Automatic defect identification system and method for woven carpet production

Abstract

The invention discloses an automatic defect identification system and method for woven carpet production, and relates to the technical field of textile industry automation, and the method comprises the steps: carrying out the synchronous scanning of the surface of transferred base cloth through a multispectral image collection device, and obtaining an original image data stream; preprocessing the image data flow and dynamically delimiting an effective analysis area; extracting texture and scattering characteristics in the effective area, and positioning a defect candidate area; synchronously collecting process parameters and carrying out space-time correlation on the process parameters and defect candidate areas; outputting defect types and causes by using an intelligent classification and traceability model; and generating a regulation instruction or a marking instruction according to the diagnosis result. According to the invention, continuous online detection of the whole process from transfer printing to cutting of the woven carpet can be realized, the interception rate of early repairable defects is improved, the systematic fault positioning time is shortened, and the automation level of a production line and the one-time qualified rate of products are remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of automation technology in the textile industry, and in particular to an automatic defect identification system and method for machine-woven carpet production. Background Technology

[0002] Machine-woven carpets, as an important component of home decoration and functional flooring materials, have extremely high requirements for pattern integrity, color reproduction, and surface smoothness during their production process. With the development of intelligent manufacturing technology, carpet production is gradually evolving towards automation and continuous processes. Among these, transfer printing, due to its high precision, high efficiency, and ability to adapt to complex patterns, has become the mainstream pattern application method. This process permanently attaches the pattern by transferring a pre-printed pattern on a special carrier to the surface of the base fabric under high temperature and pressure. During this process, the system must ensure precise alignment between the carrier and the base fabric, stable control of hot-pressing parameters, and strict guarantee of environmental cleanliness. Deviations in any step can lead to defects such as color difference, missing patterns, ghosting, or physical deformation in the final product. Therefore, building an automated identification and quality monitoring system that spans the entire process from printing, conveying, transfer printing, rolling up to cutting has become a core requirement for improving the yield and production consistency of machine-woven carpets.

[0003] Among these technologies, automated transfer printing systems based on roll-to-roll continuous production have become a key technological path in modern carpet manufacturing. This system requires the fabric to continuously pass through heated rollers at a constant tension, completing the complete transfer of the pattern from the carrier to the base fabric under a constant temperature environment of 220°C and precise pressure. Since the carrier leaves virtually no color residue after transfer, becoming waste paper, its surface condition cannot be used for subsequent quality traceability. Therefore, defect detection must be performed on the base fabric in real time after the transfer is complete. Simultaneously, to ensure transfer quality, the carrier must be absolutely clean before entering the hot-pressing zone. Tiny dust particles or fiber impurities in the environment, once adhering to the carrier surface, will solidify at high temperatures and hinder dye transfer, causing localized pattern loss. Furthermore, fabric wrinkles caused by tension fluctuations or guide roller misalignment during continuous operation can directly lead to pattern distortion or breakage. These dynamic defects are sudden and non-repeatable, making them difficult to effectively intercept using traditional offline sampling methods.

[0004] Existing technologies for defect identification in woven carpets still have significant limitations. First, most systems only have visual inspection modules at the cutting stage, failing to cover the critical intermediate processes from transfer printing to cutting, resulting in the omission of some repairable early defects. Second, existing image recognition algorithms are mostly designed for regular rectangular areas, making it difficult to adapt to edge variations in irregularly shaped carpets such as cloud shapes and free contours, easily leading to misjudgments or missed detections at corners or areas with abrupt curvature changes. Third, the systems lack the ability to analyze the correlation between transfer process parameters and defect types, failing to distinguish between systemic defects caused by equipment malfunctions and random defects caused by accidental contamination, making fault tracing difficult. Finally, although some production lines have achieved integrated printing, conveying, winding, and cutting, data silos between subsystems are severe, failing to establish a closed-loop feedback mechanism from process parameters to surface quality, and unable to achieve adaptive control based on real-time defect information. Summary of the Invention

[0005] The purpose of this invention is to provide an automatic defect identification system and method for machine-woven carpet production, so as to solve the problems of pattern defects, wrinkles and insufficient cutting accuracy in the transfer process of the prior art.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] An automatic defect identification system for machine-woven carpet production, the system comprising the following components:

[0008] The full-process image acquisition module is deployed on the continuous production line after transfer and before cutting. It is used to acquire multispectral images of the base fabric surface after transfer and generate raw image sequences containing visible light and near-infrared spectral information.

[0009] The multimodal data processing module is communicatively connected to the full-process image acquisition module. It is used to receive the original image sequence and perform image preprocessing, feature extraction and defect candidate region localization. It integrates a morphological adaptive segmentation unit to dynamically adjust the image analysis boundary according to the preset irregular carpet outline data.

