder are adjusted in real time to achieve the best product quality; wherein the real-time adjustment of the operating parameters of the extruder is expressed as follows:

[0132]

[0133] Among them, α T , α P and α V are the adjustment coefficients for temperature, pressure and speed, respectively. represents the prediction result of the integrated model at time t; Y(t) represents the actual detection result at time t; T prod (t+Δt) represents the adjusted temperature at time t+Δt; P extr (t+Δt) represents the adjusted pressure at time t+Δt; v extr (t+Δt) represents the adjusted speed at time t+Δt;

[0134] S506. Regularly evaluate the performance of the integrated model using mean square error and coefficient of determination as evaluation indicators.

[0135] The beneficial technical effects of the present invention are at least as follows:

[0136] (1) The present invention adopts a multi-level defect detection system based on computer vision and deep learning. By installing multiple high-resolution industrial cameras and thermal imagers at the extruder outlet and cooling section, image data of the product surface and interior are collected in real time. The surface image is processed using a convolutional neural network to identify surface cracks, bubbles, and uneven surface defects. For internal defects, thermal imaging data and deep learning algorithms are used for analysis to identify internal bubbles, uneven density, and other problems. Through the multi-level defect detection system, the surface and internal quality of the product can be fully covered, overcoming the limitation that traditional methods can only detect surface defects.

[0137] (2) The present invention introduces a real-time quality control system based on intelligent sensors and edge computing. Intelligent sensors, including temperature sensors, pressure sensors, displacement sensors, etc., are installed at key positions of the extruder to monitor various parameters during the extrusion process in real time. The sensor data is processed using edge computing devices to analyze changes in the extrusion process in real time. By modeling and predicting sensor data using machine learning algorithms, potential quality problems can be identified in advance, and adaptive control can be achieved to automatically adjust the operating parameters of the extruder (such as temperature, pressure, speed) to optimize product quality. This system solves the problem that traditional detection methods lack real-time monitoring and adaptive adjustment capabilities, and significantly improves production efficiency and product quality consistency.

[0138] (3) The extrusion parameter optimization system based on machine learning of the present invention collects historical production data and uses machine learning algorithm data for analysis to establish a relationship model between extrusion parameters and quality indicators. The optimization algorithm is used to optimize the extrusion parameters, find the optimal parameter combination, ensure that the extrusion parameters in the production process are always in the optimal state, realize intelligent optimization and control of the extrusion process, and solve the problem that the parameter setting in the prior art relies on experience and lacks optimization means.

[0139] (4) The present invention comprehensively solves various problems in the prior art through multi-level defect detection, real-time monitoring by intelligent sensors, and parameter optimization based on machine learning, and significantly improves the quality inspection and control level of extruded products. BRIEF DESCRIPTION OF THE DRAWINGS

[0140] The present invention is further described using the accompanying drawings, but the embodiments in the accompanying drawings do not constitute any limitation to the present invention. A person skilled in the art can obtain other drawings based on the following drawings without creative work.

[0141] Figure 1 The present invention is a flow chart of the product quality detection method for an extruder.

[0142] Figure 2 The present invention is a framework diagram of a product quality detection system for an extruder. DETAILED DESCRIPTION

[0143] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and cannot be understood as limiting the present invention.

[0144] In one or more embodiments, Figure 1 As shown, a product quality detection method for an extruder is disclosed, the method comprising the following steps S1-S5:

[0145] S1. Install high-resolution industrial cameras and thermal imagers at key positions of the extruder to collect real-time image data of the product surface and interior respectively. Install intelligent sensors to collect multimodal sensing data during the production process, pre-process and normalize the image data and multimodal sensing data, synchronize the processed data to a unified timeline, and use a distributed database to store the synchronized data.

[0146] Preferably, a high-resolution industrial camera (with a resolution of at least 4096×2160) is installed at the extruder outlet to capture the product surface image. The camera should have a high frame rate (e.g., 30 frames per second) and a high dynamic range (HDR) to capture details.

[0147] Install high-sensitivity thermal imagers (resolution of at least 640×480, sensitivity <0.05°C) at key locations of the extruder (such as the die head and cooling section) to detect internal defects of the product.

[0148] Intelligent sensors are installed at key positions of the extruder, including temperature sensors (accuracy ±0.1°C, response time <1s), pressure sensors (range 0-100MPa, accuracy ±0.1MPa) and displacement sensors (resolution 0.01mm, response time <1ms).

[0149] Ensure the time synchronization of all sensors and imaging devices by synchronizing clocks (such as PTP protocol) to reduce timing errors.

[0150] Preferably, industrial cameras and thermal imagers capture images at a fixed frequency (e.g., 30 frames per second). A trigger mechanism (e.g., photoelectric sensor) is used to ensure that each image corresponds to a different product position. The image data is transmitted to the edge computing device via a high-speed interface (e.g., GigE or USB 3.0). The collected surface image is denoised and smoothed using a Gaussian filter, as follows:

[0151]

[0152] Among them, I′ surf (x, y) is the smoothed image, G(i, j, σ) is the Gaussian kernel function, σ is the standard deviation, and k is half of the kernel size.

[0153] The thermal imaging data is calibrated and denoised, and the formula is as follows:

[0154]

[0155] Among them, T′ int (x, y) is the calibrated temperature data, T int (x, y) is the original thermal imaging data, T bias is the sensor bias, and N is the number of samples.

[0156] Furthermore, temperature, pressure and displacement sensors monitor the parameters of the extrusion process in real time, and the data is collected and transmitted to the edge computing device at a fixed frequency (e.g., 100 times per second). The sensor data is denoised and smoothed using a Kalman filter, with the following formula:

[0157]

[0158] K k =P k-1 H T (HP k-1 H T +R) -1 (4)

[0159] in, is the estimated state at the kth moment, z k is the observation value, H is the observation matrix, P k-1 is the estimation error covariance of the previous step, R is the observation noise covariance, K k is the Kalman gain.

