Loom broken warp multi-light path time polling intelligent detection method and system and storage medium

Abstract

The application discloses a loom broken end multi-light-path time-sharing polling intelligent detection method and system and a storage medium, and belongs to the technical field of textile machine detection. The application aims at the broken end detection pain point of flexible yarns such as cotton yarn, chemical fiber filament, silk, lyocell, modal and glass fiber electronic cloth, adopts the upper-middle-lower three-layer horizontal opposite laser light path time-sharing polling to collect continuous light intensity signals based on the broken end winding mechanism verified by field experiments, extracts waveform feature vectors such as peak value, pulse width, rising edge and signal energy, realizes broken end recognition and hierarchical positioning through a machine learning classifier, and adopts the shielding and sweeping dual principle to continuously shield to determine and detect the broken end winding of the flexible yarn, and to instantaneously sweep a pulse to determine and detect the abnormal signal of the glass fiber broken head sweeping from the opening area. The application solves the defects of the prior art, such as dependence on random floating of broken yarn, no periodic synchronization, loss of information in 0 / 1 binary, and inability to cover glass fiber detection.

Description

Technical Field

[0001] This invention relates to the field of intelligent detection technology for textile machinery, and in particular to a multi-optical-path time-division polling intelligent detection method, system and storage medium for warp breakage on looms. It is applicable to real-time warp breakage detection during the weaving process of flexible textile yarns such as cotton yarn, chemical fiber filament, silk, elastic yarn, lyocell (commonly known as Tencel), modal and rigid yarns such as glass fiber electronic cloth on air-jet looms, water-jet looms and rapier looms. Background Technology

[0002] In the weaving process of shuttleless looms, the warp yarns are continuously subjected to high-frequency friction and alternating tension caused by shedding and weft insertion, making warp breakage the most common production defect. If warp breakage is not detected in time, it will cause fabric defects such as missing warp, skipped patterns, and holes, leading to product downgrading or even scrapping. The above problems are particularly prominent in the weaving of fiberglass electronic cloth: fiberglass yarns are highly rigid and do not have obvious rebound or winding displacement after breaking. The traditional detection principle based on the drifting or winding displacement of broken yarns is seriously unreliable. As the core substrate of copper-clad laminates and printed circuit boards, electronic cloth has extremely high requirements for fabric quality. Warp breakage defects caused by warp breakage will directly affect the electrical performance and reliability of downstream PCBs. Therefore, the detection of warp breakage in fiberglass electronic cloth is a technical problem that the industry urgently needs to solve.

[0003] Existing menopause detection technologies can be mainly classified into the following categories: The first type is the traditional mechanical stop-warp detection. Each warp yarn is threaded through a stop-warp clip, and when a warp breaks, the clip falls and triggers the electrode to conduct. This method has a simple structure, but the weight of the stop-warp clip increases the warp tension, increases the breakage rate, and the threading process is time-consuming and has high maintenance costs, making it unsuitable for high warp density and fine yarns.

[0004] The second type is photoelectric through-beam detection. Among them, invention patent CN114836881B proposes a solution: a photoelectric sensor is arranged in the angled region between the upper and lower yarn layers, converting the received light intensity into digital signals of 0 and 1 to form a yarn light wave sequence, and a deep learning model is used to determine whether the sequence is abnormal. This solution has the following shortcomings: First, the reliability of the detection mechanism needs to be verified. This scheme relies on the random drifting of broken yarn within the included angle region to create additional light obstruction. However, the inventors, through over 1000 artificial warp breakage experiments (detailed statistics by product type are provided in the invention content section below), have confirmed that over 99% of flexible textile yarns experience entanglement with adjacent yarns after warp breakage. After entanglement, the broken yarn is driven by the adjacent yarns to produce a regular horizontal displacement, rather than randomly drifting within the included angle region. Random drifting only occurs in a very few special cases where the broken yarn is extremely long and not entangled, and is not universally applicable. Using this type of accidental phenomenon as the main detection mechanism carries the risk of missed detection in most working conditions where the broken yarn is already entangled with adjacent yarns.

[0005] Secondly, the engineering implementation of the loom shedding mechanism is insufficient. The "angle region" mentioned in this solution refers to the dynamic three-dimensional V-shaped space formed by the upper and lower warp yarns during the shedding process, the size, position, and shape of which continuously change with the movement of the heald frame. This patent fails to discuss key engineering parameters such as the sensor's front and rear mounting positions, height mounting position, and relative distance to the heald frame, nor does it explain how to ensure that the laser beam always effectively passes through the angle region. In reality, a one-centimeter deviation in the mounting position can lead to the optical path being either constantly blocked or completely unblocked, resulting in insufficient signal stability.

[0006] Third, it lacks a loom cycle synchronization mechanism. This scheme does not use an encoder or loom synchronization signal, the sequence acquisition start point is random, the periodic occlusion features of normal weaving cannot be stably extracted, and the deep learning model has difficulty learning stable patterns.

[0007] Fourth, 0 / 1 binarization loses key information. Simply binarizing continuous light intensity into 0 and 1 loses key waveform features such as peak value, pulse width, rising edge, and signal energy. It is difficult to distinguish light intensity changes caused by different reasons, such as warp breakage, dust flying by, and adjacent yarn shaking, and the false alarm rate has room for improvement.

[0008] The applicant has previously filed a series of patent applications for warp breakage detection using beams without warp stops, including CN202610310917.4 (A device and method for detecting warp breakage in textile looms without warp stops), which proposes a basic detection principle of capturing horizontal displacement of 3 to 12 cm caused by warp breakage entanglement using a horizontally opposed optical path; CN202610655684.1 (Non-contact intelligent warp breakage detection and full-process linkage production management system for all types of shuttleless looms), which proposes system-level technologies such as 940nm and 850nm dual-wavelength differential anti-interference, LoRa Mesh networking, and end-edge-cloud collaboration; and CN202610654153 (A multi-level anti-interference and error prevention control method for warp breakage detection in looms without warp stops), which proposes a three-layer anti-interference system of hardware level, signal level, and logic level. The aforementioned applications laid the technical foundation for warp breakage detection without warp breakage plates, but still have the following shortcomings: First, all of the above solutions use a single horizontal optical path, which can only determine whether there is a warp breakage, but cannot determine whether the warp breakage occurs in the upper, middle, or lower layer of warp yarns, requiring operators to check layer by layer; Second, the above solutions use threshold judgment or switch quantity detection, which cannot distinguish light intensity changes caused by different reasons such as warp breakage entanglement, dust flying, and adjacent yarn shaking, and the false alarm rate still has room for improvement under complex working conditions; Third, the above solutions do not introduce artificial intelligence classification methods, and cannot continuously optimize detection accuracy through learning; Fourth, the above solutions are all based on the warp breakage entanglement displacement mechanism of flexible yarns, and do not cover the detection scenario where rigid yarns such as glass fibers do not have obvious entanglement after breakage, but the broken end will sweep across the opening area with the movement of the heald frame, which cannot meet the warp breakage detection needs of glass fiber electronic cloth weaving. It should be noted that this application specification independently and completely describes all the necessary technical features involved, and those skilled in the art can implement the technical solution of this application without relying on the aforementioned prior applications; furthermore, none of the aforementioned prior applications disclose the core technical solutions of this application, such as "multi-layer optical path time-division polling acquisition with only a single layer of optical path lit at the same time", "extracting waveform features from unbinarized continuous light intensity signals and identifying the cause of warp breakage through a machine learning classifier", and "instantaneous sweeping determination of rigid yarn breakage as the heald frame moves across the open area". The technical solutions claimed in the claims of this application are novel relative to the aforementioned prior applications, and the aforementioned prior applications do not constitute a conflicting application with this application.

