Paper coating defect classification method and system based on infrared image recognition

Abstract

The application relates to the technical field of paper coating defect detection, and discloses a paper coating defect classification method and system based on infrared image recognition, which comprises the following steps: dividing the surface of coated paper into a plurality of detection regions; using a multi-agent distributed cooperative optimization algorithm, and distributing and sequencing detection tasks according to detection resources and progress to obtain an optimized partition detection execution sequence; based on a coupled vibration modal feature extraction algorithm and a convolutional neural network model, and under the guidance of the partition detection execution sequence, performing feature extraction on the original infrared image of each partition to obtain a multi-dimensional partition feature set containing infrared image features and aerodynamic modal features; and inputting the multi-dimensional partition features into different defect classification models respectively, and obtaining the final coating paper defect classification result of each partition through a weighted voting mode. The application realizes high-precision defect recognition of coated paper.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of paper coating defect detection, and in particular to a paper coating defect classification method and system based on infrared image recognition. BACKGROUND

[0002] With the increasing demand for paper products, its production process is also constantly improving. In order to ensure the quality of paper, the surface of the paper usually needs to be coated with a uniform layer of coating. However, in the rapid and large-scale production process, various defects may occur in the paper coating, such as uneven coating thickness, coating whitening, scratches, foreign matter, etc. These defects not only affect the visual effect of the coated product, but also weaken some functions of the coated product, such as water resistance, gloss, etc., which seriously affects the product quality, resulting in direct or indirect economic losses for the production enterprise.

[0003] In recent years, with the rapid development of computer vision technology and deep learning technology, it has become possible to use image processing and machine learning technology to automatically identify and classify coating defects, which not only improves the detection efficiency, but also improves the detection accuracy, providing important reference data for subsequent quality control and production optimization, and providing a new way for coating defect detection. The coating defect detection work gradually develops from manual operation to automatic and intelligent detection.

[0004] Among them, the image processing technology is mainly used in the classification process of paper coating defects. The coating defects are generally divided into surface quality defects and functional defects, among which the functional defects (such as coating discontinuity, coating damage, etc.) not only affect the visual effect of the coated product, but also may affect the functionality and reliability of the product. Due to the high overall quality requirement of the product, the accuracy requirement in the detection process is also correspondingly improved. This not only increases the difficulty and complexity of detection, but also leads to insufficient detail in defect classification, so that it is difficult to develop a clear treatment plan according to the classification results.

[0005] At present, there is no effective solution to the problems in the related art. SUMMARY

[0006] Therefore, the present application aims to overcome the problems of insufficient defect recognition accuracy, limited detection efficiency, and poor classification robustness in the existing paper coating defect classification technology, and provides a paper coating defect classification method and system based on infrared image recognition.

[0007] To this end, the specific technical solutions adopted by the present application are as follows:

[0008] According to one aspect of the present application, a paper coating defect classification method based on infrared image recognition is provided, comprising:

[0009] An original infrared image dataset of the coated paper is constructed according to the partition numbers, and the surface of the coated paper is divided into a plurality of detection regions according to the partition numbers;

[0010] A multi-agent distributed collaborative optimization algorithm is used, and detection tasks are allocated and sorted according to detection resources, detection progress, and task workloads of the plurality of detection regions, to obtain an optimized partition detection execution sequence;

[0011] Based on a coupled vibration modal feature extraction algorithm and a convolutional neural network model, and under the guidance of the partition detection execution sequence, features are extracted from the original infrared image of each partition to obtain a multi-dimensional partition feature set containing infrared image features and aerodynamic modal features;

[0012] The multi-dimensional partition features are respectively input into different defect classification models, and the final coated paper defect classification results of each partition are obtained through weighted voting.

[0013] Further, constructing an original infrared image dataset of the coated paper according to the partition numbers, and dividing the surface of the coated paper into a plurality of detection regions includes:

[0014] A standardized coated paper full-width original infrared image is obtained, and a spatially uniform partition grid division scheme is developed according to the required detection resolution and detection region division strategy, outputting the total number of partitions of each coated paper full-width original infrared image and the coordinate boundary parameters of each partition, and generating a globally unique partition number for each partition;

[0015] The coated paper full-width original infrared image is cut according to the partition grid coordinates, and a separate detection region subgraph is cut out for each partition, and a three-tuple set of partition number, detection region spatial coordinates and partition image subblock is obtained.

[0016] Further, using a multi-agent distributed collaborative optimization algorithm, and according to detection resources, detection progress, and task workloads of the plurality of detection regions, the detection tasks are allocated and sorted to obtain an optimized partition detection execution sequence.

[0017] The task workloads of each detection region are counted, and the state parameters of each detection agent are summarized;

[0018] The state parameters of each detection agent and the task workloads of the detection regions are combined to construct a weighted objective function, and a task priority list is generated for each detection agent, listing the candidate detection task list that the detection agent should undertake in the current round;

[0019] Based on the candidate detection task list, a multi-agent distributed collaborative optimization algorithm is applied to perform multi-round local and global scheduling optimization, and a global optimization allocation scheme is automatically generated according to preset resource utilization optimization, total progress minimization and detection task priority rules.