[0160] Furthermore, the sensor data is normalized to make the data range of each parameter consistent. The formula is as follows:

[0161]

[0162] Among them, D′ temp (t), D′ press (t), D′ disp (t) are the normalized temperature, pressure and displacement data respectively.

[0163] Furthermore, the sensor data is normalized to make the data range of each parameter consistent. The formula is as follows:

[0164]

[0165] Among them, D′ temp (t), D′ press (t), D′ disp (t) are the normalized temperature, pressure and displacement data respectively.

[0166] S2. Use convolutional neural network to process surface image data, identify surface defects in real time based on the processing results of surface image data, use deep learning algorithm to analyze thermal imaging data, identify internal defects based on the analysis results of thermal imaging data, and integrate surface defects and internal defects to obtain the final defect map, including:

[0167] S201, obtaining surface image data I' from S1 surf (x, y), divide each image into blocks, calculate the gray level co-occurrence matrix of each sub-block, obtain the texture features of each sub-block, and classify each sub-block using a support vector machine according to the texture features of each sub-block to determine whether there are defects; wherein the texture features include contrast, homogeneity and energy;

[0168] Specifically, step 1 obtains surface image data I′ surf (x, y).

[0169] Each image is processed into blocks of size m×n. The gray level co-occurrence matrix (GLCM) of each sub-block is calculated to capture the image texture features. The calculation formula of the gray level co-occurrence matrix C(i, j) is:

[0170]

[0171] Where δ is an indicator function, which takes the value 1 when the condition is true and 0 otherwise; d x and d y is the offset between adjacent pixels.

[0172] Calculate the texture features of each sub-block, including contrast, homogeneity and energy, as follows:

[0173]

[0174] Use support vector machine (SVM) to classify each sub-block to determine whether there is a defect. Training set T = {(F k ,y k )} by the eigenvector F k = [contrast, homogeneity, energy] and the corresponding label y k∈{0, 1} (0 means no defect, 1 means defect).

[0175] Classify each sub-block to obtain the defect map D surf (x, y), a value of 1 indicates a defect, and a value of 0 indicates no defect.

[0176] S202, obtaining internal thermal imaging data T′ from S1 int (x, y), use the Canny edge detection algorithm to perform edge detection on each thermal image, use the watershed algorithm to segment the image based on the edge results, divide the thermal image into different areas, and calculate the average temperature T of each area avg and temperature standard deviation T std ,Using the logistic regression model, each area is classified to determine whether there is an internal defect;

[0177] Specifically, the internal thermal imaging data T′ is obtained from step 1 int (x, y). Perform edge detection on each thermal image using the Canny edge detection algorithm. The steps are as follows:

[0178] Compute the gradient magnitude and direction of an image:

[0179]

[0180] Non-maximum suppression, thinning edges:

[0181]

[0182] Apply double threshold detection to determine true edges:

[0183]

[0184] Use the watershed algorithm to segment the image and divide the thermal image into different areas:

[0185] Label(x,y)=watershed(E′(x,y)) (17)

[0186] Calculate the average temperature T of each area avg and temperature standard deviation T std :

[0187]

[0188] Use the logistic regression model to classify each area to determine whether there are internal defects. Training set R = {(G k , z k )} By the eigenvector G k =[T avg , T std] and the corresponding label z k ∈{0, 1} (0 means no defect, 1 means defect).

[0189] Classify each area to get the defect map D int (x, y), a value of 1 indicates a defect, and a value of 0 indicates no defect.

[0190] S203, obtain the confidence map C of the surface defect through the output probability of the classifier surf (x, y) and confidence map of internal defects C int (x, y), the surface defect confidence map C surf (x, y) and internal defect confidence map C int (x, y) is fused using the Bayesian weighted fusion method to obtain the comprehensive confidence map C comb (x, y), for the comprehensive defect map D final (x, y) is post-processed to extract the boundary and shape of the defect area, the processed defect information is displayed through a visual interface, and is fed back to the control system in real time through the edge computing device;

[0191] Specifically, the confidence map C of the surface defects is calculated surf (x, y) and confidence map of internal defects C int (x, y). The confidence can be obtained through the output probability of the classifier:

[0192] C surf (x, y) = SVM_prob(F k ) (20)

[0193] C int (x, y) = LogReg_prob(G k ) (twenty one)

[0194] The surface defect confidence map C surf (x, y) and internal defect confidence map C int (x, y) are fused to obtain the comprehensive confidence map C comb (x, y), using the Bayesian weighted fusion method:

[0195]

[0196] Among them, P surf and P int are the prior probabilities of surface defects and internal defects, respectively.

[0197] According to the comprehensive confidence map C comb (x, y) generates the final defect map D final (x, y), using threshold Tcomb Binarization:

[0198]

[0199] For the comprehensive defect map D final (x, y) is post-processed to extract the boundary and shape of the defect area. Morphological operations (such as corrosion and expansion) are used to clean up the noise. The formula is as follows:

[0200] D morp (x, y) = dilate(erode(D final (x,y),SE),SE) (24)

[0201] Among them, erode and dilate represent morphological erosion and dilation operations respectively, and SE is the structural element.

[0202] The processed defect information is displayed through a visual interface and fed back to the control system in real time through edge computing devices for the next step of adaptive control.

[0203] S3. Extract features from the stored data and defect graphs, build a multi-layer perception model to generate a monitoring state vector, and use a long short-term memory network model to perform real-time monitoring and anomaly detection on the monitoring state vector. When an anomaly is detected, the anomaly information and specific feature vectors are fed back to the extruder control system, and the control system adjusts the operating parameters of the extruder based on the feedback information.

[0204] Specifically, edge computing devices, including high-performance GPUs and FPGAs, are installed next to the extruder to process large amounts of data in real time. The edge computing devices are connected to sensors and imaging devices through high-speed network interfaces (such as 10GbE) to ensure fast data transmission and processing. A real-time operating system (RTOS) is configured to ensure low-latency and high-reliability data processing.