[0010] The process parameter synchronization and correlation analysis module interacts with the production line control system in real time to synchronously acquire process parameters such as temperature, pressure, tension and linear speed during the transfer process, and establishes a correlation mapping between the time sequence of process parameters and the spatiotemporal information of defect candidate areas.

[0011] The intelligent classification and traceability decision module is connected to the multimodal data processing module and the process parameter synchronization and correlation analysis module. It is used to identify the type and trace the cause of defect candidate areas based on the correlation mapping results, and output a diagnostic report containing defect type, location, size and most likely cause.

[0012] The closed-loop feedback control module is connected to the intelligent classification and traceability decision module and the production line control system. It is used to generate adaptive control instructions based on the diagnostic report to adjust the transfer process parameters or equipment operating status in real time or in batches.

[0013] The end-to-end image acquisition module includes at least two sets of high-resolution line scan camera groups. Each camera group consists of one visible light camera and one near-infrared camera rigidly coaxially mounted. The visible light camera operates in the 400nm to 700nm wavelength range, and the near-infrared camera operates in the 850nm to 950nm wavelength range. The camera groups are arranged at fixed intervals along the production line direction to ensure seamless splicing imaging of the moving base fabric. The imaging resolution is no less than 0.1mm / pixel, and the line scanning frequency is synchronized with the production line speed with an error of less than ±0.5%.

[0014] The morphological adaptive segmentation unit in the multimodal data processing module operates as follows: First, it loads the irregular carpet contour vector data corresponding to the current production batch; second, based on the real-time position information fed back by the production line encoder, it maps the contour vector data to the physical coordinate space corresponding to each frame of the acquired image; finally, based on the mapped contour boundary, it generates a dynamic mask to strictly limit the image analysis area to the inside of the effective carpet contour, and performs Gaussian smoothing on the transition area 5 pixels wide outside the contour edge to eliminate false detections caused by edge burrs or slight alignment deviations.

[0015] Furthermore, the feature extraction operations performed by the multimodal data processing module specifically include improved local binary pattern texture analysis for the visible light image channel and scattering feature analysis based on the gray-level co-occurrence matrix for the near-infrared image channel. The improved local binary pattern expands the traditional circular neighborhood into an elliptical neighborhood adapted to the textile texture direction, with its major axis aligned with the warp and weft directions of the base fabric. The parameters of the elliptical neighborhood are defined by the following formula:

[0016]

[0017] in, The grayscale value of the center pixel. For the elliptical trajectory, the first The grayscale value of each sampling point, sampling point Major axis radius of the ellipse Pixels, minor axis radius Pixel.

[0018] Preferably, the process parameter synchronization and correlation analysis module incorporates a time-series alignment engine. This engine uses the pulse signal of the production line's main encoder as a unified timestamp to resample and align asynchronous signals from different data sources. The process parameters include the surface temperature of the heating roller assembly. Its sampling frequency is 10Hz, and its accuracy is ±0.5°C; linear pressure of the pressure roller. The sampling frequency is 50Hz, and the accuracy is ±0.1kN / m; the base fabric traveling linear velocity... With dynamic tension The sampling frequency is 100Hz. This module timestamps the image frames for each defect candidate region. Match the time sequence with process parameters to extract the time window. All process parameter data within, including Based on the dynamic calculation of linear velocity, it represents the time required for the defect location to pass through the hot pressing zone.

[0019] The intelligent classification and traceability decision-making module employs a two-stream deep neural network model based on an attention mechanism. The first stream of the model is the image feature stream, with input being multispectral image patches of defect candidate regions; the second stream is the process parameter stream, with input being aligned multidimensional process parameter sequence fragments. The model calculates the correlation weights between image features and process parameter features through a cross-modal attention layer, and its core calculation process is represented by the following formula:

[0020]

[0021] in, The query matrix for image feature vectors. and The key matrix and value matrix are the eigenvectors of the process parameters. is the dimension of the key vector. The final output layer of the model simultaneously provides the probability distribution of defect types and the probability distribution of cause classification. Defect types include color difference, missing image, ghosting, and wrinkles, while cause classification includes temperature anomaly, pressure anomaly, tension anomaly, and contaminant.

[0022] Furthermore, the closed-loop feedback control module executes a preset control strategy based on the cause classification in the diagnostic report. If the cause is determined to be a systemic equipment malfunction, the module generates a real-time control command. For example, when a regional color difference caused by temperature sensor drift is detected, the command will fine-tune the set temperature of the corresponding heating zone, with a compensation value of [value missing]. ,in This is the proportionality coefficient. To set the temperature, This represents the average measured temperature within the abnormal time window. If the cause is determined to be accidental contamination or a serious defect that cannot be repaired online, the module generates a marking instruction, sending the defect location coordinates and type information to the downstream cutting system to guide it in automatically removing the defective area during the cutting process.