[0020] The partition number, designated detection agent and scheduling execution order information of each detection task in the global optimization allocation scheme are summarized, archived and generated according to a unique sequence coding structure to obtain an optimized partition detection execution sequence.

[0021] Further, based on the candidate detection task list, a multi-agent distributed collaborative optimization algorithm is applied to perform multi-round local and global scheduling optimization, and a global optimization allocation scheme is automatically generated according to preset resource utilization optimization, total progress minimization and detection task priority rules, which includes:

[0022] According to the candidate detection task list and the current state parameters of all detection agents, a set of strategy neural network parameters and a set of judge neural network parameters of multi-agent are constructed;

[0023] The strategy neural network of each detection agent is used to output a continuous action vector of each detection agent in the current task allocation decision space as input, which is used to represent the preference of each detection task being allocated by different agents;

[0024] According to the action result, a task allocation and sequencing scheme for the current round is formed, and a joint task allocation matrix is formed using the action set of all detection agents;

[0025] The joint task allocation scheme in the joint task allocation matrix is simulated, and the resource utilization, total detection progress and priority rule implementation of all detection agents under the allocation scheme are summarized;

[0026] According to the pre-defined multi-objective reward function, an actual reward value is generated for each agent;

[0027] The expected reward value and state value are calculated using the judge neural network, and the actual reward value is subtracted to obtain a loss function, and the weight parameters of the strategy neural network and the judge neural network of each detection agent are updated by back propagation to improve the global optimization ability of the comprehensive allocation strategy;

[0028] When the convergence criteria are met, a set of task allocation and sequencing results with optimal reward value or comprehensive evaluation index is selected from the historical allocation scheme as the global optimization allocation scheme.

[0029] Further, based on the coupled vibration modal feature extraction algorithm and the convolutional neural network model, and under the guidance of the partition detection execution sequence, the original infrared image of each partition is subjected to feature extraction to obtain a multi-dimensional partition feature set containing infrared image features and aerodynamic modal features, including:

[0030] The convolutional neural network model is used to extract spatial structure features of the partition image of each detection region; Fourier transform and wavelet transform are used to extract frequency domain and texture features, and the preliminary spatial frequency domain features are formed by fusion combined with the spatial structure features;

[0031] A piezoelectric exciter is used to apply air flow to induce controlled flutter of the local paper web, and a laser Doppler vibration detector is used to monitor the dynamic response data of the detection region;

[0032] The coupled vibration modal feature extraction algorithm is used to extract the vibration modal features of the detection region, and the amplitude anomaly index and the vortex dissipation feature are quantified; the amplitude anomaly index and the vortex dissipation feature are fused into the vibration modal physical feature;

[0033] The preliminary spatial frequency domain features and the vibration modal physical features are fused to form multi-dimensional partition features in accordance with the partition detection execution sequence.

[0034] Further, the coupled vibration modal feature extraction algorithm is used to extract the vibration modal features of the detection region, and the amplitude anomaly index and the vortex dissipation feature are quantified, including:

[0035] The dynamic response data is subjected to standardization processing to obtain the regional vibration response net signal; the modal decomposition technology is used to restore the regional vibration response net signal into the main modal and high-order modal response components;

[0036] The spatial distribution function and the time history function of the main modal and high-order modal response components are extracted to form a set of space-time response functions for describing the modal behavior;

[0037] The fluid dynamics equation and the coating paper dynamics model are used to cooperatively model to calculate the vorticity field data within the detection partition range of the coating paper;

[0038] The envelope, peak value and frequency spectrum features of the space-time response function are extracted, the deviation ratio of the actual modal amplitude and the reference amplitude is calculated combined with the historical normal value and the theoretical reference standard, and the amplitude anomaly index is obtained;

[0039] The energy flow density and the vortex dissipation rate calculation method are used for the vorticity field data combined with the fluid dynamics dissipation model to quantitatively extract the vortex dissipation feature of the detection partition.

[0040] Further, the fluid dynamics equation and the coating paper dynamics model are used to cooperatively model to calculate the vorticity field data within the detection partition range of the coating paper, including:

[0041] According to the thin plate theory, a solid dynamics model of the coated paper is constructed; a Navier-Stokes equation set and boundary conditions coupled with the paper and air interface are constructed;

[0042] The Navier-Stokes equation set in the time and space domain and the solid dynamics model of the coated paper are cooperatively numerically solved to obtain spatial and temporal distribution data of a velocity field and a pressure field in a detection region;

[0043] The spatial and temporal distribution data of the velocity field and the pressure field in the detection region are used to calculate vorticity field data of a detection subregion of the coated paper.

[0044] Further, the vorticity field data is subjected to an energy flow density and a vortex dissipation rate calculation mode, combined with a fluid dynamics dissipation model, to quantitatively extract vortex dissipation features of the detection subregion, including:

[0045] According to the vorticity field data and the velocity field, and according to a kinetic energy flow density definition principle, an energy flow density field of each time-space point in the detection subregion is calculated;

[0046] According to the vorticity field data and the velocity field, and according to a fluid dynamics dissipation model, vortex dissipation rate field data of the detection subregion is calculated;

[0047] Based on the energy flow density field and the vortex dissipation rate field data, vortex dissipation features are extracted and reduced.