[0205] Further, the edge computing device receives the synchronization data transmitted from step 1 and step 2, including the surface image data I′ surf (x, y), thermal imaging data T′ int (x, y), and sensor data D′ temp (t), D′ press (t), D′ disp (t).

[0206] Match and parse all data by timestamp to ensure data consistency and synchronization. The formula is as follows:

[0207]

[0208] Among them, D sync(t) is the synchronization data at time t, TD is the original data timestamp set, t1 and t2 are the two most recent timestamps, and D(t1) and D(t2) are the corresponding data values.

[0209] Furthermore, the received data is processed and analyzed in real time to extract useful information. Convolutional Neural Network (CNN) is used to extract features from surface images:

[0210] F surf =CNN(I′ surf (x,y)) (26)

[0211] Among them, F surf is the feature vector of the surface image, I′ surf (x, y) is the surface image after preprocessing.

[0212] Furthermore, the temperature features of the thermal imaging data are extracted:

[0213] F int =[T avg , T std ] (27)

[0214] Among them, T avg is the average temperature, T std is the temperature standard deviation, T′ int (x, y) is the preprocessed thermal imaging data.

[0215] Extract features from sensor data and calculate the rate of change of real-time parameters:

[0216]

[0217] Among them, D′ temp (t), D′ press (t) and D′ disp (t) are the normalized temperature, pressure and displacement data, respectively, and Δt is the time interval.

[0218] Furthermore, a multi-layer perception model is established to fuse the extracted feature data and generate the overall monitoring state vector S(t):

[0219] S(t)=[F surf , F in t, ΔD temp (t), ΔD press (t), ΔD disp (t)] (31)

[0220] Among them, F surf is the feature vector of the surface image, F int is the characteristic vector of thermal imaging data, ΔDtemp (t), ΔD press (t) and ΔD disp (t) are the rates of change of temperature, pressure and displacement data respectively.

[0221] The long short-term memory network (LSTM) model is used for real-time monitoring and anomaly detection. The model input is the current and historical state vector S(t), and the output is the anomaly score A(t):

[0222] A(t)=LSTM(S(tT), S(t-T+1),…,S(t)) (32)

[0223] Among them, τ is the time window length.

[0224] Determine whether the system has an anomaly based on the anomaly score A(t). When A(t) exceeds the threshold T A When an alarm is triggered, the abnormal information is recorded:

[0225] if A(t)>T A then trigger alarm (33)

[0226] Among them, T A The threshold for scoring anomalies.

[0227] It can be understood that when an abnormality is detected, the abnormal information and the specific feature vector are fed back to the extruder control system, and the control system adjusts the operating parameters (such as temperature, pressure and speed) of the extruder according to the feedback information to correct the abnormality. The formula is as follows:

[0228] ΔP extr =K p ΔD press (t) (34)

[0229] ΔT prod =K t ΔD temp (t) (35)

[0230] Δv extr =K v ΔD disp (t) (36)

[0231] Among them, K p , K t and K v are the control gain coefficients for pressure, temperature and speed adjustment, ΔP extr , ΔT prod and Δv extr They are the adjustment amounts for extruder pressure, temperature and speed respectively.

[0232] Furthermore, real-time monitoring data and anomaly detection results are transmitted to the cloud database via high-speed networks for storage and analysis. Visualization tools are used to display monitoring data, anomaly information, and adjustment parameters to operators in the form of charts to ensure real-time visualization and tracking.

[0233] S4. Based on the real-time monitoring data and comparing and analyzing it with the historical data, an adaptive control model is constructed to adjust the operating parameters of the extruder in real time according to the analysis results.

[0234] Specifically, obtain the real-time monitoring data from step 3, including the state vector S(t) and the anomaly score A(t). Compare the current state with the historical state and calculate the state error E(t):

[0235]

[0236] in, is the predicted ideal state vector.

[0237] Furthermore, the adaptive control model is established, and the steps are as follows:

[0238] According to the state error E(t), the control target C(t) is defined as:

[0239] C(t)=K c ·E(t) (38)

[0240] Among them, K c is the control gain matrix, which is used to adjust the influence of the error on the control target;

[0241] Decompose the control target C(t) into the temperature control target C T (t), pressure control target C P (t) and speed control target C V (t):

[0242] C T (t) = [c T1 (t), c T2 (t),…,c Tn (t)] (39)

[0243] C P (t) = [c P1 (t), c P2 (t),…,c Pn (t)] (40)

[0244] C V (t) = [c V1 (t), c V2 (t), ..., c Vn (t)] (41)

[0245] Among them, c Ti (t), c Pi (t) and c Vi (t) are the target values ​​of each control variable at time t.

[0246] Furthermore, a PID controller is designed for each control target, and the operating parameters of the extruder are adjusted to achieve the control target. The temperature PID controller formula is as follows:

[0247]

[0248] e T (t) = C T (t)-T prod (t) (43)

[0249] Among them, u T (t) is the temperature control input, e T (t) is the temperature error, K pT , K iT and K dT They are the proportional, integral and derivative gains of temperature control respectively.

[0250] The pressure PID controller formula is as follows:

[0251]

[0252] e P (t) = C P (t)-P extr (t) (45)

[0253] Among them, u P (t) is the pressure control input, e P (t) is the pressure error, K pP , K iP and K dP They are the proportional, integral and differential gains of pressure control respectively.

[0254] The speed PID controller formula is as follows:

[0255]

[0256] e V (t) = C V (t)-v extr (t) (47)

[0257] Among them, u V (t) is the speed control input, e V (t) is the speed error, K pV , KiV and K dV They are the proportional, integral and differential gains of speed control respectively.

[0258] Further, a control signal is generated, and the control input u T (t),u P (t) and u V (t) Convert into actual control instructions to adjust the operating parameters of the extruder.