[0009] The third category is machine vision inspection. This method uses industrial cameras to photograph the fabric surface or warp area and identifies warp breaks through image processing. This solution suffers from high cost, large data volume, stringent requirements for light sources and environment, and complex maintenance. Furthermore, the fabric surface is wide and the warp breaks are subtle, resulting in insufficient reliability in detection.

[0010] Therefore, the industry urgently needs a weaving machine warp breakage detection solution that is based on reliable physical mechanisms, has engineering feasibility, high detection accuracy, and low false alarm rate. Summary of the Invention

[0011] Purpose of the invention The purpose of this invention is to overcome the shortcomings of existing technologies, such as unreliable detection mechanisms, lack of synchronization mechanisms, serious information loss, and inability to cover the detection of rigid yarns such as glass fiber. This invention provides an intelligent detection method, system, and storage medium based on the physical mechanism of warp breakage and entanglement displacement verified by field experiments, employing multi-layer optical path time-division polling to acquire continuous light intensity waveforms, accurately identifying warp breakage events through an AI classifier, and simultaneously covering the detection of warp breakage in flexible textile yarns and glass fiber electronic cloth through the synergistic effect of occlusion and sweeping. Technical solution

[0012] To achieve the above objectives, the present invention adopts the technical solutions described in claims 1 to 13.

[0013] The core technical concept of this invention is based on the following physical findings verified by field experiments: Mechanism of warp breakage and entanglement displacement: The inventors conducted over 1000 artificial warp breakage experiments on actual weaving sites of various flexible textile yarns, including cotton yarn, chemical fiber filament, silk, elastic yarn, Lyocell (commonly known as Tencel), and Modal. These experiments covered different linear densities from 10 to 80 counts, different models of air-jet and water-jet looms, and different speeds from 600 to 1000 rpm. Statistical results show that after warp breakage in flexible textile yarns, over 99% of the broken ends will entangle with adjacent warp yarns. After entanglement, the broken end is driven by the adjacent yarn, and under the combined action of the heald frame opening movement and weft vibration, it produces a regular horizontal displacement. This horizontal displacement blocks the horizontally opposed laser beam path, producing a characteristic light intensity change waveform. This mechanism is universal and reliable, unlike the accidental phenomenon of random drifting.

[0014] Mechanism of warp breakage sweeping: For rigid yarns such as glass fiber electronic cloth, after warp breakage, the broken end does not show obvious entanglement or horizontal displacement, but after losing tension, it will float or droop, sweeping across the shed opening area under the influence of the heald frame's up-and-down movement. During normal weaving, there are no warp yarns in the opening area, and the optical path is unobstructed; after warp breakage, the free broken end periodically sweeps across the optical path with the heald frame's movement, generating a clearly characteristic instantaneous falling pulse. This mechanism makes the laser optical path scheme of this invention applicable not only to flexible yarns but also to detecting warp breaks in rigid yarns such as glass fiber.

[0015] Based on the above mechanism, the technical solution of the present invention includes six core aspects: First, a multi-layer optical path time-division polling method is used. Horizontally opposed laser optical paths are arranged at three heights: upper, middle, and lower. Only one layer is lit at a time, and the three layers are polled sequentially. This design avoids inter-layer optical crosstalk and achieves the localization of the broken wavelength level. If further suppression of ambient light interference is required, the dual-wavelength differential technology disclosed in the applicant's existing patent CN202610655684.1 can be used, but this technology is not an essential technical feature of this invention.

[0016] Second, continuous light intensity signal acquisition. Instead of using 0 / 1 binarization, continuous light intensity signals are acquired at a sampling rate of no less than 50kHz and a resolution of no less than 12 bits, thus preserving the waveform information completely.

[0017] Third, precise phase synchronization of the spindle. Phase locking is achieved using a spindle encoder, and acquisition is triggered within the stable range of the shed opening, ensuring that the signals acquired in each cycle are comparable.

[0018] Fourth, waveform feature extraction and AI classification. Multi-dimensional waveform feature vectors, such as peak value, pulse width, rising edge, and signal energy, are extracted and input into a random forest or one-dimensional convolutional neural network classifier to achieve accurate identification of menstrual interruption events.

[0019] Fifth, multi-cycle verification and incremental self-learning. Cross-cycle verification reduces false alarms, while incremental learning at the edge continuously improves classification accuracy.

[0020] Sixth, the dual principles of blocking and sweeping are used in synergy. When multiple optical paths are arranged along the warp direction, the optical path near the back beam uses the continuous blocking judgment principle to detect warp breaks and entanglements in flexible yarns, while the optical path near the heald frame uses the instantaneous sweeping judgment principle to detect abnormal signals when the broken end sweeps across the opening area. The dual principles can simultaneously cover the detection of warp breaks in both flexible yarns and rigid yarns such as glass fiber.