[0048] Further, the multi-dimensional subregion features are respectively input into different defect classification models, and through a weighted voting manner, a final coated paper defect classification result of each subregion is obtained, including:

[0049] The multi-dimensional subregion features are input into different pre-trained defect classification models to obtain defect prediction results and class probability scores of the defect classification models;

[0050] According to a preset model weighting strategy, the class probability scores of each defect classification model are respectively summed according to weights to obtain weighted voting scores of each class;

[0051] The class with the highest weighted score is selected as the final coated paper defect classification result of the corresponding detection subregion.

[0052] According to another aspect of the present application, there is also provided a paper coating defect classification system based on infrared image recognition, comprising a detection area division module, an execution sequence determination module, a multi-dimensional partition feature acquisition module and a defect classification module: the detection area division module is used to construct a partition number-archived original infrared image data set of the coated paper, and divide the surface of the coated paper into a plurality of detection areas according to the partition number; the execution sequence determination module is used to distribute and sort the detection tasks by using a multi-agent distributed collaborative optimization algorithm and according to the detection resources, the detection progress and the respective task workloads of the plurality of detection areas, to obtain an optimized partition detection execution sequence; the multi-dimensional partition feature acquisition module is used to extract features from the original infrared image of each partition based on a coupled vibration modal feature extraction algorithm and a convolutional neural network model, and obtain a multi-dimensional partition feature set containing infrared image features and aerodynamic modal features under the guidance of the partition detection execution sequence; and the defect classification module is used to input the multi-dimensional partition features into different defect classification models respectively, and obtain the final coated paper defect classification results of each partition by using a weighted voting method.

[0053] The present application includes the following beneficial effects:

[0054] (1) In each detection partition, the present application not only fuses the high-level spatial structure and texture image features extracted by the convolutional neural network, but also introduces physical indicators such as vibration modal features, energy flow density and vortex dissipation, to realize the synchronous and accurate identification of surface quality and functional defects. In the detection of high-difficulty defects such as functionality (such as micro-damage and coating discontinuity), the present application surpasses the single image processing paradigm, improves the sensitivity and discrimination of physical nature abnormalities, and effectively reduces the missed detection and false detection rates.

[0055] (2) By introducing a multi-agent distributed collaborative optimization algorithm, the present application optimizes the partition distribution and execution sequence of the detection tasks in real time in combination with the partition task workloads and the real-time state of the detection agents, which not only maximizes the resource utilization rate and shortens the overall detection period, but also has strong general adaptability and can dynamically adapt to different production rhythms, abnormal working conditions or resource scheduling constraints.

[0056] (3) The present application integrates different types of defect classification models (such as convolutional neural network, random forest classification model and extreme gradient boosting classification model) into a multi-model integrated decision system, which collectively votes through the respective responses to multi-source deep features, and adopts weighted probability fusion, effectively alleviating the performance fluctuations of single models under specific categories or interference working conditions, and improving the robustness and generalization ability of the overall classification. BRIEF DESCRIPTION OF DRAWINGS

[0057] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed in the embodiments will be briefly introduced as follows. Obviously, the drawings in the following description only constitute some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained based on these drawings without creative effort.

[0058] Figure 1 is a flow chart of a paper coating defect classification method based on infrared image recognition according to an embodiment of the present application;

[0059] Figure 2 is a principle block diagram of a paper coating defect classification system based on infrared image recognition according to an embodiment of the present application;

[0060] Figure 3 is a flow chart of partition detection execution sequence optimization in a paper coating defect classification method based on infrared image recognition according to an embodiment of the present application.

[0061] Legend of reference signs:

[0062] 1, detection area division module; 2, execution sequence determination module; 3, multi-dimensional partition feature acquisition module; 4, defect classification module. DETAILED DESCRIPTION

[0063] In order to make the purpose, technical solutions and advantages of the present application more clear, the present application will be further described in detail below in combination with the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not intended to limit the present application.

[0064] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which the present application belongs. The terms used herein are only for the purpose of describing the embodiments of the present application and are not intended to limit the present application.

[0065] According to an embodiment of the present application, a paper coating defect classification method and system based on infrared image recognition are provided.

[0066] The present application will be further described in combination with the drawings and specific embodiments. As shown in Figure 1 and Figure 3 According to an embodiment of the present application, a paper coating defect classification method based on infrared image recognition is provided, which comprises:

[0067] S1, constructing a partition number archived original infrared image data set of coated paper, and dividing the surface of the coated paper into a plurality of detection areas according to the partition number.

[0068] In one embodiment, the original infrared image data set of the partition number archive of the coated paper is constructed, and the coated paper surface is divided into a plurality of detection areas according to the partition number, including:

[0069] The standardized coated paper full-width original infrared image is obtained, and a spatially uniform partition grid division scheme is formulated according to the required detection resolution and the detection area division strategy, and the total number of partitions of each coated paper full-width original infrared image and the coordinate boundary parameters of each partition are output, and a globally unique partition number is generated for each partition; the coated paper full-width original infrared image is cut according to the partition grid coordinates, and a separate detection area subgraph is cut out for each partition, and a three-tuple set of partition number, detection area spatial coordinates and partition image subblock is obtained by combination.