[0023] A method for automatically identifying defects in machine-woven carpet production, the specific steps of which are as follows:

[0024] Step S110: Using a multispectral image acquisition device deployed on the production line after transfer and before cutting, the continuously moving surface of the transferred base fabric is synchronously scanned to obtain the raw image data stream containing visible light and near-infrared bands.

[0025] Step S120: Real-time preprocessing of the acquired raw image data stream, including non-uniformity correction, brightness equalization and image stitching, and calling the pre-stored irregular carpet outline data, dynamically delineating the effective analysis area of ​​each frame image through the morphological adaptive segmentation algorithm to eliminate background interference.

[0026] Step S130: Within the defined effective analysis area, texture feature extraction based on improved local binary mode and feature extraction based on near-infrared scattering are performed in parallel. Through multi-feature fusion and threshold segmentation algorithms, candidate regions of defects in the image are initially located, and their pixel coordinates and timestamps are recorded.

[0027] Step S140: Real-time synchronous acquisition of temperature, pressure, tension and linear speed parameters during the transfer process, and spatiotemporal alignment with the defect candidate area located in step S130 using a unified time reference to construct a process parameter context fragment centered on the defect.

[0028] Step S150: The multispectral image patch of the defect candidate region and its associated process parameter context fragment are input together into the trained intelligent classification and tracing model. The model fuses image and parameter features through a cross-modal attention mechanism and outputs the specific type diagnosis of the defect and the most likely cause tracing result.

[0029] Step S160: Based on the diagnosis and tracing results output in step S150, determine the nature of the defect: if it is a controllable systemic defect, generate a corresponding process parameter adjustment instruction and feed it back to the transfer equipment control system in real time; if it is a defect that cannot be repaired online, generate a defect marking instruction containing precise location information and send it to the subsequent cutting and sorting units.

[0030] Compared with the prior art, the beneficial technical effects of the present invention are as follows:

[0031] This invention enables continuous online inspection of the entire process from the transfer of woven carpets to the cutting stage. Through multispectral imaging and morphological adaptive segmentation technology, it effectively covers the contour area of ​​irregularly shaped carpets, increasing the interception rate of early repairable defects to over 95%, thus avoiding waste caused by defects flowing into subsequent processes.

[0032] This invention establishes a multidimensional spatiotemporal correlation model between defect image features and transfer process parameters. Through a cross-modal attention neural network, it achieves accurate identification of defect types and intelligent tracing of the root causes of production anomalies, reducing the average location time of systemic faults from several hours of traditional manual investigation to minutes.

[0033] This invention constructs a closed-loop feedback control link from defect identification to process control, which can adaptively fine-tune key parameters such as transfer temperature and pressure based on real-time diagnostic results, or guide the cutting system to accurately remove defective areas, significantly improving the automation level of the production line and the first-pass yield of products, and increasing the overall yield by more than 3% under high-speed continuous production conditions.

[0034] This invention breaks down data silos between traditional production subsystems through full-process data integration and intelligent analysis, forming a complete quality control system covering "perception-analysis-decision-execution", providing a reliable technical foundation for the digital and intelligent manufacturing of machine-woven carpets. Attached Figure Description

[0035] Figure 1 This is a schematic diagram of the overall technical solution architecture of the automatic defect identification system for machine-woven carpet production proposed in this invention;

[0036] Figure 2 This is a schematic diagram of the core principle framework of intelligent defect classification and source tracing decision based on cross-modal attention mechanism in this invention. Detailed Implementation

[0037] The features and exemplary embodiments of various aspects of the present invention will now be described in detail. To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are merely intended to explain the present invention and not to limit the present invention. For those skilled in the art, the present invention can be practiced without some of these specific details. The following description of the embodiments is merely to provide a better understanding of the present invention by illustrating examples of the invention.

[0038] Example 1

[0039] The system and method described in this invention are put into operation on continuous production lines for machine-woven carpets, particularly in the section between the transfer printing process and the cutting process. The base fabric in this section has already undergone pattern transfer and is in a continuous, high-speed movement, with a linear speed reaching 20 to 30 meters per minute. The carpet products flowing through the production line contain a large number of non-rectangular irregular contours, such as circles, ellipses, or customized shapes with complex curved boundaries. The goal of this embodiment is to achieve fully automated, high-precision, online identification of defects on the surface of the moving base fabric, and to complete closed-loop quality control from defect discovery to cause tracing to process adjustment or defect removal.

[0040] See Figure 1 The overall architecture of this system includes a full-process image acquisition module, a multimodal data processing module, a process parameter synchronization and correlation analysis module, an intelligent classification and traceability decision-making module, and a closed-loop feedback control module. These modules interact with the production line main control system via industrial Ethernet, forming a real-time online distributed intelligent detection and control system.