[0021] Seventh, time-sharing polling, continuous light intensity signals, and AI classification form an interdependent and collaborative closed loop, overcoming the technical bias in this field**. It has long been widely believed in this field that: due to the high spindle speed of looms, the single-cycle detection window being only a few milliseconds to tens of milliseconds, and the limited computing power of embedded controllers, only simple 0 / 1 binary threshold judgments with minimal computational load can be used. Acquiring and processing continuous light intensity signals would be impossible in real time due to the excessive data volume; this is the technical bias in this field. This invention takes a different approach and achieves continuous waveform-level detection with limited computing power through the specific combination of three technical features: First, multi-layer optical path time-division polling ensures that only one layer is lit at a time, eliminating crosstalk between layers from the source and orderly separating the waveforms of each layer on the time axis to form a distinguishable time-series waveform with inter-layer switching boundaries; Second, the continuous light intensity signal (rather than 0 / 1 binarization) completely preserves the details discarded by binarization, such as light intensity amplitude, inter-layer switching transients, and pulse edges; Third, the AI ​​classifier uses the multi-dimensional features of the aforementioned time-series waveform, such as switching transients, edge slopes, and inter-layer correlations, to accurately distinguish between real warp breaks and inter-layer switching transients, dust flying over, and adjacent yarn shaking. These three elements are interdependent and indispensable: without time-division polling, the superposition of multiple signal layers makes it impossible to distinguish the levels in continuous waveforms; if binarization is used, the transient and amplitude details of switching are lost, and the AI ​​classifier loses its discriminative feature basis; if thresholding is still used instead of AI classification, it is impossible to distinguish warp breaks from various types of interference in continuous waveforms, and the information obtained from continuous acquisition cannot be converted into reliable judgments. It is this collaborative closed loop that enables this invention to achieve level positioning, cause differentiation, and full coverage of flexible / rigid yarns without increasing hardware computing power requirements, achieving technical effects that existing binarization thresholding schemes cannot achieve, and is not a simple superposition of common knowledge. Beneficial effects

[0022] 1. Reliable detection mechanism: Based on the warp breakage and winding displacement mechanism verified by field experiments, the detection accuracy of warp breakage and winding conditions of flexible yarn is no less than 99.4%, and the vast majority of warp breaks can be effectively detected, unlike unreliable solutions that rely on random drift.

[0023] 2. Complete information: Continuous light intensity signal acquisition retains all waveform features. Combined with multi-dimensional feature extraction and AI classification, it can distinguish different causes such as warp breakage, dust, and adjacent yarn shaking, resulting in a low false alarm rate. The measured false alarm rate is 2.60% when the fiberglass breakage sweeps through at 600 to 900 rpm and only 1.13% when it is at 450 to 500 rpm (see Example 6 for details).

[0024] 3. Location capability: The three-layer optical path time-division polling can determine the layer where the broken warp is located. Combined with the lateral position coding, the broken warp coordinates can be located, shortening the time for operators to find the head.

[0025] 4. Project Feasibility: All project details, including optical path installation location, polling timing, synchronization method, and sampling parameters, have been clearly defined, allowing for direct implementation.

[0026] 5. High adaptability: Baseline adaptive calibration compensates for dust and aging, and edge incremental self-learning continuously optimizes the model to adapt to different yarn types and loom conditions.

[0027] 6. Dual-principle synergy covers multiple categories, especially suitable for fiberglass electronic cloth: Combining the blocking principle and the scanning principle, it can detect warp breaks and entanglements in flexible textile yarns such as cotton yarn, chemical fiber, silk, Lyocell (commonly known as Tencel), and Modal, as well as detect abnormal signals when the broken ends of rigid yarns such as fiberglass electronic cloth sweep across the open area after warp breaks. This breaks through the limitation of traditional beam detection, which is only applicable to flexible yarns, and provides a low-cost and highly reliable warp break detection solution for high-value-added fabrics such as fiberglass electronic cloth.

[0028] The technical methods of this invention and existing detection technologies are compared in the table below (qualitative comparison): As can be seen from the table above, this invention is not a simple superposition of existing technical means, but is based on the broken warp winding displacement mechanism and broken head sweeping mechanism verified by field experiments. It organically combines multi-layer optical path time-division polling, continuous waveform acquisition, spindle phase synchronization and AI classification, and produces a combined technical effect that cannot be achieved simultaneously by existing technologies: "cause differentiation, level positioning, lateral positioning, and detection of both flexible and rigid yarns". Attached Figure Description

[0029] In each of the accompanying drawings, the same reference numerals indicate the same component. In different accompanying drawings, the meaning of the same numerical reference numerals is subject to the accompanying description and they are independent of each other.

[0030] Figure 1 This is a schematic diagram of the overall system architecture and the installation positions of the three-layer optical path of the present invention, wherein: 1 is the upper laser transmitter, 2 is the middle laser transmitter, 3 is the lower laser transmitter, 4 is the upper photoelectric receiver, 5 is the middle photoelectric receiver, 6 is the lower photoelectric receiver, the dashed arrow indicates the optical path of the laser from the transmitter to the receiver, 7 is the transmission drive control unit, 8 is the reception processing control unit, and 7 and 8 are bidirectional communication connections. Figure 2 This is a schematic diagram of the physical mechanism of warp breakage winding displacement, where: 1 is the back beam, 2 is the long winding area formed by the broken end and the adjacent yarn, 3 is the end of the broken end after displacement, 4 is the heald frame, and 5 is the horizontally opposed laser detection optical path (the broken end after winding displacement blocks this optical path). Figure 3This is a schematic diagram of the laser detection principle for free yarn at a long included angle in the shed. In the diagram: 1 is the back beam, 2 is the upper warp yarn, 3 is the lower warp yarn, 4 is the upper free yarn, 5 is the lower free yarn, 6 is the upper laser transmitting and receiving unit, 7 is the lower laser transmitting and receiving unit, 8 is the upper horizontal detection optical path, 9 is the lower horizontal detection optical path, and 10 is the heald frame. Figure 3 The instantaneous scanning detection principle demonstrated can be applied to the deployment scheme of multiple optical paths distributed along the warp direction as described in claim 9, wherein each optical path adopts... Figure 3 The horizontally opposed laser detection structure is shown. Detailed Implementation

[0031] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments. Figure 1 The overall system architecture and the installation positions of the three-layer optical path of the present invention are shown. Figure 2 The physical mechanism of warp breakage and winding displacement is shown. Figure 3 The principle of laser detection for free wires at long included angles in the shed is shown.