[0070] Through the partition number archive and the spatial grid division mechanism, the spatial resolution and information organization efficiency of the infrared detection data are improved. Each detection partition corresponds to a standardized detection unit, which is helpful for subsequent realization of regional quality defect precise positioning, batch tracing and historical comparison analysis. At the same time, the binding of the partition number and the spatial coordinates enables the defect detection, result reporting and production control to focus on the fine-grained physical space as needed, realizes the rapid locking, hierarchical management and visualized tracing of defects. For example, on the coated paper, it is divided into 20 rows x 10 columns according to the detection resolution requirement, totaling 200 partitions. For each partition, such as the 8th row and the 5th column, a unique number No.0805 is assigned, and its physical coordinate range (such as the pixel point of the upper left corner / the lower right corner) is recorded, and the corresponding infrared image subblock is cut out.

[0071] S2, using a multi-agent distributed collaborative optimization algorithm, and according to the detection resources and progress, allocating and sorting the detection tasks to obtain an optimized partition detection execution sequence.

[0072] In one embodiment, a multi-agent distributed collaborative optimization algorithm is used, and the detection tasks are allocated and sorted according to the detection resources, the detection progress and the task workload of each of the plurality of detection areas, to obtain an optimized partition detection execution sequence.

[0073] The task workload of each detection area is counted, and the state parameters of various detection agents are summarized, including operation speed parameters, load capacity parameters, available time period parameters, and operation state parameters of detection hardware; a weighted objective function is constructed in combination with the state parameters of the detection agents and the task workload of the detection area, a task priority list is generated for each detection agent, and a candidate detection task list is listed for the detection agent to preferentially undertake in the current round; based on the candidate detection task list, a multi-agent distributed collaborative optimization algorithm (such as a multi-agent deep deterministic policy gradient model, a multi-agent genetic algorithm collaborative optimization model, etc.) is applied to perform multi-round local and global scheduling optimization, and a global optimization allocation scheme is automatically generated according to the preset resource utilization optimization, shortest total progress, and detection task priority rules; the partition number, designated detection agent, and scheduling execution sequence information of each detection task in the global optimization allocation scheme are summarized, archived and generated according to a unique sequence coding structure, and an optimized partition detection execution sequence is obtained.

[0074] The multi-agent distributed collaborative optimization algorithm is an algorithm for realizing adaptive, dynamic, and optimal allocation and scheduling of coating paper partition detection tasks through distributed collaboration and game among multiple autonomous agents (Agents) and using intelligent algorithms such as reinforcement learning or evolutionary optimization. It allows multiple intelligent individuals to work and cooperate in a shared environment and continuously iteratively solve the optimal global scheme while learning. The algorithm includes an Agent state perception module, a multi-agent policy decision module, a joint task allocation simulation / execution module, a global multi-objective reward mechanism, a referee network / batch archiving and policy reverse updating and self-learning. Mainly, each agent has an independent or partially shared policy neural network that outputs actions (allocation, priority, etc. decision vectors) after inputting global or local states; according to the collective action, the actual arrangement of work is arranged, or the simulation is run or the actual execution is performed; the resource utilization rate, total task progress, and task priority are evaluated to determine the pros and cons of the current scheduling, and a quantitative incentive is given to all agents; the value function and loss are calculated to evaluate the decision effect, which is used to accelerate training and experience replay; gradient descent / genetic evolution is used to gradually optimize the neural network or control strategy.

[0075] In one embodiment, based on the candidate detection task list, a multi-agent distributed collaborative optimization algorithm is applied to perform multi-round local and global scheduling optimization, and a global optimization allocation scheme is automatically generated according to the preset resource utilization optimization, shortest total progress, and detection task priority rules, including:

[0076] According to the candidate detection task list and the current state parameters of all detection agents, a set of strategy neural network parameters and a set of judge neural network parameters of the multi-agent are constructed; the strategy neural network of each detection agent is used to output a continuous action vector of each detection agent for the current task allocation decision space as an input of the candidate detection task and the current resource state, which is used to represent the preference of each detection task allocated by different agents; according to the action result, a task allocation and sorting scheme of the current round is formed, and a joint task allocation matrix is formed by using the action set of all detection agents; the joint task allocation scheme in the joint task allocation matrix is simulated, and the resource utilization, total detection progress and priority rule implementation of all detection agents under the allocation scheme are summarized; according to the pre-defined multi-objective reward function (such as detection resource utilization, total progress shortest, task priority weighting and the like), actual reward values are generated for each agent as the reinforcement learning training signal of the strategy network; the expected reward value and the state value are calculated by using the judge neural network, and the difference between the actual reward value and the expected reward value is obtained to obtain a loss function, and the weight parameters of the strategy neural network and the judge neural network of each detection agent are updated by back propagation to improve the global optimization ability of the comprehensive allocation strategy; when the convergence standard is met, a set of task allocation and sorting results with the optimal reward value or comprehensive evaluation index is selected from the historical allocation scheme as the global optimization allocation scheme.