[0041] The end-to-end image acquisition module is physically deployed within a closed inspection box between the transfer equipment exit and the cutting equipment entrance. At the core of this module are two sets of high-resolution line scan cameras, arranged at a fixed interval of 1.5 meters along the production line. Each camera set consists of one visible light camera and one near-infrared camera rigidly coaxially mounted via a precision mechanical structure, ensuring perfect alignment of the images taken by both cameras at the same physical location. The visible light camera operates in the 400-700 nanometer wavelength range, capturing color, pattern, and visible texture details on the carpet surface. The near-infrared camera operates in the 850-950 nanometer wavelength range, utilizing the specific absorption and scattering characteristics of organic dyes and base fabric fibers in the near-infrared band to detect internal defects that are difficult to detect with the naked eye, such as uniformity of transfer layer thickness, tiny bubbles, or residual moisture. The imaging system's resolution is set to 0.05 millimeters per pixel to meet the requirements for identifying minute defects. The line scan frequency is triggered in real-time by pulse signals from the production line's main encoder, ensuring strict synchronization between the scan frequency and the base fabric speed, with a synchronization error controlled within ±0.3%. The raw image data acquired by the camera group is transmitted in real time to the multimodal data processing module via a gigabit Ethernet interface. The data stream format is a raw image sequence with 12-bit depth, accompanied by timestamps and location coordinates provided by the encoder.

[0042] The multimodal data processing module, deployed on an industrial server, is responsible for receiving and processing the raw image data stream from the image acquisition module. Its workflow begins with image preprocessing. The preprocessing operation first performs non-uniformity correction on the raw image to eliminate fixed-pattern noise caused by inconsistencies in the responses of individual pixels from the camera sensor. Correction coefficients are obtained by calibrating with a uniform white board and stored in a lookup table. Subsequently, brightness equalization is performed using an algorithm based on Retinex theory to compensate for local brightness variations caused by uneven lighting on the production line or minor undulations on the fabric surface, ensuring the stability of subsequent feature extraction. After correction and equalization, the system stitches together image strips acquired by adjacent camera groups in real time to generate a continuous, complete image frame covering the entire width of the fabric.

[0043] After preprocessing, the morphological adaptive segmentation unit begins operation. This unit first loads the contour vector data of the irregularly shaped carpet corresponding to the current production batch from the production management system. This data, typically exported from CAD design files, contains precise contour boundary coordinates. Secondly, based on the absolute position information of the base fabric fed back in real-time by the production line encoder, the system dynamically maps the contour vector data to the pixel coordinate space corresponding to each frame of the acquired image. This mapping process considers the camera's installation position, viewing angle, and image distortion, and is achieved through a pre-calibrated perspective transformation matrix. Finally, based on the mapped contour boundaries, the system generates a binary dynamic mask. This mask strictly confines the image analysis area to the effective carpet contour, completely masking the background area outside the contour. To handle potential burrs or slight alignment deviations caused by mechanical vibration at the contour edges, the system performs Gaussian smoothing on the five-pixel-wide transition area outside the contour boundary, with the standard deviation of the smoothing kernel set to 1.5 pixels. This operation effectively avoids false texture features caused by sharp edge transitions, thereby eliminating false detections.

[0044] Within the effective analysis region defined by the dynamic mask, the system performs multimodal feature extraction in parallel. For the visible light image channel, texture analysis based on an improved local binary pattern is performed. Traditional local binary patterns use circular neighborhood sampling, but the texture of woven carpets has a clear warp and weft directionality. Therefore, this invention improves the sampling neighborhood to an ellipse. The major axis of the ellipse is strictly aligned with the warp and weft direction of the base fabric after production line calibration. The specific parameters of the elliptical neighborhood are: sixteen sampling points, a major axis radius of three pixels, and a minor axis radius of one pixel. For each pixel in the image, its gray value is used as the center value. The difference between the gray values ​​of the sixteen sampling points on the elliptical trajectory and the center value is calculated, and then binarized according to a threshold function to finally form a sixteen-bit binary pattern code. After rotation invariance and uniform pattern dimensionality reduction processing, this pattern code generates a fifty-nine-dimensional texture feature histogram to characterize the micro-texture structure of the region. For the near-infrared image channel, scattering feature analysis based on the gray-level co-occurrence matrix is ​​performed. The system calculates gray-level co-occurrence matrices in four directions (0 degrees, 45 degrees, 90 degrees, and 135 degrees) at a distance of one pixel, and extracts four statistical quantities—contrast, correlation, energy, and homogeneity—from these matrices to form a sixteen-dimensional scattering feature vector, which is used to characterize the uniformity of the material's internal structure.