[0032] Hardware configuration: Optical path detection: A three-layer 650nm red laser beam path (laser wavelength selectable range from 600nm to 950nm, this embodiment uses 650nm), each layer contains a 5mW laser emitter and a photodiode receiver, horizontally mounted opposite each other on both sides of the warp yarn arrangement area, with a vertical spacing of 15mm between layers; the 650nm 5mW visible light laser belongs to Class... With a 3R safety rating, direct eye contact with the laser beam must be avoided during installation; a laser protection shield is recommended. The use of 650nm visible red light allows for direct visual observation of the beam position during installation and calibration, significantly reducing the difficulty of multi-path alignment and adjustment, making it particularly suitable for dense deployments of three-layer optical paths and multiple detection units. This spacing is determined based on the warp yarn vertical oscillation envelope analysis: in high-speed weaving, the vertical amplitude of the warp yarn affected by airflow and tension fluctuations typically does not exceed 3mm. After adding a 3x safety margin, the interlayer spacing is set to 15mm, ensuring that normal warp yarn oscillation will not simultaneously block adjacent optical paths, avoiding interlayer crosstalk. When arranging multiple detection units along the loom width, a 5% to 10% overlap area is set in the optical path coverage of adjacent detection units. This overlap ratio is calculated based on the laser beam divergence angle (typically 0.1mrad) and detection distance (1.5 to 2.5m), ensuring no detection blind spots at the junction of adjacent detection units while avoiding signal redundancy caused by excessive overlap. Synchronization: 1024-line incremental spindle encoder with phase locking accuracy of ±0.5°; Acquisition: 12-bit ADC, sampling rate 50kHz; Control computation: STM32F103RET6 microcontroller (72MHz, 512KB Flash, 64KB RAM, no hardware floating-point unit), running a random forest classifier (100 decision trees, maximum depth 12, 8-dimensional feature input). Theoretical estimates indicate that a single inference requires traversing 100 trees, with an average tree depth of approximately 8 layers, and approximately 800 comparison operations. At a 72MHz clock speed, the worst-case inference time is approximately 2.5 milliseconds, and the average is approximately 1.2 milliseconds. Optimization using tree structure array storage and feature pre-sorting results in a memory footprint of approximately 8KB, meeting the 64KB RAM constraint. Communication: Data is uploaded to the host computer via RS485 Modbus RTU protocol.

[0033] Workflow: S1. The encoder calculates the spindle angle in real time and outputs a trigger signal within the stable range of the shed opening (spindle 155° to 195°). S2. The three optical paths are polled in the order of top → middle → bottom, with each layer lit for 200 microseconds and inter-layer blanking for 20 microseconds, resulting in a polling period of 660 microseconds. When the lower limit of the polling period is 0.5 milliseconds, each layer can still be lit for about 150 microseconds after deducting the blanking time of the three layers. At a sampling rate of 50kHz, each layer can obtain 7 to 8 sampling points each time it is lit, which is sufficient to extract the light intensity envelope and peak features. Therefore, the polling period range of 0.5 to 10 milliseconds specified in claim 2 can be achieved. S3 and ADC collect the light intensity signal at the receiving end at a sampling rate of 50kHz. Each layer of optical path collects 10 sampling points during each lighting period (200 microseconds). Within the stable range of the shed opening (155° to 195° of the main shaft, which corresponds to about 6.7 milliseconds at 1000rpm), the three layers of optical path are polled about 10 times, and about 100 sampling points are collected per layer per weaving cycle to form a continuous light intensity waveform. S4. Extract the 6-dimensional basic feature vector of claim 1 from the continuous waveform: signal peak value, pulse width, rise time, fall time, signal energy (waveform area), and baseline drift; and further superimpose extended features such as peak / baseline ratio and pulse interval variation coefficient as described in claim 4. In this embodiment, the random forest uses a total of 8-dimensional feature vectors, including 6-dimensional basic features and 2-dimensional extended features. S5. The feature vector is input into the random forest classifier. The classifier outputs detailed classification results and confidence scores for five categories: "normal, broken warp and entanglement, dust interference, adjacent yarn shaking, and broken end sweeping". It maps "broken warp and entanglement, broken end sweeping" to broken warp events and outputs the broken warp type (entanglement type or sweeping type). It maps "dust interference, adjacent yarn shaking" to interference and critical samples to suspected. The above five detailed classification results are output by the same multi-class classifier at one time, without the need to set up separate classifiers for the two types of broken warp: entanglement type and sweeping type. S6. When two consecutive cycles in the same detection unit interval of the same level are determined to be a broken meridian, confirm the broken meridian event, output the level and coordinates, and trigger the shutdown and audible and visual alarms. S7. Perform baseline calibration every 10 minutes, collect baseline values ​​under unobstructed conditions, and update using an exponentially weighted moving average; Environmental linkage calibration: When the temperature and humidity sensor detects an ambient temperature change exceeding 5°C or a relative humidity change exceeding 15%, trigger an additional baseline calibration to compensate for the drift of optoelectronic device parameters caused by temperature and humidity changes; Anomaly self-check: When the baseline value deviation between adjacent detection units exceeds 20%, it is determined that the detection unit may have lens contamination or hardware failure, triggering automatic recalibration and reporting the abnormal status; S8. After the waveform characteristics and manual review results of each menopause event are screened for validity, they are uploaded to the cloud server. The cloud gathers valid samples from multiple looms every month to perform incremental training. The updated random forest model is then distributed to the edge computing units of each loom. The edge unit only performs real-time inference and does not perform model retraining.

[0034] To prevent catastrophic forgetting caused by incremental learning, model updates employ an A / B dual-model and regression validation mechanism: the edge computing unit simultaneously stores the currently used model (Model A) and the newly deployed model (Model B) from the cloud. Model B is first validated offline on a local historical sample regression test set (covering all categories of cotton yarn, chemical fiber, elastic yarn, regenerated cellulose yarn, and glass fiber, while retaining a certain proportion of interfering negative samples such as dust and adjacent yarn shaking). Only when the overall classification accuracy of Model B on the regression test set is not lower than that of Model A, and the accuracy of any yarn category does not decrease, is Model B switched to the used model; otherwise, Model A is retained and the validation results are sent back to the cloud, where samples are supplemented and the model is retrained, thus ensuring that each model update only improves and does not regress.