[0077] By using the multi-agent distributed collaborative optimization algorithm, the complex coated paper partition detection task is decomposed and intelligently allocated by mobilizing multiple heterogeneous detection agents. Each detection area is first assigned a specific task workload indicator. According to the detection workload and agent capability parameters, a weighted multi-objective optimization function (covering resource utilization, partition priority, total detection progress, etc.) is constructed. Through reinforcement learning or evolutionary intelligent algorithms such as multi-agent deep deterministic policy gradient model or multi-agent genetic collaborative optimization model, the strategy neural network of the agent outputs the action corresponding to the task allocation in each round, forms the allocation matrix, and combines the judge neural network to give reward update based on historical performance and immediate results. This mechanism takes into account both local and global optimization in iterations, continuously learns itself until it converges to a globally optimized task allocation and sorting. Under the multi-index trade-off (such as maximizing equipment utilization, minimizing total detection time, and detecting critical partitions), compared with the traditional static or manual allocation mode, the present application can adapt to the task load in real time, significantly improve the overall throughput and mutual backup fault tolerance of the detection pipeline.

[0078] For example, five parallel detection devices with different parameters are set up on a coated paper production line: A device has strong computing power but high load, B device is better for local coating thickness and pinhole detection in edge area, suitable for smaller partition tasks, C, D, E have different available time and health status. The current paper web is divided into 10 unequal-width detection partitions, and the task amount of each partition is different. Collect the current state of all detection devices, based on multi-agent deep reinforcement learning optimization algorithm, evaluate the preference of each device for each task "action" in each round, and comprehensively adjust the strategy based on historical data and real-time simulation feedback (such as device utilization, total progress, priority). After thousands of rounds of iterative optimization, a set of optimal partition task allocation and sequence list is finally output, realizing the maximum exertion of device capacity, priority completion of key tasks, and the shortest overall detection time.

[0079] S3, based on the coupled vibration modal feature extraction algorithm and the convolutional neural network model, and under the guidance of the partition detection execution sequence, the original infrared image of each partition is feature extracted to obtain a multi-dimensional partition feature set containing infrared image features and aerodynamic modal features.

[0080] In one embodiment, based on the coupled vibration modal feature extraction algorithm and the convolutional neural network model, and under the guidance of the partition detection execution sequence, the original infrared image of each partition is feature extracted to obtain a multi-dimensional partition feature set containing infrared image features and aerodynamic modal features, including:

[0081] A convolutional neural network model is used to extract spatial structure features of the partition image of each detection area; Fourier transform and wavelet transform are used to extract frequency domain and texture features, and combined with spatial structure features, preliminary spatial frequency domain features are formed by fusion; a piezoelectric exciter is used to apply controllable airflow to induce controlled flutter of the local paper web, and a laser Doppler vibration meter is used to monitor the dynamic response data of the detection area during the sampling period; a coupled vibration modal feature extraction algorithm, Kooptics algorithm (combined with Navier-Stokes fluid dynamics and thin plate vibration model), is used to extract vibration modal features of the detection area, and to quantify amplitude anomaly index and vortex dissipation features; the amplitude anomaly index and the vortex dissipation features are fused into vibration modal physical features; the preliminary spatial frequency domain features and the vibration modal physical features are fused to form multi-dimensional partition features that meet the partition detection execution sequence.

[0082] The coupled vibration modal feature extraction algorithm is applied to the dynamic detection of the partition of the coated paper by applying the fluid-solid coupling dynamics principle to the vibration intrinsic characteristics of the paper web under the air dynamic disturbance, that is, the air flow disturbance and vortex dissipation of the detection area are modeled by using the Navier-Stokes equation set, and the structural dynamic response of the coated paper is described by using the thin plate theory. In actual detection, the piezoelectric exciter is used to locally apply air flow to the area to induce controlled flutter, and the laser Doppler vibration instrument is used to obtain the time domain response of the area vibration; the response waveform is subjected to modal decomposition to obtain the spatial-time distribution of the main mode and the high-order mode, and the amplitude anomaly (compared with the reference normal paper response) and the vortex energy dissipation characteristics are calculated in combination with the aerodynamic data. The above reflects the physical nature of the local stiffness, damage, uneven thickness and surface functional abnormalities of the coated paper under the action of aerodynamics. The core of the coupled vibration modal feature extraction algorithm is the modal decomposition module, the fluid-solid coupling modeling module and the feature extraction module, that is, the original response is separated into basic mode and high-order mode components by using signal transformation or modal identification technology, and the behavior mode in time and space is output; the fluid (Navier-Stokes) and solid (thin plate) dynamics models are solved to obtain the aerodynamic parameters such as vortex field, velocity field and pressure field; the quantitative features are extracted from the physical indicators such as amplitude anomaly, vortex energy flow density and dissipation rate, and finally the complete and physically meaningful multi-dimensional feature vector is formed as the input of the coated paper defect diagnosis and classification by fusing the spatial frequency domain features.