[0259] P extr (t+Δt)=P extr (t)+u P (t) (48)

[0260] T prod (t+Δt)=T prod (t)+u T (t) (49)

[0261] v extr (t+Δt)=v extr (t)+u V (t) (50)

[0262] Among them, P extr (t+Δt), T prod (t+Δt) and v extr (t+Δt) are the adjusted pressure, temperature and speed respectively. The control instructions are sent to the extruder control system in real time to ensure the timeliness and accuracy of the adjustment.

[0263] S5. Extract key features from stored data and real-time monitoring data, select multiple machine learning algorithms for integration and use weighted average method to fuse the prediction results of multiple models, use optimization algorithm to predict and optimize extrusion parameters, and combine the operating parameters adjusted in real time by adaptive control model to obtain the optimal extruder operating parameter combination and the optimal product quality test results, including:

[0264] S501, extract key features from the stored data and real-time monitoring data, and construct a feature vector X(t); wherein the key features include temperature T prod 、Pressure P extr and speed v extr ; The feature vector X(t) is expressed as follows:

[0265] X(t)=[T prod (t), P extr (t), v extr (t), ΔT prod (t), ΔP extr (t), Δv extr(t), …] (51)

[0266] Among them, T prod (t) represents the extruder production temperature at time t, P extr (t) represents the extrusion pressure at time t, v extr (t) represents the extrusion speed at time t, ΔT prod (t) represents the rate of change of temperature, expressed as the temperature change from time t to the previous time divided by the time interval, ΔP extr (t) represents the rate of change of pressure, expressed as the pressure change between time t and the previous time divided by the time interval, Δv extr (t) represents the rate of change of speed, which is expressed as the change in speed from time t to the previous time divided by the time interval;

[0267] The product quality test results are used as labels Y(t):

[0268] Y(t)=[D surf (t), D int (t)] (52)

[0269] Among them, D surf (t) represents the surface defect detection result at time t, D int (t) represents the internal defect detection result at time t;

[0270] S502, selecting multiple machine learning models for integration to obtain an integrated model, wherein the multiple machine learning models include a support vector machine, a random forest, and a gradient boosting decision tree;

[0271] Specifically, each model is trained with the following formula:

[0272]

[0273] in, The prediction results of the SVM model at time t. The prediction results of the random forest model at time t. Prediction result of GBDT model at time t. X(t): feature vector at time t.

[0274] S503, using the weighted average method to fuse the prediction results of multiple models of the integrated model, expressed as follows:

[0275]

[0276] in, At time t, the prediction result of the integrated model. SVM : Weight of the SVM model. wRF : Weights of the random forest model. w GBDT : The weight of the GBDT model. w SVM +w RF +w GBDT =1: The sum of the weights is 1.

[0277] S504, using incremental SVM to perform online learning and real-time optimization on the integrated model, as shown below:

[0278]

[0279] Among them, w t+1 represents the updated weight vector at time t+1; w t represents the weight vector at time t; η represents the learning rate; y t represents the actual label at time t; σ represents the activation function; x t Represents the feature vector at time t;

[0280] S505. According to the prediction result of the optimized integrated model, the operating parameters of the extruder are adjusted in real time to achieve the best product quality; wherein the real-time adjustment of the operating parameters of the extruder is expressed as follows:

[0281]

[0282] Among them, α T , α P and α V are the adjustment coefficients for temperature, pressure and speed, respectively. represents the prediction result of the integrated model at time t; Y(t) represents the actual detection result at time t; T prod (t+Δt) represents the adjusted temperature at time t+Δt; P extr (t+Δt) represents the adjusted pressure at time t+Δt; v extr (t+Δt) represents the adjusted speed at time t+Δt;

[0283] S506. Use mean square error and determination coefficient as evaluation indicators to regularly evaluate the performance of the integrated model, as shown below:

[0284]

[0285] Where Yi: the actual value of the i-th sample. The predicted value of the i-th sample. The mean of all actual values. N: number of samples.

[0286] Furthermore, based on the evaluation results, the model is updated regularly to ensure the stability and efficiency of the model in long-term use.

[0287] In one or more embodiments, Figure 2 As shown, a product quality detection system for an extruder is disclosed, the system comprising:

[0288] The data acquisition module 101 is used to install a high-resolution industrial camera and a thermal imager at a key position of the extruder to respectively collect image data of the product surface and the inside in real time, and to install an intelligent sensor to collect multimodal sensor data in the production process, pre-process and normalize the image data and the multimodal sensor data, synchronize the processed data to a unified time axis, and use a distributed database to store the synchronized data;

[0289] The defect analysis module 102 is used to process the surface image data using a convolutional neural network, identify surface defects in real time according to the processing results of the surface image data, analyze the thermal imaging data using a deep learning algorithm, identify internal defects according to the analysis results of the thermal imaging data, and integrate the surface defects and the internal defects to obtain a final defect map;

[0290] The quality detection module 103 is used to extract features from the stored data and defect graph, build a multi-layer perception model to generate a monitoring state vector, and use a long short-term memory network model to perform real-time monitoring and abnormality detection on the monitoring state vector. When an abnormality is detected, the abnormal information and specific feature vectors are fed back to the extruder control system, and the control system adjusts the operating parameters of the extruder according to the feedback information;

[0291] The parameter adjustment module 104 is used to compare and analyze the real-time monitoring data with the historical data, and to build an adaptive control model to adjust the operating parameters of the extruder in real time according to the analysis results;

[0292] The quality inspection operation optimization module 105 is used to extract key features from the stored data and real-time monitoring data, select multiple machine learning algorithms for integration and use the weighted average method to fuse the prediction results of multiple models, use the optimization algorithm to predict and optimize the extrusion parameters, and combine the operating parameters adjusted in real time by the adaptive control model to obtain the optimal extruder operating parameter combination and the optimal product quality inspection result;

[0293] The steps performed by the quality inspection operation optimization module include:

[0294] S501, extract key features from the stored data and real-time monitoring data, and construct a feature vector X(t); wherein the key features include temperature T prod 、Pressure P extr and speed vextr ; The feature vector X(t) is expressed as follows:

[0295] X(t)=[T prod (t), P extr (t), v extr (t), ΔT prod (t), ΔP extr (t), Δv extr (t), ...]