[0045] After feature extraction, the system locates candidate defect regions using a multi-feature fusion and threshold segmentation algorithm. First, the visible light texture feature vector and the near-infrared scattering feature vector are concatenated at the feature layer to form a 75-dimensional fused feature vector. Then, an anomaly detection algorithm based on Mahalanobis distance is used to calculate the deviation between the fused feature vector of each sliding window (e.g., 32 pixels by 32 pixels) in the image and the "normal feature distribution" learned from a large number of normal samples. Regions with deviations exceeding a dynamic threshold are marked as candidate defect regions. The dynamic threshold is adaptively adjusted based on the average texture complexity of the current production batch's base fabric. For each located candidate region, the system precisely records the pixel coordinates of its bounding rectangle in the image, the region area, and the timestamp of the corresponding image frame. This information is encapsulated into a data structure and transmitted to subsequent modules.

[0046] The process parameter synchronization and correlation analysis module interacts with the programmable logic controller and distributed input / output stations of the production line control system in real time. This module incorporates a high-precision timing alignment engine. The engine uses the pulse signal emitted by the main encoder of the production line, which is strictly synchronized with the displacement of the base fabric, as a unified hardware timestamp source for the entire system, with a frequency of 1 kHz. Asynchronous process parameter signals from different sensors, including temperature, pressure, tension, and linear velocity, are all fed into this timing alignment engine.

[0047] Temperature parameters are derived from multiple platinum resistance temperature sensors installed on the surface of the heating roller assembly in the transfer equipment, measuring the surface temperature of the area where the heating roller contacts the base fabric. Temperature data is sampled at a frequency of 10 Hz, with a measurement accuracy of ±0.5 degrees Celsius, and is input to the system via an analog input module. Pressure parameters are derived from the pressure transmitter of the pressure roller hydraulic system, measuring the linear pressure applied to the base fabric, sampled at a frequency of 50 Hz, with an accuracy of ±0.1 kN / m. Linear velocity and dynamic tension parameters are derived from the encoder signal of the drive motor and the feedback signal of the tension sensor, respectively, both sampled at a frequency of 100 Hz. The timing alignment engine resamples all input signals, unifying them to an equally spaced time sequence based on encoder pulses, eliminating timing misalignments caused by communication delays or asynchronous sampling periods.

[0048] After time-series alignment, this module performs spatiotemporal association between defects and process parameters. For each defect candidate region reported by the multimodal data processing module, its timestamp is extracted. Based on the linear velocity of the base fabric at that timestamp, the system reverse-calculates the time required for the physical location of the defect to pass through the core action section of the transfer hot-pressing zone, defining this time as half the width of the association time window. Specifically, the time window is set to the half-width duration before and after the defect generation timestamp. The system extracts all temperature, pressure, tension, and linear velocity data within this time window from the aligned process parameter time-series database, forming a multidimensional process parameter context fragment centered on the defect. This fragment not only contains the instantaneous values ​​of the parameters but also their statistical characteristics within the window, such as mean, variance, and gradient change trends. Finally, each defect candidate region and its associated process parameter context fragment are bound together, forming a data object containing both image and parameter evidence, ready to be sent to the decision module.

[0049] The intelligent classification and source tracing decision module adopts a two-stream deep neural network model based on an attention mechanism. Its core principle framework can be found in [link to relevant documentation]. Figure 2 The model uses historical production data during the training phase, including labeled defect images and their corresponding verified records of process parameter anomalies. The first stream of the model is the image feature stream, which takes as input multispectral image patches of defect candidate regions, uniformly scaled to 224 pixels by 224 pixels, containing both visible and near-infrared channels. This stream uses a pre-trained convolutional neural network backbone to extract deep visual features. The second stream is the process parameter stream, which takes as input as correlated multidimensional process parameter sequence fragments, first extracting local temporal pattern features through a one-dimensional convolutional layer.

[0050] The core of the model is a cross-modal attention layer. This layer uses the image feature vector extracted by the previous generation as the query and the process parameter feature vector extracted by the next generation as the key and value. The specific calculation process is as follows: First, the image feature vector is linearly transformed into a query matrix, and the process parameter feature vector is linearly transformed into a key matrix and a value matrix, respectively. Then, the dot product of the query matrix and the key matrix is ​​calculated and scaled by dividing by the square root of the key vector dimension to prevent gradient vanishing. Next, a function is applied to the scaled dot product result to obtain the attention weight distribution of the image features to each process parameter feature time step. Finally, these attention weights are used to perform a weighted summation of the value matrix to generate a context vector that incorporates relevant process parameter information. This process can be formally represented as: the attention output is equal to the product of the query matrix and the transpose of the key matrix, divided by the square root of the key vector dimension, and then multiplied by the value matrix. Through this mechanism, the model can dynamically focus on the process parameter anomaly segments most relevant to the current visual defect. For example, when the image displays color difference, the model will assign higher attention weights to the abnormal temperature time window.