[0083] In one embodiment, the coupled vibration modal feature extraction algorithm is used to extract the vibration modal features of the detection area, and the amplitude anomaly index and the vortex dissipation characteristics are quantified, which includes:

[0084] The dynamic response data is subjected to standardization processing to obtain the net signal of the area vibration response; the modal decomposition technology (such as orthogonal decomposition or empirical mode decomposition) is used to restore the net signal of the area vibration response to the main mode and high-order mode response components; the spatial distribution function and the time history function of the main mode and high-order mode response components are extracted to form a set of time-space response functions for describing the modal behavior; the Navier-Stokes fluid dynamics equation and the coated paper dynamics model are used to cooperatively model the vortex field data in the detection partition range of the coated paper; the envelope, peak value and frequency spectrum features of the time-space response functions are extracted, the deviation ratio of the actual modal amplitude to the reference amplitude is calculated in combination with the historical normal value and the theoretical reference standard, and the amplitude anomaly index is obtained; the vortex dissipation characteristics of the detection partition are quantitatively extracted by using the energy flow density and the vortex dissipation rate calculation method in combination with the fluid dynamics dissipation model, and the quantitative description of the physical instability anomaly is realized.

[0085] In one embodiment, the vorticity field data in the detection area of the coated paper is calculated by using the fluid dynamics equation in cooperation with a coated paper dynamics model, including:

[0086] According to the thin plate theory, a solid dynamics model of the coated paper is constructed, a Navier-Stokes equation set coupled with the paper and air interface and boundary conditions are constructed, the Navier-Stokes equation set and the solid dynamics model of the coated paper in the time and space domain are cooperatively numerically solved to obtain the spatial and temporal distribution data of the velocity field and the pressure field in the detection area, and the vorticity field data (i.e., the vorticity field time and space distribution data) of the detection area of the coated paper is calculated by using the spatial and temporal distribution data of the velocity field and the pressure field in the detection area.

[0087] In one embodiment, the vorticity field data is calculated by using the energy flow density and the vortex dissipation rate, and the vortex dissipation features of the detection area are quantitatively extracted by using the fluid dynamics dissipation model, including:

[0088] According to the vorticity field data and the velocity field, and according to the kinetic energy flow density definition principle, the energy flow density field of each time and space point in the detection area is calculated, according to the vorticity field data and the velocity field, and according to the fluid dynamics dissipation model, the vortex dissipation rate field data of the detection area is calculated, and the vortex dissipation features are extracted and reduced based on the energy flow density field and the vortex dissipation rate field data.

[0089] By the time and space joint modeling and the modal feature extraction, the application can not only accurately capture the inherent dynamic behavior of the paper web under normal and abnormal states, but also quantize the microstructure changes caused by the air dynamic coupling, and enhance the detection capability for the hidden problems such as the paper surface microdefects, delamination and thickness abnormalities. In addition, by the calculation of the energy flow density and the vortex dissipation rate, the local energy dissipation and the flow field instability under the air flow disturbance can be finely described, and the quantitative early warning for the potential local quality hidden danger can be realized.

[0090] In actual application, when a certain area is randomly inspected, firstly, the air flow exciter is triggered to trigger the controlled disturbance, and the response curve of the area is obtained by using the vibration meter. After the data is subjected to empirical mode decomposition, the main modal peak value is 0.8 mm, and the high-order modal locally exists abnormal increase. By comparison calculation with the historical normal sample, the amplitude abnormality index is 0.25, which indicates that there is a structural micro-variation. Then, the local time and space vorticity distribution is obtained by numerically simulating the velocity field and the pressure field, and the energy flow density and the vortex dissipation of each grid point are calculated. For example, the energy flow density of a certain abnormal area is significantly higher than that of the normal area, and the vortex dissipation rate is 1.5 times that of the normal area, which can assist in determining that there are problems such as structural looseness, abnormal fiber arrangement or local material damage in the area, for subsequent intelligent grading processing or manual re-inspection confirmation.

[0091] S4, input the multi-dimensional partition features into different defect classification models respectively, and obtain the final coating paper defect classification result of each partition through weighted voting.

[0092] In one embodiment, the method comprises:

[0093] inputting the multi-dimensional partition features into different pre-trained defect classification models, obtaining the defect prediction result (single defect or multiple defects) and the probability score of each category of the defect classification model, summing up the probability scores of each category of the defect classification model according to a preset model weighting strategy, and obtaining the weighted voting score of each category; and selecting the category with the highest weighted score as the final coating paper defect classification result of the corresponding detection partition.

[0094] The present application integrates the advantages and discriminant details of different defect detection models through the integration of multi-model weighted voting, avoids local errors caused by misjudgment or bias of a single model, and improves the generalization performance and decision accuracy of the model through weighted fusion.

[0095] As shown in FIG. 1, Figure 2 According to another embodiment of the present application, a paper coating defect classification system based on infrared image recognition is also provided, which comprises a detection area division module 1, an execution sequence determination module 2, a multi-dimensional partition feature acquisition module 3, and a defect classification module 4.

[0096] The detection area division module 1 is used to construct a partition number-archived original infrared image dataset of the coating paper, and divide the surface of the coating paper into a plurality of detection areas according to the partition numbers.

[0097] The execution sequence determination module 2 is used to utilize a multi-agent distributed collaborative optimization algorithm, and allocate and sort the detection tasks according to the detection resources, the detection progress, and the task workload of each of the plurality of detection areas, to obtain an optimized partition detection execution sequence.

[0098] The multi-dimensional partition feature acquisition module 3 is used to acquire the multi-dimensional partition features based on a coupled vibration modal feature extraction algorithm and a convolutional neural network model, and extract features from the original infrared image of each partition under the guidance of the partition detection execution sequence, to obtain a multi-dimensional partition feature set containing infrared image features and aerodynamic modal features.