[0296] Among them, T prod (t) represents the extruder production temperature at time t, P extr (t) represents the extrusion pressure at time t, v extr (t) represents the extrusion speed at time t, ΔT prod (t) represents the rate of change of temperature, expressed as the temperature change from time t to the previous time divided by the time interval, ΔP extr (t) represents the rate of change of pressure, expressed as the pressure change between time t and the previous time divided by the time interval, Δv extr (t) represents the rate of change of speed, which is expressed as the change in speed from time t to the previous time divided by the time interval;

[0297] The product quality test results are used as labels Y(t):

[0298] Y(t)=[D surf (t), D int (t)]

[0299] Among them, D surf (t) represents the surface defect detection result at time t, D int (t) represents the internal defect detection result at time t;

[0300] S502, selecting multiple machine learning models for integration to obtain an integrated model, wherein the multiple machine learning models include a support vector machine, a random forest, and a gradient boosting decision tree;

[0301] S503, using a weighted average method to fuse the prediction results of multiple models of the integrated model;

[0302] S504, using incremental SVM to perform online learning and real-time optimization on the integrated model, as shown below:

[0303]

[0304] Among them, w t+1 represents the updated weight vector at time t+1; w t represents the weight vector at time t; η represents the learning rate; yt represents the actual label at time t; σ represents the activation function; x t Represents the feature vector at time t;

[0305] S505. According to the prediction result of the optimized integrated model, the operating parameters of the extruder are adjusted in real time to achieve the best product quality; wherein the real-time adjustment of the operating parameters of the extruder is expressed as follows:

[0306]

[0307] Among them, α T , α P and α V are the adjustment coefficients for temperature, pressure and speed, respectively. represents the prediction result of the integrated model at time t; Y(t) represents the actual detection result at time t; T prod (t+Δt) represents the adjusted temperature at time t+Δt; P extr (t+Δt) represents the adjusted pressure at time t+Δt; v extr (t+Δt) represents the adjusted speed at time t+Δt;

[0308] S506. Regularly evaluate the performance of the integrated model using mean square error and coefficient of determination as evaluation indicators.

[0309] It is worth noting that the specific working process of the product quality detection system for an extruder provided in the embodiment of the present invention is the same as the process of the product quality detection method for an extruder described in the above embodiment, and will not be described in detail here.

[0310] Compared with the prior art, the product quality inspection system for an extruder provided by an embodiment of the present invention installs a high-resolution industrial camera and a thermal imager at a key position of the extruder to respectively collect image data of the product surface and the interior in real time, and installs an intelligent sensor to collect multimodal sensing data in the production process, pre-processes and normalizes the image data and the multimodal sensing data, synchronizes the processed data to a unified time axis, and uses a distributed database to store the synchronized data; uses a convolutional neural network to process the surface image data, and identifies surface defects in real time according to the processing results of the surface image data; uses a deep learning algorithm to analyze the thermal imaging data, identifies internal defects according to the analysis results of the thermal imaging data, and integrates the surface defects and the internal defects to obtain a final defect map; performs special analysis from the stored data and the defect map Feature extraction, building a multi-layer perception model to generate a monitoring state vector, and using a long short-term memory network model to perform real-time monitoring and anomaly detection on the monitoring state vector. When an anomaly is detected, the anomaly information and specific feature vectors are fed back to the extruder control system, and the control system adjusts the operating parameters of the extruder according to the feedback information; based on the real-time monitoring data and historical data, the data are compared and analyzed, and an adaptive control model is built to adjust the operating parameters of the extruder in real time according to the analysis results; key features are extracted from the stored data and real-time monitoring data, a variety of machine learning algorithms are selected for integration, and the prediction results of multiple models are fused using the weighted average method, and the optimization algorithm is used to predict and optimize the extrusion parameters. Combined with the operating parameters adjusted in real time by the adaptive control model, the optimal extruder operating parameter combination and the optimal product quality detection results are obtained.

[0311] The embodiment of the present invention also provides a product quality detection device for an extruder, comprising a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, wherein when the processor executes the computer program, the steps in the embodiment of the product quality detection method for an extruder are implemented, for example Figure 1 or, the processor implements the functions of each module in the above-mentioned system embodiments when executing the computer program.

[0312] Exemplarily, the computer program may be divided into one or more modules, which are stored in the memory and executed by the processor to complete the present invention. The one or more modules may be a series of computer program instruction segments capable of completing specific functions, which are used to describe the execution process of the computer program in the product quality detection device for an extruder.

[0313] The product quality detection device for the extruder can be a computing device such as a desktop computer, a notebook, a PDA, and a cloud server. The product quality detection device for the extruder can include, but is not limited to, a processor and a memory. Those skilled in the art can understand that the product quality detection device for the extruder can also include an input and output device, a network access device, a bus, etc.

[0314] The processor may be a central processing unit (CPU), other general processors, digital signal processors (DSP), application specific integrated circuits (ASAC), field programmable gate arrays (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general processor may be a microprocessor or any conventional processor, etc. The processor is the control center of the product quality detection device for the extruder, and various interfaces and lines are used to connect the various parts of the entire product quality detection device for the extruder.

[0315] The memory can be used to store the computer program and / or module, and the processor realizes various functions of the product quality inspection device for the extruder by running or executing the computer program and / or module stored in the memory, and calling the data stored in the memory. The memory can mainly include a program storage area and a data storage area, wherein the program storage area can store an operating system, an application required for at least one function, etc.; the data storage area can store data created according to the operation of the air-conditioning controller, etc. In addition, the memory can include a high-speed random access memory, and can also include a non-volatile memory, such as a hard disk, a memory, a plug-in hard disk, a smart memory card (SmartMedaaCard, SMC), a secure digital (SecureDagatal, SD) card, a flash card (FlashCard), at least one disk storage device, a flash memory device, or other volatile solid-state storage devices.