[0051] The fused feature vectors are fed into subsequent fully connected layers for classification. The model output layer uses two parallel classification heads. The first classification head outputs the probability distribution of defect types, including categories such as color difference, missing image, ghosting, wrinkles, stains, and bubbles. The second classification head outputs the probability distribution of defect causes, including categories such as abnormal temperature, abnormal pressure, abnormal tension, dye contamination, foreign matter in the base fabric, and mechanical vibration of equipment. The model ultimately outputs a structured diagnostic report, including the most likely defect type, the most likely cause, a confidence score, and precise pixel location and physical size estimation of the defect.

[0052] The closed-loop feedback control module receives the diagnostic report output by the intelligent classification and traceability decision-making module and executes corresponding operations based on the preset control strategy library. The control strategies are divided into two categories: real-time process adjustment and defect labeling and removal.

[0053] If the diagnostic report determines that the defect is caused by a systemic, controllable equipment parameter anomaly, the module generates a real-time control command. For example, when a regional, periodic color difference is detected due to a slight drift in the temperature sensor of a certain heating zone, the control logic is triggered. The system calculates the deviation between the average measured temperature and the set temperature within the abnormal time window. Subsequently, a temperature compensation command is generated, with the compensation value being a proportional coefficient multiplied by the temperature deviation. The proportional coefficient is adaptively learned based on historical control effects, with an initial value set to 0.5. This command is sent to the temperature controller of the transfer equipment via the industrial network to fine-tune the set temperature of the corresponding heating zone, achieving feedforward compensation and preventing similar defects from occurring in subsequent products.

[0054] If the diagnostic report determines that the defect is caused by accidental contamination, irreversible material defects, or severe mechanical damage that cannot be repaired online, the module generates a defect marking instruction. This instruction includes the precise physical coordinates of the defect, the defect type code, and a suggested handling method. This instruction is sent in real-time to the downstream CNC cutting system. Upon receiving the instruction, the cutting system's path planning software automatically avoids the defective area during cutting and layout, or adds special markings around the defective area to guide the robotic arm or manual labor to remove the defective sheet in subsequent sorting processes. For continuous defects, the system can generate a stop-and-inspection request to notify maintenance personnel to intervene.

[0055] The entire system operates in parallel, pipeline-like fashion. Image acquisition and preprocessing, feature extraction and candidate region localization, process parameter synchronization and correlation, intelligent classification and decision-making, and closed-loop feedback control are connected via a high-speed data bus, forming a processing pipeline with millisecond-level response. The system's human-machine interface displays the production line's comprehensive quality indicators, defect distribution heatmaps, process parameter trends, and alarm information in real time, providing production managers with a global situational awareness.

[0056] Example 2

[0057] Based on the stable operation of the system described in Example 1, this example optimizes the application of the system for ultra-high-end custom woven carpet production lines with small batches, multiple varieties, and extremely complex patterns. These production lines are characterized by frequent product changes, more diverse irregular contours, and patterns containing a large number of intricate gradient colors and tiny floral designs, placing higher demands on the sensitivity and specificity of defect detection.

[0058] To address this, a third high-resolution area array spectral camera was added to the end-to-end image acquisition module. Deployed two meters downstream of the line array camera group, this camera operates within the spectral range of 400 to 1000 nanometers and possesses imaging capabilities for sixteen discrete spectral channels. Its operating mode is as follows: when the line array camera group and subsequent algorithms detect a suspicious area but the confidence level is below a threshold, or when the system identifies a special sensitive pattern in the current product based on production order information, the area array spectral camera is triggered to perform fixed-point, high-resolution spectral imaging of the specific area. The imaging data provides a continuous spectral reflectance curve for each pixel, used to identify metamerism caused by minute batch differences in dyes that is difficult to distinguish under wide-band imaging, or to accurately analyze the microscopic inhomogeneities of the chemical composition of the transfer layer.

[0059] In the multimodal data processing module, the functionality of the morphological adaptive segmentation unit has been enhanced. In addition to loading contour vector data, it also loads the product's digital pattern design file. The unit semantically segments different color block regions and texture regions in the pattern design and sets differentiated feature extraction parameters and detection sensitivity thresholds for each region. For example, for large areas of solid color, a stricter texture uniformity standard is adopted; for complex gradient regions, the focus is on detecting color continuity. This enables the detection algorithm to "understand" the image content, significantly reducing the probability of misjudging normal design patterns as defects.