[0035] Based on Example 1, the microcontroller is upgraded to an STM32F407RET6 (with FPU, 1MB Flash, 192KB RAM), and the classifier is replaced with a one-dimensional convolutional neural network: Network structure: Input layer (waveform sequence of length 64) → Convolutional layer 1 (16 3×1 convolutional kernels, ReLU) → Max pooling layer (2×1) → Convolutional layer 2 (32 3×1 convolutional kernels, ReLU) → Global average pooling → Fully connected layer (32 neurons, ReLU) → Output layer (Softmax, three classification); The model is quantized using INT8, and the inference time is approximately 3 milliseconds. The feature input is the normalized original waveform sequence. Without manual feature extraction, the CNN automatically learns the features. From about 100 sampling points per cycle per layer, 32 sampling points are extracted before and after the triggering of the meridian break event to form a waveform sequence of length 64 as the input to the CNN. The starting point of the extraction is determined by the triggering position of the falling edge of the light intensity. Training dataset construction: During actual loom operation, light intensity waveform samples were collected through artificial warp breakage experiments and natural warp breakage events. Each sample was labeled as one of the following categories: "Normal," "Warp Breakage Entanglement," "Dust Interference," "Adjacent Yarn Shaking," or "Broken Yarn Sweeping." The dataset was stratified by yarn type, loom model, and speed range to ensure a balanced number of samples in each category, with no fewer than 500 training samples per yarn type. The training set and test set were divided in an 8:2 ratio. For sample validity screening, the signal-to-noise ratio (SNR) was used, with only samples having an SNR of at least 10 dB retained in the training set to ensure the quality of the training samples.

[0036] Waveform feature vector definition: The waveform feature vector includes the following dimensions: (1) signal peak value, which is the difference between the maximum and minimum values ​​of the light intensity signal within the sampling window; (2) pulse width, which is the duration during which the light intensity is lower than 70% of the peak value; (3) rise time, which is the time it takes for the light intensity to rise from 10% to 90% of the peak value; (4) fall time, which is the time it takes for the light intensity to fall from 90% to 10% of the peak value; (5) signal energy, which is the integral of the light intensity signal with respect to time within the sampling window; (6) baseline drift, which is the difference in light intensity between the start and end points of the sampling window. All of the above features are calculated directly from the continuous light intensity signal without binarization processing. Furthermore, the waveform feature vector may also include the following dimensions: (7) the ratio of signal peak value to baseline value, i.e. the ratio of signal peak value to current light intensity baseline value, used to characterize the degree of occlusion; (8) the pulse interval variation coefficient, i.e. the ratio of the standard deviation to the mean of multiple consecutive pulse intervals, used to characterize the regularity of the breakage jitter after the warp is broken and wound; (9) the correlation coefficient of multi-layer optical path signals, i.e. the correlation coefficient of light intensity signals of upper, middle and lower optical paths in the same time window, specifically the Pearson correlation coefficient, or the Spearman rank correlation coefficient and other correlation measures (collectively referred to as multi-layer optical path signal correlation coefficient in the claims), used to distinguish between single warp breakage occlusion and multi-layer simultaneous interference; (10) the peak value of the signal difference between adjacent periods, i.e. the peak value of the difference in light intensity signals in the same phase of two adjacent weaving periods, used to detect the gradual occlusion that changes over time.

[0037] Training process: The random forest classifier uses the Bagging ensemble method with the Gini coefficient as the splitting criterion. During training, feature subsets and sample subsets are randomly selected, and the final classification result is output through voting among multiple decision trees. The one-dimensional convolutional neural network uses the cross-entropy loss function and the Adam optimizer, with an initial learning rate of 0.001, a batch size of 32, and at least 50 training epochs. Training is terminated early when the accuracy on the validation set no longer improves. After training, the model parameters are quantized into 8-bit fixed-point numbers and deployed to edge computing units.

[0038] Experimental Verification: The inventors conducted over 1000 artificial warp breakage experiments on actual weaving sites using various flexible yarns, including cotton yarn, synthetic filament, elastic yarn, and lyocell (commonly known as Tencel). Approximately 400 experiments were conducted with cotton yarn, 300 with synthetic filament, 150 with elastic yarn, and 150 with regenerated cellulose yarns such as lyocell and modal. The remainder were mixed yarns and repeatability verification experiments, covering different linear densities from 10 to 80 counts, different models of air-jet and water-jet looms, and different speeds from 600 to 1000 rpm. Statistical results show that over 99% of the broken ends of flexible textile yarns will entangle with adjacent warp yarns and produce a regular horizontal displacement. The detection scheme based on this mechanism can reliably detect warp breakage events under the above conditions. This invention is applicable to flexible textile yarns such as cotton yarn, polyester-cotton blended yarn, pure synthetic fiber filaments (polyester, nylon, polypropylene), silk, elastic yarn (spandex core-spun yarn), and regenerated cellulose filaments (lyocell (commonly known as Tencel), modal, viscose). These yarns share the common characteristics of having a certain degree of softness and entanglement; after warp breakage, the probability of the broken end entangled with adjacent yarns exceeds 99%.

[0039] The testing parameters for different yarn specifications are adapted as follows: For coarse yarns, the warp breakage and entanglement displacement are large, and the light intensity changes significantly. A simple threshold can be used for detection, and a random forest classifier with 50 decision trees is sufficient. For fine yarns and silk, the entanglement displacement is small, and the light intensity changes are weak. It is necessary to increase the sampling rate to above 50kHz, increase the feature dimension, and use a random forest with 200 trees or a one-dimensional CNN classifier to improve detection sensitivity.

[0040] For rigid inorganic fibers such as glass fiber, since there is no rebound or obvious entanglement after breakage, the entanglement displacement mechanism scheme of this embodiment is not directly applicable. However, claims 9 and embodiment 6 of this invention have clearly disclosed a technical solution for detecting warp breakage of rigid yarns such as glass fiber by instantaneous scanning judgment principle. Therefore, the detection of warp breakage of rigid yarns such as glass fiber is still within the overall protection scope of this invention. The three-layer optical path described in claim 1 is a preferred embodiment. In practical applications, it can be simplified to a single-layer or two-layer optical path depending on cost budget and detection accuracy requirements. It should be noted that the single-layer and two-layer schemes are simplified variations of the three-layer scheme. Although they do not possess the complete hierarchical positioning capability of the three-layer scheme, their core technical concepts of continuous light intensity signal acquisition, waveform feature extraction, and AI classification are consistent with this invention.

[0041] (1) Single-layer scheme: Only the middle layer optical path is retained, and the upper and lower layers are eliminated. This scheme has the simplest structure and is the most convenient to install and debug. It is suitable for yarn varieties such as coarse cotton yarn and chemical fiber filaments with large displacement (more than 5cm) due to warp breakage and winding. After warp breakage, the broken end has a high probability of passing through the middle layer optical path area, regardless of whether it winds upward or downward. The single-layer scheme loses the hierarchical positioning capability, but it can still distinguish warp breakage and winding from dust interference through an AI classifier. Its detection performance is better than the traditional threshold switch scheme.