[0316] Wherein, if the module integrated with the product quality detection equipment for the extruder is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on such an understanding, the present invention implements all or part of the process in the above-mentioned embodiment method, and can also be completed by instructing the relevant hardware through a computer program. The computer program can be stored in a computer-readable storage medium, and the computer program can implement the steps of the above-mentioned various method embodiments when executed by the processor. Wherein, the computer program includes a computer program code, and the computer program code can be in the form of source code, object code, executable file or some intermediate form. The computer-readable medium may include: any entity or device capable of carrying the computer program code, a recording medium, a U disk, a mobile hard disk, a magnetic disk, an optical disk, a computer memory, a read-only memory (ROM, Read-Only Memory), a random access memory (RAM, Random Access Memory), an electric carrier signal, a telecommunication signal, and a software distribution medium.

[0317] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiments can be implemented by instructing the relevant hardware through a computer program, and the program can be stored in a computer-readable storage medium, and when the program is executed, it can include the processes of the embodiments of the above-mentioned methods. The storage medium can be a disk, an optical disk, a read-only memory (ROM) or a random access memory (RAM), etc.

[0318] The above is a preferred embodiment of the present invention. It should be pointed out that a person skilled in the art can make several improvements and modifications without departing from the principle of the present invention. These improvements and modifications are also considered to be within the scope of protection of the present invention.

Claims

1. A product quality detection method for an extruder, characterized in that: The method comprises: S1. Install high-resolution industrial cameras and thermal imagers at key locations of the extruder to collect real-time image data of the product surface and interior respectively. Install intelligent sensors to collect multimodal sensor data during the production process, pre-process and normalize the image data and multimodal sensor data, synchronize the processed data to a unified timeline, and use a distributed database to store the synchronized data; S2. Use a convolutional neural network to process the surface image data, identify surface defects in real time based on the processing results of the surface image data, use a deep learning algorithm to analyze the thermal imaging data, identify internal defects based on the analysis results of the thermal imaging data, and integrate the surface defects and internal defects to obtain the final defect map; S3, extracting features from the stored data and defect graph, building a multi-layer perception model to generate a monitoring state vector, and using a long short-term memory network model to perform real-time monitoring and anomaly detection on the monitoring state vector. When an anomaly is detected, the anomaly information and specific feature vectors are fed back to the extruder control system, and the control system adjusts the operating parameters of the extruder according to the feedback information; S4, based on the real-time monitoring data and comparing and analyzing with the historical data, construct an adaptive control model to adjust the operating parameters of the extruder in real time according to the analysis results; S5. Extract key features from the stored data and real-time monitoring data, select multiple machine learning algorithms for integration and use the weighted average method to fuse the prediction results of multiple models, use the optimization algorithm to predict and optimize the extrusion parameters, and combine the operating parameters adjusted in real time by the adaptive control model to obtain the optimal extruder operating parameter combination and the optimal product quality test results; Wherein, the S5 includes: S501, extract key features from the stored data and real-time monitoring data, and construct a feature vector X(t); wherein the key features include temperature T prod 、Pressure P extr and speed v extr ; The feature vector X(t) is expressed as follows: X(t)=[T prod (t),P extr (t),v extr (t),△T prod (t),△P extr (t),Δv extr (t),…] Among them, T prod (t) represents the extruder production temperature at time t, P extr (t) represents the extrusion pressure at time t, V extr (t) represents the extrusion speed at time t, △T prod (t) represents the rate of change of temperature, which is the temperature change between time t and the previous time divided by the time interval, △P extr (t) represents the rate of change of pressure, expressed as the pressure change between time t and the previous time divided by the time interval, Δv extr (t) represents the rate of change of speed, which is expressed as the change in speed from time t to the previous time divided by the time interval; The product quality test results are used as labels Y(t): Y(t)=[D surf (t),D int (t)] Among them, D surf (t) represents the surface defect detection result at time t, D int (t) represents the internal defect detection result at time t; S502, selecting multiple machine learning models for integration to obtain an integrated model, wherein the multiple machine learning models include a support vector machine, a random forest, and a gradient boosting decision tree; S503, using a weighted average method to fuse the prediction results of multiple models of the integrated model; S504, using incremental SVM to perform online learning and real-time optimization on the integrated model, as shown below: Among them, w t+1 represents the updated weight vector at time t+1; w t represents the weight vector at time t; η represents the learning rate; y t represents the actual label at time t; σ represents the activation function; x t Represents the feature vector at time t; S505. According to the prediction result of the optimized integrated model, the operating parameters of the extruder are adjusted in real time to achieve the best product quality; wherein the real-time adjustment of the operating parameters of the extruder is expressed as follows: Among them, α T , α P and α V are the adjustment coefficients for temperature, pressure and speed, respectively. represents the prediction result of the integrated model at time t; Y(t) represents the actual detection result at time t; T prod (t+△t) represents the adjusted temperature at time t+△t; P extr (t+△t) represents the adjusted pressure at time t+△t; v extr (t+△t) represents the adjusted speed at time t+△t; S506. Regularly evaluate the performance of the integrated model using mean square error and coefficient of determination as evaluation indicators.

2. The product quality detection method for an extruder according to claim 1, characterized in that: The intelligent sensor includes a temperature sensor, a pressure sensor and a displacement sensor; The preprocessing includes: denoising the collected surface image and smoothing it with a Gaussian filter; calibrating and denoising the thermal imaging data; using temperature sensors, pressure sensors and displacement sensors to monitor the parameters of the extrusion process in real time, collecting data at a fixed frequency and transmitting it to the edge computing device, denoising the sensor data, and smoothing it with a Kalman filter.