[0060] The process parameter synchronization and correlation analysis module expands the data sources. In addition to basic transfer process parameters, it integrates environmental temperature and humidity sensor data, flow and pressure data from the dye supply system, and moisture content detection data of the base fabric during the pretreatment process. The time-series alignment engine needs to handle more heterogeneous signals, and the correlation model has been upgraded from two-dimensional spatiotemporal correlation to multi-dimensional causal correlation analysis. The system uses statistical methods such as Granger causality tests to analyze the lead-lag relationship between various parameter anomalies and the final defect occurrence on a longer time scale, thereby discovering indirect, delayed root causes of failures.

[0061] The intelligent classification and traceability decision-making module employs incremental learning and personalized adaptation of its model. Upon completion of each new product batch, the system uses all accumulated data from that batch, including inspection results and final quality inspector verification, as a learning sample package. The model utilizes these new samples for fine-tuning, enabling it to quickly adapt to the texture, color, and typical defect patterns of the new product. Simultaneously, the model maintains a lightweight feature subnetwork for each long-term partner or product series, facilitating rapid switching and personalized optimization of the inspection model.

[0062] The closed-loop feedback control module incorporates an adaptive optimization mechanism based on reinforcement learning into its strategy library. The module not only executes preset control commands but also feeds back the effects of each control action to an agent. By analyzing the sequence of "control action - process parameter change - subsequent defect rate change," the agent learns which control strategy will yield better overall quality and cost-effectiveness in the long run under complex and dynamic production environments. For example, it learns whether to immediately make a small adjustment when there are slight temperature fluctuations, or to tolerate brief fluctuations to avoid oscillations caused by frequent actions of the control system.

[0063] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Therefore, all equivalent changes made in accordance with the structure, shape, and principle of the present invention should be covered within the scope of protection of the present invention.

Claims

1. An automatic defect identification system for machine-woven carpet production, characterized in that, The system includes the following components: The full-process image acquisition module is deployed on the continuous production line after transfer and before cutting. It is used to acquire multispectral images of the base fabric surface after transfer and generate raw image sequences containing visible light and near-infrared spectral information. The multimodal data processing module is used to receive the original image sequence and perform image preprocessing, feature extraction and defect candidate region localization. It integrates a morphological adaptive segmentation unit to dynamically adjust the image analysis boundary according to the preset irregular carpet outline data. The process parameter synchronization and correlation analysis module is used to synchronously acquire the temperature, pressure, tension and linear speed process parameters during the transfer process, and establish a correlation mapping between the time sequence of process parameters and the spatiotemporal information of defect candidate areas. The intelligent classification and traceability decision module is used to identify the type and trace the cause of defect candidate areas based on the association mapping results, and output a diagnostic report. The closed-loop feedback control module is used to generate adaptive adjustment instructions based on the diagnostic report to adjust the transfer process parameters or equipment operating status in real time or in batches.

2. The automatic defect identification system for machine-woven carpet production according to claim 1, characterized in that, The full-process image acquisition module includes at least two sets of high-resolution line scan camera groups. Each camera group consists of one visible light camera and one near-infrared camera rigidly coaxially mounted. The visible light camera operates in the wavelength range of 400nm to 700nm, and the near-infrared camera operates in the wavelength range of 850nm to 950nm. The camera groups are arranged at fixed intervals along the production line direction, with an imaging resolution of not less than 0.1mm / pixel, and the line scanning frequency is synchronized with the production line speed.

3. The automatic defect identification system for machine-woven carpet production according to claim 1, characterized in that, The morphological adaptive segmentation unit in the multimodal data processing module operates as follows: First, it loads the irregular carpet contour vector data corresponding to the current production batch; second, based on the real-time position information fed back by the production line encoder, it maps the contour vector data to the physical coordinate space corresponding to each frame of the acquired image; finally, based on the mapped contour boundary, it generates a dynamic mask to strictly limit the image analysis area to the inside of the effective carpet contour, and performs Gaussian smoothing on the transition area 5 pixels wide outside the contour edge.

4. The automatic defect identification system for machine-woven carpet production according to claim 1, characterized in that, The feature extraction operations performed by the multimodal data processing module specifically include improved local binary pattern texture analysis for the visible light image channel and scattering feature analysis based on gray-level co-occurrence matrix for the near-infrared image channel. The improved local binary model expands the traditional circular neighborhood into an elliptical neighborhood that adapts to the textile texture direction, with its major axis aligned with the warp and weft directions of the base fabric. The parameters of the elliptical neighborhood are given by the formula... Define, where The grayscale value of the center pixel. For the elliptical trajectory, the first The grayscale value of each sampling point, sampling point Major axis radius of the ellipse Pixels, minor axis radius Pixel.

