Cylinder yarn type recognition system and method based on machine vision

Abstract

The application relates to the technical field of pattern recognition, and discloses a bobbin yarn type recognition system and method based on machine vision, which comprises the following steps: yarn vibration is excited by non-contact airflow, two-way acoustic signals containing environmental noise are synchronously collected by using double microphones, adaptive noise cancellation is performed on the two-way signals to extract target acoustic signals, and then a frequency spectrum fingerprint is generated based on the target acoustic signals for comparison and recognition. The application directly acquires acoustic resonance characteristics determined by the physical nature of yarn, avoids the problem that the prior art is interfered by light and stains due to the dependence on optical appearance, can distinguish bobbins of yarns with no visual difference, and through the built-in self-calibration and multi-modal analysis mechanism, the recognition result is consistent under different working conditions, and the yarn quality can be comprehensively evaluated.

Description

Technical Field

[0001] This invention relates to a machine vision-based yarn type identification system, belonging to the field of pattern recognition technology. Background Technology

[0002] Currently, this method identifies products by collecting two-dimensional visual features such as color, texture, and shape. For products with significant visual differences, this method can meet basic identification needs and is therefore used in some production processes. However, in high-end textiles and other fields, it is often necessary to distinguish yarns with different blending ratios or material compositions but highly consistent visual appearances. In such scenarios, the identification method relying on optical images will face its inherent limitations. This is because the type of yarn is uniquely determined by its three-dimensional physical properties such as density and fiber structure, while optical images are merely two-dimensional projections of these three-dimensional physical properties under specific lighting and angles. This projection process itself causes the loss of key physical information.

[0003] Looking further, what the algorithm learns from these incomplete projection information is a statistical association between visual features and material types. This association is not based on physical causality. In actual production environments, when the surface of the yarn package is subject to electrostatic adsorption of trace amounts of dust or slight humidity differences during process flow, causing subtle changes in its optical properties that are imperceptible to the human eye, the aforementioned statistical association is easily disrupted, thereby directly affecting the stability of the recognition results.

[0004] Specifically, the shortcomings of existing identification methods are mainly reflected in the following aspects: 1. The identification basis is indirect and volatile optical appearance, thus it cannot reliably distinguish yarn packages with different material properties but highly consistent visual characteristics; 2. The stability of the identification process is easily affected by factors such as light, stains, or deformation in the production line environment. Even if higher resolution cameras are used to obtain more detailed images, the fundamental problem caused by the loss of physical information in optical projection cannot be solved. Therefore, how to establish an identification method that can directly measure the intrinsic physical properties of yarn packages to obtain stable identification basis that can characterize its material nature, thereby avoiding interference from optical appearance and environmental factors, and achieving stable identification of various types of yarn packages, has become the technical problem to be solved by this invention. Summary of the Invention

[0005] This invention provides a machine vision-based yarn type identification system, the main purpose of which is to solve the problem that existing technologies rely on optical appearance for identification, which cannot reliably distinguish yarns with different material properties but similar visual appearances and are easily affected by environmental interference.

[0006] To achieve the above objectives, the present invention provides a machine vision-based yarn type identification system, the system comprising:

[0007] One excitation unit;

[0008] A pickup unit, comprising two microphones arranged at a preset distance;

[0009] A processing unit is connected to an excitation unit and a pickup unit. The processing unit is configured to: control the excitation unit to apply a high-frequency pulsed airflow to the yarn of the bobbin, causing forced vibration of the yarn; control two microphones in the pickup unit to simultaneously acquire two acoustic signals, including the acoustic signal generated by the forced vibration of the yarn and ambient noise; execute an adaptive noise cancellation algorithm based on the differential information between the two acoustic signals to generate a target acoustic signal that has suppressed structural conducted noise and ambient noise; generate a spectral fingerprint reflecting the resonance characteristics of the yarn using a fast Fourier transform based on the target acoustic signal; and compare the spectral fingerprint with a reference fingerprint database to output an identification result characterizing the type of bobbin.

[0010] Preferably, the excitation unit includes a miniature air pump and a precision nozzle, and the processing unit controls the high-frequency pulsed airflow generated by the excitation unit to be a pulsed airflow that sweeps the frequency in the range of 1 kHz to 20 kHz.

[0011] Preferably, the processing unit is further configured to: identify a fundamental frequency resonant and at least one harmonic resonant from the spectral fingerprint; and generate a dimensionless proportional fingerprint, wherein the dimensionless proportional fingerprint is composed of the frequency ratio R between the harmonic resonant and the fundamental frequency resonant. n =f n+1 / f1 is formed, where f1 is the frequency of the fundamental resonant peak, f n+1 The frequency of the nth harmonic resonance peak is used, and the processing unit specifically compares the values ​​based on a dimensionless proportional fingerprint.