3. The product quality detection method for an extruder according to claim 1, characterized in that: The step S2 specifically includes: S201, obtaining surface image data I′surf(x, y) from S1, performing block processing on each image, calculating the gray level co-occurrence matrix of each sub-block, obtaining the texture feature of each sub-block, and classifying each sub-block using a support vector machine according to the texture feature of each sub-block to determine whether there is a defect; wherein the texture features include contrast, homogeneity and energy; S202, obtaining internal thermal imaging data T′ from S1 in t(x, y), use the Canny edge detection algorithm to perform edge detection on each thermal image, use the watershed algorithm to segment the image based on the edge results, divide the thermal image into different areas, and calculate the average temperature T of each area avg and temperature standard deviation T std ,Using the logistic regression model, each area is classified to determine whether there is an internal defect; S203, obtain the confidence map C of the surface defect through the output probability of the classifier surf (x, y) and confidence map of internal defects C int (x, y), the surface defect confidence map C surf (x, y) and internal defect confidence map C int (x, y) is fused using the Bayesian weighted fusion method to obtain the comprehensive confidence map C comb (x, y), for the comprehensive defect map D final (x, y) is post-processed to extract the boundary and shape of the defect area, the processed defect information is displayed through a visual interface, and is fed back to the control system in real time through the edge computing device.

4. The product quality detection method for an extruder according to claim 3, characterized in that: The image is divided into m×n sub-blocks, and the gray level co-occurrence matrix C(i, j) is calculated as follows: Where δ represents the indicator function, which takes the value 1 when the condition is true and takes the value 0 otherwise; d x and d y Indicates the offset of adjacent pixels; I′ surf (x, y) represents the smoothed surface image, I′ surf (x+d x ,y+d y ) represents the smoothed surface image based on the offset of neighboring pixels; Calculate the texture features of each sub-block, including contrast, homogeneity and energy, as follows: Use support vector machine to classify each sub-block to determine whether there are defects: Training set T = {(F k , yk)} is given by the eigenvector F k =[contrast, homogeneity, energy] and the corresponding label yk∈{0,1}, where 0 means no defect and 1 means defect; Classify each sub-block to obtain the defect map D surf (x, y), a value of 1 indicates a defect, and a value of 0 indicates no defect.

5. The product quality detection method for an extruder according to claim 4, characterized in that: The S202 specifically includes: Calculate the gradient strength and direction of the internal image, expressed as follows: Among them, G x , G y represents the corresponding coordinates of the internal image gradient intensity, and G represents the direction of the internal image; Non-maximum suppression, thinning edges, is expressed as follows: Apply double threshold detection to determine the true edge, which is expressed as follows: Among them, E′(x, y) represents the real edge; Use the watershed algorithm to segment the image and divide the thermal image into different areas: Label(x,y)=watershed(E′(x,y)) Among them, Label(x, y) represents the segmented area, and watershed represents the watershed function; Calculate the average temperature T of each area avg and temperature standard deviation T std : Where N represents the number of samples; Use the logistic regression model to classify each area to determine whether there are internal defects: Training set R = {(G k , z k )} By the eigenvector G k =[T avg , T std ] and the corresponding label z k ∈{0, 1}, where 0 means no defect and 1 means defect; Classify each area to get the defect map D int (x, y), a value of 1 indicates a defect, and a value of 0 indicates no defect.

6. The product quality detection method for an extruder according to claim 5, characterized in that: The S203 specifically includes: The confidence map C of the surface defect is calculated by the output probability of the classifier surf (x, y) and confidence map of internal defects C int (x, y) confidence, expressed as follows: C surf (x,y)=SVM_prob(F k ) C int (x,y)=LogReg_prob(G k ) The surface defect confidence map C is transformed into surf (x, y) and internal defect confidence map C int (x, y) are fused to obtain the comprehensive confidence map C comb (x, y), expressed as follows: Among them, P surf and P int denote the prior probabilities of surface defects and internal defects respectively; According to the comprehensive confidence map C comb (x, y) generates the final defect map D final (x, y), using threshold T comb Binarization is performed, which is expressed as follows: Use morphological operations to analyze the integrated defect map D final (x, y) is post-processed to extract the boundary and shape of the defect area; wherein the post-processing is to clean up the noise, and the formula is as follows: D morp (x, y) dilate(erode(D final (x, y), SE), SE) Among them, erode and dilate represent morphological erosion and dilation operations respectively, and SE represents the structural element; The processed defect information is displayed through a visual interface and fed back to the control system in real time through edge computing devices.

7. The product quality detection method for an extruder according to claim 1, characterized in that: In S3, the real-time monitoring and anomaly detection specifically include: A multi-layer perception model is established to fuse the extracted feature data and generate the overall monitoring state vector S(t), which is expressed as follows: S(t)=[F surf ,F int ,△D temp (t),△D press (t),△D disp (t)] Among them, F surf The feature vector representing the surface image, F int The characteristic vector representing the thermal imaging data, △D temp (t), △D press (t) and △D disp (t) represents the rate of change of temperature, pressure and displacement data respectively. The long short-term memory network model is used for real-time monitoring and anomaly detection. The model input is the current and historical state vector S(t), and the output is the anomaly score A(t), which is expressed as follows: A(t)=LSTM(S(t-τ), S(t-T+1),…,S(t)) Where T represents the time window length. According to the anomaly score A(t), it is judged whether the system has an anomaly. When A(t) exceeds the threshold T A When an alarm is triggered, the abnormal information is recorded, as shown below: if A(t)>T A then trigger alarm Among them, T A Threshold for scoring anomalies; Among them, when an abnormality is detected, the abnormal information and specific feature vectors are fed back to the extruder control system, and the control system adjusts the operating parameters of the extruder according to the feedback information to correct the abnormality, as shown below: △P extr =K p ·△D press (t) △T prod =K t ·△D temp (t) △V extr =K v ·△D disp (t) Among them, K p , K t and K v Respectively represent the control gain coefficients of pressure, temperature and speed adjustment, △P extr , △T prod and △V extr Respectively represent the adjustment amount of extruder pressure, temperature and speed.