5. The automatic defect identification system for machine-woven carpet production according to claim 1, characterized in that, The process parameter synchronization and correlation analysis module has a built-in time alignment engine. This engine uses the pulse signal of the main encoder of the production line as a unified timestamp to resample and align asynchronous signals from different data sources. The process parameters include the surface temperature of the heating roller assembly. Its sampling frequency is 10Hz and its accuracy is Linear pressure of the pressure roller The sampling frequency is 50Hz, and the accuracy is ±0.1kN / m; the base fabric traveling linear velocity... With dynamic tension The sampling frequency is 100Hz; this module timestamps the image frames of each defect candidate region. Match the time sequence with process parameters to extract the time window. All process parameter data within, including Based on the dynamic calculation of linear velocity, it represents the time required for the defect location to pass through the hot pressing zone.

6. The automatic defect identification system for machine-woven carpet production according to claim 1, characterized in that, The intelligent classification and traceability decision-making module adopts a two-stream deep neural network model based on the attention mechanism; the first stream of the model is the image feature stream, and the input is a multispectral image patch of the defect candidate region; The second stream is the process parameter stream, with the input being an aligned multidimensional process parameter sequence fragment. The model calculates the correlation weights between image features and process parameter features through a cross-modal attention layer, and its core calculation process is defined by the formula... It means that among them The query matrix for image feature vectors. and The key matrix and value matrix are the eigenvectors of the process parameters. The dimension of the key vector; the final output layer of the model simultaneously provides the probability distribution of defect type and cause classification.

7. The automatic defect identification system for machine-woven carpet production according to claim 1, characterized in that, The closed-loop feedback control module executes a preset control strategy based on the cause classification in the diagnostic report. If the cause is determined to be a systemic equipment malfunction, the module generates a real-time control command. When a regional color difference caused by temperature sensor drift is detected, the command will fine-tune the set temperature of the corresponding heating zone, with a compensation value of [value missing]. ,in This is the proportionality coefficient. To set the temperature, The average measured temperature within the abnormal time window; if the cause is determined to be accidental contamination or a serious defect that cannot be repaired online, the module generates a marking instruction and sends the defect location coordinates and type information to the downstream cutting system.

8. A method for automatically identifying defects in machine-woven carpet production, applied to the automatic defect identification system for machine-woven carpet production as described in any one of claims 1-7, characterized in that, The method includes the following steps: Step S110: Using a multispectral image acquisition device deployed on the production line after transfer and before cutting, the continuously moving surface of the transferred base fabric is synchronously scanned to obtain the raw image data stream containing visible light and near-infrared bands. Step S120: Real-time preprocessing of the acquired raw image data stream, including non-uniformity correction, brightness equalization and image stitching, and calling the pre-stored irregular carpet outline data, dynamically delineating the effective analysis area of ​​each frame image through the morphological adaptive segmentation algorithm to eliminate background interference. Step S130: Within the defined effective analysis area, texture feature extraction based on improved local binary mode and feature extraction based on near-infrared scattering are performed in parallel. Through multi-feature fusion and threshold segmentation algorithms, candidate regions of defects in the image are initially located, and their pixel coordinates and timestamps are recorded. Step S140: Real-time synchronous acquisition of temperature, pressure, tension and linear speed parameters during the transfer process, and spatiotemporal alignment with the defect candidate area located in step S130 using a unified time reference to construct a process parameter context fragment centered on the defect. Step S150: The multispectral image patch of the defect candidate region and its associated process parameter context fragment are input together into the trained intelligent classification and tracing model. The model fuses image and parameter features through a cross-modal attention mechanism and outputs the specific type diagnosis of the defect and the most likely cause tracing result. Step S160: Based on the diagnosis and tracing results output in step S150, determine the nature of the defect: if it is a controllable systemic defect, generate a corresponding process parameter adjustment instruction and feed it back to the transfer equipment control system in real time; if it is a defect that cannot be repaired online, generate a defect marking instruction containing precise location information and send it to the subsequent cutting and sorting units.

9. The method for automatically identifying defects in machine-woven carpet production according to claim 8, characterized in that, In step S120, the morphological adaptive segmentation algorithm specifically includes: loading the irregular carpet contour vector data; mapping the contour vector data to the pixel coordinate space corresponding to each frame of the acquired image according to the real-time position information fed back by the production line encoder; generating a dynamic mask based on the mapped contour boundary, strictly limiting the image analysis area to the inside of the effective carpet contour, and performing Gaussian smoothing on the transition area 5 pixels wide outside the contour edge.

10. The method for automatically identifying defects in machine-woven carpet production according to claim 8, characterized in that, In step S150, the intelligent classification and tracing model is a two-stream deep neural network model based on an attention mechanism; the model calculates the correlation weights between image features and process parameter features through a cross-modal attention layer, and its core calculation process is defined by the formula... It means that among them The query matrix for image feature vectors. and The key matrix and value matrix are the eigenvectors of the process parameters. The dimension of the key vector; the final output layer of the model simultaneously provides the probability distribution of defect type and cause classification.

Images