[0042] (2) Two-layer scheme: retain the upper and lower layers, or retain the middle layer and one of the layers. This scheme achieves a compromise between cost and inspection coverage and is suitable for medium-count yarn varieties. The two-layer scheme can still achieve rough layer positioning (upper layer warp breakage or lower layer warp breakage), and the installation and debugging difficulty is between that of single-layer and three-layer.

[0043] The simplified scheme retains the same polling timing, sampling rate, AI classifier, and multi-cycle confirmation as Embodiment 1, with only the number of optical path layers reduced accordingly. The linkage control (S5) of the simplified scheme is implemented as follows: In the single-layer scheme, lacking hierarchical positioning capability, after confirming a broken wire, it only outputs the lateral coordinates corresponding to the detection unit that triggered the broken wire, stops the machine, and triggers an alarm, without outputting hierarchical information; the two-layer scheme outputs a rough hierarchy of "upper layer" or "lower layer" and its corresponding lateral coordinates, then stops the machine and triggers an alarm.

[0044] As a specific implementation of the multi-path optical distribution along the warp direction as described in Independent Claim 9 (this implementation adopts a front-to-back distribution deployment dimension, and does not require the three-layer structure of Claim 1, but falls within the protection scope of Independent Claim 9), multiple sets of horizontal optical paths can be arranged along the warp direction (from the back beam to the heald frame) within the maximum opening area of ​​the shed, with all optical paths at the same height (middle layer of the shed), without upper and lower layers. This implementation adopts a pulse-triggered sampling mode, that is, within each shed opening cycle, a preset number of laser pulses are emitted and the light intensity response is collected synchronously. The collected pulse response sequence can still be spliced ​​into a continuous waveform in chronological order for AI classification; the number of laser pulses is configured differently according to the opening amount and warp breakage detectability at each position. (1) Optical path near the back beam: The warp yarn at this position is just drawn out from the back beam, and the opening is small (about 1 to 2 cm). The probability of broken warp ends falling into the opening area is low. Only 1 to 2 laser pulses need to be emitted each time the shed opening cycle. (2) Optical path at the middle position: The opening at this position is moderate (about 2 to 3 cm), and 3 to 4 laser pulses are emitted each time the opening is cycled; (3) Optical path near the heald frame: The opening is the largest at this position (up to 4 to 5 cm), and the probability of broken wires falling into the opening area is the highest. 6 to 12 laser pulses are emitted each opening cycle to increase the sampling density in order to capture the short-lived broken wire occlusion signal.

[0045] The advantage of the front-to-back multi-path solution is that all optical paths are at the same height, and installation and calibration only require alignment with a single horizontal plane, significantly reducing debugging difficulty compared to the top-to-bottom three-layer solution. Its limitations lie in the loss of hierarchical positioning capability and the potential for missed detection of broken warp threads that are wound upwards and do not fall into the central opening area. Therefore, this solution is suitable for applications with high requirements for ease of installation and low requirements for hierarchical positioning. The front-to-back multi-path solution can be combined with an AI classifier to uniformly extract features from the continuous light intensity waveforms of each optical path and classify them accordingly.

[0046] As another optional deployment method, three horizontal optical paths are arranged sequentially along the warp direction from the back beam to the heald frame. The three optical paths work together using two different detection principles: (1) First path (near the back beam): Continuous occlusion is used for detection. At this location, the warp yarns have not yet formed a clear opening and are densely packed. After a warp break, the broken end becomes entangled with adjacent yarns, causing horizontal displacement and continuously blocking the light path. The detection logic is as follows: When the system starts, it collects the light intensity reference value V_ref1 under normal weaving conditions; when the ADC sampled value is continuously lower than V_ref1×60% for more than 100 milliseconds, and this condition is met for two consecutive weaving cycles, it is determined to be a warp breakage and entanglement event. This path mainly detects warp breakage and entanglement in flexible yarns such as cotton and chemical fibers.

[0047] (2) Second path (middle position of the shed): The instantaneous sweep judgment principle is adopted. At this position, the upper and lower warp yarns have separated to form an opening. During normal weaving, there are no warp yarns in the opening area, and the ADC reference value V_ref2 is the maximum light transmittance. When the glass fiber or flexible yarn breaks, the broken end loses tension and floats or droops. When it moves up and down with the heald frame, it sweeps through the opening area, generating an instantaneous light intensity drop pulse. The movement mechanism of the broken glass fiber end is as follows: After the warp breaks, the broken end on the weave side is released from tension control and droops under its own weight or floats under the action of airflow. At the same time, it is driven by the friction of the adjacent yarn movement and periodically passes through the opening area with the opening of the heald frame. According to the inventor's observation on the glass fiber loom site using high-speed camera (1000 frames / second), under the conditions of 7628 and 2116 specification glass fiber cloth and loom speed of 600 to 900 rpm, the movement amplitude of the broken end in the opening area after the warp breaks can reach 2 to 5 cm, which is enough to sweep through the horizontally arranged laser light path and generate a detectable instantaneous light intensity drop pulse. The judgment logic is as follows: when the ADC sampled value is lower than V_ref2×70%, the pulse duration is between 2 ms and 50 ms, and the slope of the pulse falling edge exceeds the threshold predetermined by the calibration experiment (corresponding to the characteristic of a single warp yarn sweeping quickly), it is judged as a warp breakage and head breakage sweeping event. This path can detect yarns such as glass fibers that have broken without obvious entanglement, but whose broken ends sweep across the open area with the movement of the heald frame.

[0048] (3) Third path (closest to the frame, with the largest opening): The instantaneous sweep judgment principle is also adopted, and the judgment logic is the same as that of the second path. The reference value is V_ref3. Since the opening is the largest at this position, the path swept by the broken end is the longest and the signal is the most obvious. It can be used as confirmation and redundancy for the second path.

[0049] The three-channel collaborative judgment logic works as follows: when the first channel detects continuous obstruction and the second or third channel simultaneously detects a sweeping pulse, a warp break is confirmed with high confidence; when the second and third channels detect the same warp break sweeping pulse within adjacent time windows, a warp break is confirmed (applicable to glass fiber non-tangle scenarios); when only the second or only the third channel detects a sweeping pulse, it is marked as a suspected warp break event; a warp break is confirmed when the sweeping pulse recurs at the same position for two consecutive weaving cycles. This three-channel solution, combining the obstruction and sweeping principles, can simultaneously cover warp break detection for both flexible yarns and rigid yarns such as glass fiber.

[0050] ADC sampling parameters: All three optical paths use 12-bit ADCs with a sampling rate of no less than 50kHz. The reference value is automatically calibrated every 10 minutes, and data from periods of abnormal meridian interruption are excluded during calibration.