[0099] The defect classification module 4 is used to input the multi-dimensional partition features into different defect classification models respectively, and obtain the final coating paper defect classification result of each partition through weighted voting.

[0100] In summary, in each detection partition, the application not only extracts high-level spatial structure and texture image features by fusion convolutional neural network, but also introduces vibration modal features, energy flow density and vortex dissipation and other physical indicators, to realize the synchronous and accurate identification of surface quality and functional defects. In the detection of high-difficulty defects such as function (such as micro-damage, coating discontinuity), the application surpasses the single image processing paradigm, improves the sensitivity and discrimination of physical nature abnormalities, and effectively reduces the missed detection and false detection rate. The application introduces a multi-agent distributed collaborative optimization algorithm, combines the partition task workload and the real-time state of the detection agent, and optimizes the partition allocation and execution order of the detection task in real time, which not only maximizes the resource utilization and shortens the overall detection cycle, but also has strong general adaptability and can dynamically adapt to different production rhythms, abnormal working conditions or resource scheduling constraints. Different types of defect classification models (such as convolutional neural network, random forest classification model, extreme gradient boosting classification model, etc.) are included in the multi-model integrated decision system, the responses of each to the multi-source deep features are collectively voted, and the weighted probability fusion is adopted, which effectively alleviates the performance fluctuation of a single model under certain categories or interference conditions, and improves the robustness and generalization ability of the overall classification.

[0101] The above only describes the preferred embodiments of the application and is not intended to limit the application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the application shall be included in the protection scope of the application.

Claims

1. A paper coating defect classification method based on infrared image recognition, characterized by, include: A dataset of original infrared images archived with partition numbers of coated paper is constructed, and the surface of the coated paper is divided into several detection regions according to the partition numbers; specifically including: Obtain standardized full-frame raw infrared images of coated paper, and formulate a spatially uniform partitioning grid scheme according to the required detection resolution and detection area division strategy. Output the total number of partitions and the coordinate boundary parameters of each partition in the full-frame raw infrared image of coated paper, and generate a globally unique partition number for each partition. The original infrared image of the coated paper is cut out according to the grid coordinates of the partitions. Each partition is cut out into a separate detection area sub-image, and the partition number, detection area spatial coordinates and partition image sub-block triplet set are combined to obtain the partition number, detection area spatial coordinates and partition image sub-block triplet set. By utilizing a multi-agent distributed collaborative optimization algorithm, and based on detection resources, detection progress, and the workload of each of the several detection regions, detection tasks are allocated and sorted to obtain an optimized partitioned detection execution sequence. Based on a coupled vibration mode feature extraction algorithm and a convolutional neural network model, and guided by a partition detection execution sequence, features are extracted from the original infrared images of each partition to obtain a multi-dimensional partition feature set containing both infrared image features and aerodynamic mode features; specifically including: A convolutional neural network model is used to extract spatial structure features from the partitioned images of each detection region; Fourier transform and wavelet transform are used to extract frequency domain and texture features, and combined with spatial structure features, they are fused to form preliminary spatial frequency domain features; An airflow is applied using a piezoelectric vibrator to induce controlled vibration in a localized area of ​​the paper web, and a laser Doppler vibration meter is used to monitor the dynamic response data of the detection area. The vibration modal features of the detection area are extracted using a coupled vibration modal feature extraction algorithm, and the amplitude anomaly index and vortex dissipation features are quantified. The amplitude anomaly index and vortex dissipation features are then fused into the vibration modal physical features. The preliminary spatial frequency domain features are fused with the physical features of vibration modes to form a multi-dimensional partition feature that conforms to the partition detection execution sequence. The multidimensional partition features are input into different defect classification models, and the final coating paper defect classification result for each partition is obtained through weighted voting.

2. The paper coating defect classification method based on infrared image recognition according to claim 1, characterized in that, The optimized partitioned detection execution sequence, obtained by utilizing a multi-agent distributed collaborative optimization algorithm and allocating and sorting detection tasks based on detection resources, detection progress, and the workload of each of the several detection regions, includes: The workload of each detection area is calculated, and the state parameters of each detection agent are summarized. By combining the state parameters of each detection agent with the workload of the detection area, a weighted objective function is constructed to generate a task priority list for each detection agent, and a list of candidate detection tasks that the detection agent should take priority in the current round is listed. Based on the candidate detection task list, a multi-agent distributed collaborative optimization algorithm is applied to perform multiple rounds of local and global scheduling optimization. Based on the preset rules of optimal resource utilization, shortest total progress and detection task priority, a global optimization allocation scheme is automatically generated. The partition number, designated detection agent and scheduling execution sequence information of each detection task in the global optimization allocation scheme are summarized, archived and generated according to a unique sequence coding structure, and an optimized partition detection execution sequence is obtained.