[0012] Preferably, the processing unit is further configured to: measure the transit time from the generation of the high-frequency pulsed airflow excitation in the excitation unit to the capture of the acoustic signal by the pickup unit; calculate the sound speed of the current environment based on the transit time and the fixed distance between the excitation unit and the pickup unit; and normalize and calibrate the spectral fingerprint using the sound speed of the current environment and the standard sound speed.

[0013] Preferably, the processing unit is further configured to: control the excitation unit to apply a non-resonant instantaneous acoustic pulse excitation to the yarn package; control the pickup unit to capture the acoustic echo signal reflected by the yarn package; and determine the winding density of the yarn package by analyzing the time-domain attenuation characteristics of the acoustic echo signal.

[0014] Preferably, the processing unit is further configured to: after outputting the recognition result, control the excitation unit to apply energy excitation at the frequency of the strongest resonance peak in the spectral fingerprint; after the energy excitation stops, control the pickup unit to capture the time-domain decay signal of the yarn vibration; and generate data characterizing the physical integrity of the yarn based on the decay rate of the time-domain decay signal.

[0015] Preferably, the system further includes an electrostatic probe; and the processing unit is configured to: after initially determining the yarn type through spectral fingerprinting, control the excitation unit to rub the yarn with airflow, and confirm or correct the initially determined yarn type based on the triboelectric potential characteristics sensed by the electrostatic probe.

[0016] Preferably, the processing unit executes an adaptive noise cancellation algorithm, including the following logic: defining one of the two acoustic signals as the main signal; defining the other as the reference signal; adjusting the filter coefficients through iterative calculation to minimize the mean square error between the main signal and the filtered reference signal; and subtracting the filtered reference signal from the main signal to generate the target acoustic signal. The processing unit compares the spectral fingerprint with the benchmark fingerprint database, specifically: extracting the frequencies and amplitudes of multiple resonance peaks from the spectral fingerprint to form a feature vector; calculating the cosine similarity between the feature vector and the standard fingerprint vector stored in the benchmark fingerprint database; and outputting the yarn type corresponding to the standard fingerprint vector when the cosine similarity is greater than the matching threshold.

[0017] Preferably, the benchmark fingerprint library includes acoustic fingerprints of various standard yarn bobbin materials. The acoustic fingerprints are generated by performing multiple excitation and pickup operations on standard yarn bobbin samples under a standard environment with constant temperature and humidity, and then statistically averaging and extracting features from the collected acoustic signals.

[0018] A machine vision-based method for identifying yarn type, comprising the following steps:

[0019] Step a: Apply high-frequency pulsed airflow to the yarn of the bobbin to induce forced vibration in the yarn;

[0020] Step b: Two microphones arranged at a preset distance are used to simultaneously collect two acoustic signals, including the acoustic signal generated by the forced vibration of the yarn and the ambient noise.

[0021] Step c: Based on the differential information between the two acoustic signals, an adaptive noise cancellation algorithm is executed to generate a target acoustic signal that has suppressed structural conducted noise and environmental noise.

[0022] Step d: Based on the target acoustic signal, a spectral fingerprint reflecting the resonance characteristics of the yarn is generated by fast Fourier transform;

[0023] Step e involves comparing the spectral fingerprint with a benchmark fingerprint database to output an identification result characterizing the yarn type. Compared to existing technologies, the advantages of this invention are:

[0024] 1. This invention uses non-contact high-frequency pulsed airflow to induce forced vibration in the yarn, and utilizes two microphones arranged at a preset distance to synchronously collect acoustic signals. Then, based on the differential information of the two signals, an adaptive noise cancellation algorithm is executed. This combination of technical features enables the system to accurately separate the weak target acoustic signal generated solely by yarn vibration from noisy industrial background noise. Since the subsequently generated spectral fingerprint is directly derived from this target acoustic signal, it represents the intrinsic physical properties determined by the yarn's own density and elastic modulus. Therefore, the identification method of this invention avoids the identification uncertainty caused by existing technologies that rely on variable optical appearances such as light, color, and surface stains. Especially for yarn packages that are visually indistinguishable but have subtle differences in physical material, it can also achieve stable and reliable type differentiation.

[0025] 2. After obtaining the spectral fingerprint of the yarn's resonance characteristics, this invention does not use its absolute frequency value as the final identification basis. Instead, it further calculates the frequency ratio between the fundamental frequency and each harmonic to form a dimensionless proportional fingerprint. At the same time, in each identification operation, the system also uses the transit time of the excitation signal from the excitation unit to the pickup unit to reconstruct the sound speed of the current environment in real time, and uses this to normalize and calibrate the entire spectral fingerprint. These two algorithm-level processing methods work together to ensure that the identification results are not affected by the internal tension changes of the yarn due to consumption, and also avoid external measurement drift caused by changes in sound speed due to fluctuations in workshop temperature and humidity. Ultimately, this ensures that the identification results of the same yarn are consistent throughout its entire service life and under different environmental conditions.