[0051] Statistical analysis of fiberglass breakage detection: The inventors conducted hundreds of manual warp breakage tests on 7628 and 2116 specification fiberglass electronic cloth weaving machines, focusing on the condition where the broken warp yarns did not entangle after breaking but only oscillated up and down with the heald frame. These tests were performed separately for high and low speed conditions, combined with continuous observation during multiple shifts of normal weaving. Using all test samples as a unified statistical standard, the sum of the detection accuracy, false alarm rate (the proportion of normal weaving mistakenly judged as warp breakage), and missed detection rate (the proportion of actual warp breakages not detected) was 100%. The actual test results under different speed conditions are as follows: (a) High speed operating conditions (loom spindle 600 to 900 rpm): detection accuracy rate is 95.07%, false alarm rate is 2.60%, and missed detection rate is 2.33%; (ii) Lower speed operating conditions (loom spindle 450 to 500 rpm): detection accuracy rate is 97.57%, false alarm rate is 1.13%, and missed detection rate is 1.30%.

[0052] At lower speeds, the broken yarn moves slower along the optical path with the heald frame, the duration of a single blockage is longer, and the waveform characteristics are more fully developed, resulting in higher detection accuracy and lower false alarm and false negative rates. Both sets of data jointly verify the principle that detection performance improves as the speed decreases, and also indicate that this solution has stable detection capabilities within the commonly used 450-900 rpm speed range for fiberglass electronic cloth. For the warp breakage and entanglement of flexible yarns such as cotton and chemical fibers, the broken yarn produces a regular and continuous displacement after entanglement with adjacent yarns, resulting in significant and stable signal characteristics. Its detection accuracy is no less than 99.4%, significantly higher than that of fiberglass. The above measured data show that the scanning determination principle of this invention has a stable and reproducible detection capability for rigid yarns such as fiberglass that do not entangle after breakage.

[0053] Those skilled in the art should understand that the present invention is not limited to the specific embodiments described above. The number of optical path layers can be adjusted to 2 to 5 layers as needed, or multiple optical paths can be arranged along the warp direction to adopt a dual principle of blocking and scanning for collaborative detection; the polling sequence, sampling rate, classifier parameters, and judgment threshold can be adjusted according to the loom model and yarn type, and these modifications all fall within the protection scope of the claims of the present invention.

Claims

1. A multi-optical-path time-division polling intelligent detection method for warp breakage in looms, characterized in that, Includes the following steps: S1. Multi-layer optical path time-division polling acquisition: Three layers of laser beam-beaming optical paths (upper, middle, and lower) are arranged horizontally in the warp yarn arrangement area of ​​the loom. The three optical paths are distributed at intervals along the vertical direction of the warp yarns. Only one layer of optical path is lit at any given time. The three optical paths are polled sequentially according to a preset time sequence. At least two detection units are distributed along the width of the loom, and each detection unit corresponds to a horizontal interval. The continuous light intensity signal of the receiving end of each optical path is acquired. The continuous light intensity signal refers to the receiving end signal that retains the continuous change information of the light intensity amplitude without binarization or thresholding. After analog-to-digital conversion, it is the original light intensity sampling sequence that retains the amplitude information. This is different from the 0 / 1 switch signal that only outputs high and low levels. S2. Warp Break Feature Extraction: The continuous light intensity signal is preprocessed to extract waveform feature vectors including signal peak value, pulse width, rise time, fall time, signal energy, and baseline drift. The warp break features include warp break entanglement displacement and warp break sweeping characteristics. The waveform feature vectors are used to identify both entanglement and sweeping warp breaks. The entanglement displacement refers to the physical phenomenon where the broken end of a flexible textile yarn entangles with adjacent yarns after warp breakage, causing horizontal displacement of the broken end and thus blocking the horizontally opposed light path, resulting in continuous light intensity changes. The sweeping of the broken end refers to the physical phenomenon where the broken end floats or droops after warp breakage and sweeps across the open area with the heald frame, generating a momentary light intensity drop pulse. This is applicable to rigid yarns that do not exhibit obvious entanglement behavior after breakage. Rigid yarns refer to yarns where the broken end does not form a stable entanglement with adjacent yarns after breakage but moves with the heald frame in a free state, including glass fiber yarns. For the specific material range of the flexible and rigid yarns, please refer to the instruction manual. S3. AI Classification and Recognition: Input the waveform feature vector into a trained machine learning classifier based on the waveform feature vector, and output the warp breakage event judgment result, the level of the warp breakage, and the type of warp breakage. The types of warp breakage include warp breakage entanglement type and warp breakage sweeping type; S4, Multi-cycle Confirmation and Positioning: When the warp breakage feature recurs in the same level and the same detection unit interval within two or more consecutive weaving cycles, the warp breakage event is confirmed and the warp breakage level and the lateral coordinates corresponding to the detection unit are output; S5, Linkage Control: After confirming the warp breakage, a stop signal is sent to the loom control system, and the warp breakage position is indicated by the alarm indicator module.

2. The method according to claim 1, characterized in that, The three-layer optical path in step S1 adopts a laser-guided optical path, and the laser wavelength is selected from the visible light band or the near-infrared band. Each layer of the optical path includes a transmitter and a receiver, which are horizontally opposite each other on both sides of the warp yarn arrangement area. In the at least two detection units distributed along the width direction, each detection unit includes a set of independent three-layer optical path transmitter and receiver components. The detection units are arranged at equal intervals along the width direction of the loom, and the optical path coverage of adjacent detection units has a 5% to 10% overlap area to eliminate detection blind spots. The polling period of the three-layer optical path is 0.5 to 10 milliseconds, the illumination duration of each layer of the optical path is 100 to 500 microseconds, and the inter-layer switching blanking interval is 10 to 50 microseconds.

3. The method according to claim 1, characterized in that, Step S1 further includes detecting and acquiring phase lock: acquiring the encoder signal of the loom spindle, triggering the acquisition of light intensity signals of each layer of optical path in the stable phase interval of the shed opening, with a phase lock accuracy better than ±1°; the sampling rate of the continuous light intensity signal is not less than 50kHz, and the sampling resolution is not less than 12 bits.

4. The method according to claim 1, characterized in that, The waveform feature vector in step S2 also includes: the ratio of signal peak value to baseline, pulse area, pulse interval variation coefficient, multi-layer optical path signal correlation coefficient, and the differential peak value of the difference in light intensity signal of the same phase between two adjacent weaving cycles; the preprocessing includes baseline drift removal, digital filtering, and outlier removal.