8. The product quality detection method for an extruder according to claim 1, characterized in that: In S4, the adaptive control model is constructed as follows: According to the state error E(t), the control target C(t) is defined as follows: C(t)=K c ·E(t) Among them, K c represents the control gain matrix, which is used to adjust the effect of the error on the control target; Decompose the control target C(t) into the temperature control target C T (t), pressure control target C P (t) and speed control target C V (t): C T (t)=[c T1 (t),c T2 (t),…,c Ti (t),c Tn (t)] C P (t)=[c P1 (t),c P2 (t),…,c Pi (t),c Pn (t)] C V (t)=[c V1 (t),c V2 (t),...,c Vi (t),c Vn (t)] Among them, c Ti (t), c Pi (t) and c Vi (t) are the target values ​​of each control variable at time t.

9. The product quality detection method for an extruder according to claim 8, characterized in that: A PID controller is designed for each control target, and the operating parameters of the extruder are adjusted to achieve the control target. The temperature PID controller formula is as follows: e T (t)=C T (t)-T prod (t) Among them, u T (t) represents the temperature control input, e T (t) represents the temperature error, K pT , K iT and K dT They represent the proportional, integral and derivative gains of temperature control respectively; The pressure PID controller formula is as follows: e P (t)=C P (t)-P extr (t) Among them, u P (t) represents the pressure control input, e P (t) represents the pressure error, K pP , K iP and K dP They represent the proportional, integral and differential gains of pressure control respectively; The speed PID controller formula is as follows: e V (t)=C V (t)-V extr (t) Among them, u V (t) represents the speed control input, e V (t) represents the speed error, K pV , K iV and K dV Respectively represent the proportional, integral and differential gains of speed control; Control input u T (t),u P (t) and u V (t) is converted into actual control instructions to adjust the operating parameters of the extruder, as shown below: P extr (t+△t)=P extr (t)+u P (t) T prod (t+△t)=T prod (t)+u T (t) v extr (t+△t)=v extr (t)+u V (t) Among them, P extr (t+△t), T prod (t+△t) and V extr (t+△t) represent the adjusted pressure, temperature and speed respectively.

10. A product quality detection system for an extruder, characterized in that: The system comprises: The data acquisition module is used to install high-resolution industrial cameras and thermal imagers at key positions of the extruder to collect real-time image data of the product surface and interior respectively. At the same time, intelligent sensors are installed to collect multimodal sensor data during the production process, and the image data and multimodal sensor data are pre-processed and normalized. The processed data are synchronized to a unified timeline and the synchronized data is stored using a distributed database; The defect analysis module is used to process the surface image data using a convolutional neural network, identify surface defects in real time based on the processing results of the surface image data, analyze the thermal imaging data using a deep learning algorithm, identify internal defects based on the analysis results of the thermal imaging data, and integrate surface defects and internal defects to obtain the final defect map; The quality inspection module is used to extract features from the stored data and defect graphs, build a multi-layer perception model to generate a monitoring state vector, and use a long short-term memory network model to perform real-time monitoring and anomaly detection on the monitoring state vector. When an anomaly is detected, the anomaly information and specific feature vectors are fed back to the extruder control system, and the control system adjusts the operating parameters of the extruder according to the feedback information; The parameter adjustment module is used to compare and analyze the real-time monitoring data with the historical data, and to build an adaptive control model to adjust the operating parameters of the extruder in real time according to the analysis results; The quality inspection operation optimization module is used to extract key features from the stored data and real-time monitoring data, select multiple machine learning algorithms for integration and use the weighted average method to fuse the prediction results of multiple models, use the optimization algorithm to predict and optimize the extrusion parameters, and combine the operating parameters adjusted in real time by the adaptive control model to obtain the optimal extruder operating parameter combination and the optimal product quality inspection results; The steps performed by the quality inspection operation optimization module include: S501, extract key features from the stored data and real-time monitoring data, and construct a feature vector X(t); wherein the key features include temperature T prod 、Pressure P extr and speed v extr ; The feature vector X(t) is expressed as follows: X(t)=[T prod (t),P extr (t),v extr (t),△T prod (t),△P extr (t),△v extr (t),…] Among them, T prod (t) represents the extruder production temperature at time t, P extr (t) represents the extrusion pressure at time t, v extr (t) represents the extrusion speed at time t, △T prod (t) represents the rate of change of temperature, which is the temperature change between time t and the previous time divided by the time interval, △P extr (t) represents the rate of change of pressure, which is the pressure change between time t and the previous time divided by the time interval, △v extr (t) represents the rate of change of speed, which is expressed as the change in speed from time t to the previous time divided by the time interval; The product quality test results are used as labels Y(t): Y(t)=[D surf (t),D int (t)] Among them, D surf (t) represents the surface defect detection result at time t, D int (t) represents the internal defect detection result at time t; S502, selecting multiple machine learning models for integration to obtain an integrated model, wherein the multiple machine learning models include a support vector machine, a random forest, and a gradient boosting decision tree; S503, using a weighted average method to fuse the prediction results of multiple models of the integrated model; S504, using incremental SVM to perform online learning and real-time optimization on the integrated model, as shown below: Among them, w t+1 represents the updated weight vector at time t+1; w t represents the weight vector at time t; η represents the learning rate; y t represents the actual label at time t; σ represents the activation function; x t Represents the feature vector at time t; S505. According to the prediction result of the optimized integrated model, the operating parameters of the extruder are adjusted in real time to achieve the best product quality; wherein the real-time adjustment of the operating parameters of the extruder is expressed as follows: Among them, α T , α P and α V are the adjustment coefficients for temperature, pressure and speed, respectively. represents the prediction result of the integrated model at time t; Y(t) represents the actual detection result at time t; T prod (t+△t) represents the adjusted temperature at time t+△t; P extr (t+△t) represents the adjusted pressure at time t+△t; v extr (t+△t) represents the adjusted speed at time t+△t; S506. Regularly evaluate the performance of the integrated model using mean square error and coefficient of determination as evaluation indicators.

Images