3. The paper coating defect classification method based on infrared image recognition according to claim 2, characterized in that, The multi-agent distributed collaborative optimization algorithm is applied based on the candidate detection task list to perform multi-round local and global scheduling optimization, and a global optimization allocation scheme is automatically generated according to preset resource utilization optimization, shortest total progress and detection task priority rules, including: According to the candidate detection task list and the current state parameters of all detection agents, a set of strategy neural network parameters and a set of judge neural network parameters of the multi-agent are constructed; The strategy neural network of each detection agent is used to output a continuous action vector of each detection agent in the current task allocation decision space as input, which represents the preference of each detection agent for the current task allocation. According to the action result, a task allocation and sorting scheme for the current round is formed, and a joint task allocation matrix is formed using the action set of all detection agents. The joint task allocation scheme in the joint task allocation matrix is simulated, and the resource utilization, total detection progress and priority rule implementation of all detection agents under the allocation scheme are summarized. According to the pre-defined multi-objective reward function, the actual reward value of each agent is generated. The expected reward value and state value are calculated using the judge neural network, and the difference between the actual reward value and the expected reward value is obtained to form a loss function, which is used to update the weight parameters of the strategy neural network and the judge neural network of each detection agent to improve the global optimization capability of the comprehensive allocation strategy. When the convergence criteria are met, a set of task allocation and sorting results with the optimal reward value or comprehensive evaluation index is selected from the historical allocation schemes as the global optimization allocation scheme.

4. The paper coating defect classification method based on infrared image recognition according to claim 1, characterized in that, The vibration modal characteristics of the detection area are extracted using the coupled vibration modal feature extraction algorithm, and the amplitude anomaly index and vortex dissipation characteristics are quantified, including: The dynamic response data is standardized to obtain the area vibration response net signal; the modal decomposition technology is used to restore the area vibration response net signal to the main modal and high-order modal response components; The spatial distribution function and time history function of the main modal and high-order modal response components are extracted to form a set of time-space response functions for describing the modal behavior; The fluid dynamics equation and the coating paper dynamics model are used to cooperatively model the vortex field data within the coating paper detection partition range; The envelope, peak value and frequency spectrum characteristics of the time-space response function are extracted, and the deviation ratio of the actual modal amplitude and the reference amplitude is calculated based on the historical normal value and the theoretical reference standard to obtain the amplitude anomaly index; The energy flow density and vortex dissipation rate calculation method are used to quantitatively extract the vortex dissipation characteristics of the detection partition based on the fluid dynamics dissipation model.

5. The method for classifying paper coating defects based on infrared image recognition according to claim 4, characterized in that, The fluid dynamics equation and the coating paper dynamics model are used to cooperatively model the vortex field data within the coating paper detection partition range, including: According to the thin plate theory, a solid dynamics model of the coated paper is constructed; a Navier-Stokes equation set coupled with the paper and air interface and boundary conditions are constructed; The Navier-Stokes equation set in the time and space domain and the solid dynamics model of the coated paper are cooperatively numerically solved to obtain spatial and temporal distribution data of a velocity field and a pressure field in a detection region; The spatial and temporal distribution data of the velocity field and the pressure field in the detection region are used to calculate vorticity field data of a detection subregion of the coated paper.

6. The paper coating defect classification method based on infrared image recognition according to claim 4, characterized in that, The vorticity field data is calculated by using an energy flow density and a vortex dissipation rate calculation method, combined with a fluid dynamics dissipation model, to quantitatively extract vortex dissipation features of the detection subregion, including: According to the vorticity field data and the velocity field, and according to the kinetic energy flow density definition principle, an energy flow density field of each spatiotemporal point in the detection subregion is calculated; According to the vorticity field data and the velocity field, and according to the fluid dynamics dissipation model, vortex dissipation rate field data of the detection subregion is calculated; Based on the energy flow density field and the vortex dissipation rate field data, vortex dissipation features are extracted and reduced.

7. The method for classifying paper coating defects based on infrared image recognition according to claim 1, characterized in that, The multi-dimensional subregion features are respectively input into different defect classification models, and the final coated paper defect classification results of each subregion are obtained by using a weighted voting method, including: The multi-dimensional subregion features are input into different pre-trained defect classification models to obtain defect prediction results and probability scores of each category of the defect classification models; According to a preset model weighting strategy, the category probability scores of each defect classification model are summed according to weights to obtain weighted voting scores of each category; The category with the highest weighted score is selected as the final coated paper defect classification result of the corresponding detection subregion.

8. A paper coating defect classification system based on infrared image recognition for implementing the paper coating defect classification method based on infrared image recognition according to any one of claims 1 to 7, characterized in that The system includes a detection region division module, an execution sequence determination module, a multi-dimensional subregion feature acquisition module, and a defect classification module: The detection region division module is configured to construct a subregion number-archived original infrared image dataset of the coated paper, and divide the surface of the coated paper into a plurality of detection regions according to the subregion numbers; The execution sequence determination module is configured to use a multi-agent distributed cooperative optimization algorithm, and allocate and sort detection tasks according to detection resources, detection progress, and task workloads of the plurality of detection regions, to obtain an optimized subregion detection execution sequence; The multi-dimensional subregion feature acquisition module is configured to use a coupled vibration modal feature extraction algorithm and a convolutional neural network model, and extract features from the original infrared image of each subregion under the guidance of the subregion detection execution sequence, to obtain a multi-dimensional subregion feature set containing infrared image features and aerodynamic modal features; The defect classification module is configured to input the multi-dimensional subregion features into different defect classification models, and obtain the final coated paper defect classification results of each subregion by using a weighted voting method.

Images