5. The method according to claim 1, characterized in that, The machine learning classifier based on waveform feature vectors is a random forest classifier or a one-dimensional convolutional neural network classifier; the random forest classifier outputs the classification result by ensemble voting of multiple decision trees, and the one-dimensional convolutional neural network takes a fixed-length normalized original waveform sampling sequence as input. The original waveform sampling sequence is obtained by intercepting continuous light intensity signals and is different from the waveform feature vector described in claim 1. After extracting local features through convolutional layers, the classification result is output through fully connected layers.

6. The method according to claim 1, characterized in that, The multi-cycle confirmation in step S4 includes: when the classifier determines that the warp breakage has occurred for 2 to 3 consecutive weaving cycles in the same detection unit interval at the same level, the warp breakage event is confirmed; a single determination is marked as a suspected event, and only recorded without alarm.

7. The method according to claim 1, characterized in that, It also includes dynamic compensation of light intensity baseline: Under the full life cycle adaptive calibration framework, the light intensity baseline under the state of no light obstruction is collected every 5 to 30 minutes, and the baseline parameters are updated by an exponential weighted moving average algorithm to compensate for baseline drift caused by dust attenuation, lens contamination and light source aging; the full life cycle adaptive calibration framework includes four stages: initial benchmark calibration, dynamic light attenuation compensation, environmental linkage calibration and anomaly self-check.

8. The method according to claim 1, characterized in that, It also includes edge acquisition and cloud-based collaborative incremental self-learning: the edge computing unit uses the waveform features and manual review results corresponding to each rupture determination as training samples, and uploads them to the cloud server after validity screening; the validity screening means that only samples that are manually confirmed as real rupture events and whose signal quality meets the preset signal-to-noise ratio requirements are included in the training set, and the signal-to-noise ratio is the ratio of the signal peak root mean square to the noise root mean square, and its determination threshold is specified in the specification; The cloud server aggregates valid samples from multiple loom nodes, performs incremental training on the classifier, and then distributes the updated model to the edge computing units of each loom. The edge computing units are responsible for real-time inference and do not perform complete model retraining.

9. A multi-optical path detection method for warp breakage on a loom, characterized in that, include: At least two horizontally opposed laser beams are distributed along the warp direction from the back beam to the heald frame in the shed area of ​​the loom. Each beam is time-divisionally polled and illuminated to collect the unbinded continuous light intensity signal from the receiving end. Waveform features are extracted from the continuous light intensity signal and identified as warp breakage using a machine learning classifier. The first beam closer to the back beam uses a continuous occlusion determination logic, while the subsequent beams closer to the heald frame use an instantaneous sweep determination logic. The continuous occlusion determination is: when the light intensity at the receiving end is continuously lower than 50% to 70% of the reference value for a duration exceeding 50 to 200 milliseconds, it is determined to be a warp breakage and entanglement of the flexible yarn. The instantaneous sweep determination is: when the light intensity at the receiving end is lower than 200 milliseconds for a duration exceeding 200 milliseconds, it is determined to be a warp breakage and entanglement of the flexible yarn. A momentary falling pulse between milliseconds and 50 milliseconds, with a falling edge slope exceeding a threshold predetermined by calibration experiments for the corresponding yarn variety, is determined to be a warp break or broken end sweeping through the open area. The reference value is the light intensity reference of each optical path collected during normal weaving at system startup. The momentary sweep determination logic is applicable to yarn varieties such as glass fiber that, after breakage, do not show obvious entanglement but the broken end floats or droops and sweeps through the open area with the heald frame movement. The multi-path optical paths distributed front and rear also perform collaborative verification: when the first path near the back beam detects a breakage... When a subsequent optical path that is continuously blocked and close to the heald frame detects a momentary sweeping pulse, it is confirmed with high confidence as a break in the flexible yarn. When two or more subsequent optical paths detect sweeping pulses originating from the same break in adjacent time windows, it is confirmed as a break in the rigid yarn. When only a single optical path detects a sweeping pulse, it is first marked as suspected, and it is confirmed only when it recurs at the same position for two consecutive weaving cycles. The optical paths distributed along the warp direction and the multi-layered optical paths layered on top and bottom are two independent deployment dimensions, and one of them can be used or used in combination.

10. The method according to claim 1, characterized in that, The alarm indication in step S5 includes at least one of audible and visual alarm, screen display alarm, or remote communication alarm; the audible and visual alarm is achieved by flashing indicator lights corresponding to the broken meridian level and lateral coordinate position and emitting a buzzer sound.

11. A multi-optical-path time-division polling intelligent detection system for warp breakage in looms, characterized in that, include: The multi-layer optical path detection module includes three horizontally opposed laser emitting and receiving units (upper, middle, and lower layers) for time-division multiplexing and polling continuous light intensity signals from the warp yarn arrangement area; the main shaft synchronization module is used to acquire the loom main shaft phase signal and output a phase-locking trigger signal; and the signal processing module is used to preprocess the continuous light intensity signal and extract waveform feature vectors. The AI ​​classification module runs a trained machine learning classifier based on waveform feature vectors, performs classifier inference operations, and outputs the warp breakage determination result, the level of the warp breakage, and the type of warp breakage. The dynamic baseline compensation module periodically collects the light intensity baseline under unobstructed conditions within the full lifecycle adaptive calibration framework, updates the baseline parameters using an exponentially weighted moving average algorithm, and compensates for baseline drift caused by dust attenuation, lens contamination, and light source aging. The full lifecycle adaptive calibration framework includes four stages: initial baseline calibration, dynamic light attenuation compensation, environmental linkage calibration, and anomaly self-checking. The multi-cycle confirmation module verifies warp breakage events across cycles and outputs the warp breakage location. The edge computing module is responsible for task scheduling and edge-side inference deployment management. It collects waveform samples of warp breakage events, uploads them to the cloud server after validity screening, receives updated models from the cloud and performs local model updates and version management, and interacts with the host computer or cloud server through a communication interface. The linkage control module is used to send stop signals to the loom and trigger alarm indications.

12. The system according to claim 11, characterized in that, The edge computing module uses a microcontroller to run a random forest classifier, or a microcontroller with a floating-point unit to run a quantized one-dimensional convolutional neural network classifier; the edge computing module communicates with the host computer via LoRa or RS485, or interacts with the cloud server via Ethernet.

13. A computer-readable storage medium, characterized in that, The storage medium stores computer-executable instructions that, when executed by a processor, implement the method according to any one of claims 1 to 10.

Images