[0026] 3. This invention utilizes the same set of excitation and pickup units, and achieves non-destructive detection of multi-dimensional information of yarn packages by switching the operating logic of the processing units. After completing material identification by analyzing the steady-state spectral fingerprint, the system can then apply single-frequency energy excitation and analyze the time-domain decay signal of subsequent vibrations to obtain data characterizing the physical integrity of the yarn, or apply instantaneous acoustic pulses and analyze the time-domain characteristics of their echo signals to determine the winding density of the yarn packages. This working mode of switching between frequency domain analysis and time domain analysis allows the single system to expand its function from simple material identification to quality flaw detection and structural evaluation without increasing any hardware costs, providing multi-dimensional technical support for quality control in the production process. Attached Figure Description

[0027] Figure 1 This is a flowchart of the yarn type identification and quality assessment method of the present invention;

[0028] Figure 2 This is a schematic diagram of the time-domain attenuation signal characterizing the physical integrity of the yarn according to the present invention;

[0029] Figure 3 This is a logical diagram illustrating the switching between different operating modes of the system of the present invention. Detailed Implementation

[0030] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be described in further detail below; obviously, the described embodiments are some embodiments of the present invention, but not all embodiments; based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0031] This invention provides a machine vision-based yarn type identification system. Functionally, the system includes an excitation unit, a pickup unit, and a processing unit connected to them. During operation, the processing unit directs the excitation unit to apply non-contact energy excitation to the yarn of the test yarn and controls the pickup unit to capture the response signal characterizing the yarn's physical properties induced by the excitation. Subsequently, the processing unit analyzes the response signal to generate a spectral fingerprint reflecting the yarn's resonance characteristics. Finally, by comparing this spectral fingerprint with a preset benchmark fingerprint database, the identification result characterizing the yarn type is output. In an application scenario that distinguishes high-end blended yarns with consistent visual appearance, methods relying on optical image analysis cannot find stable representations from two-dimensional appearances such as color or texture. The limitations of traditional methods make them difficult to apply, especially when there are changes in lighting conditions in the production line environment or when there are traces of dirt on the yarn surface, resulting in insufficient reliability of the identification results. To address this challenge, the excitation unit of this invention is configured as a pneumatic assembly consisting of a miniature air pump and a precision nozzle. The processing unit controls the excitation unit to apply a high-frequency pulsed airflow excitation to the yarn of the bobbin, which performs a linear frequency sweep in the frequency range of 1 kHz to 20 kHz. The duration is calibrated to 50 milliseconds. The pressure of the airflow is controlled by a proportional valve to a threshold that can effectively excite the yarn to produce forced vibration without disturbing its physical structure. In this way, through a non-contact and energy-controlled pneumatic excitation, the system prepares the initial vibration conditions carrying the physical information of the yarn for subsequent acoustic analysis.

[0032] Given the presence of environmental noise and structural conducted noise generated by equipment operation in industrial settings, the energy of which may be far greater than the weak target acoustic signal generated by yarn vibration, the acquisition of the target signal poses a technical challenge. To address this, the pickup unit of this invention is configured to include two omnidirectional MEMS microphones arranged side-by-side at a preset spacing of 1 to 2 centimeters. The midpoint of the line connecting these two microphone arrays is spatially aligned with the intersection of the aforementioned high-frequency pulsed airflow and the yarn. The processing unit is configured to acquire two acoustic signals in parallel using a synchronous clock, defining one as the main signal and the other as the reference signal. In terms of signal processing logic, the processing unit executes an adaptive noise cancellation algorithm based on minimum mean square error. This algorithm utilizes the physical difference between highly correlated noise components in the two signals and the target signal components with phase differences due to the distance to the sound source. By iteratively adjusting the filter coefficients, it subtracts the filtered reference signal, which matches the noise components, from the main signal. Through this differential acoustic cancellation mechanism, the system can separate the target acoustic signal generated by yarn vibration from the background noise, thereby providing data with a high signal-to-noise ratio for subsequent spectral analysis.

[0033] Before this system is deployed in practical applications, its sensor probe, which includes the excitation unit and the pickup unit, needs to undergo a preliminary geometric and aerodynamic parameter calibration procedure. This procedure first solidifies the physical structure of the nozzle to have an outlet inner diameter of 0.1 to 0.5 mm and an aspect ratio of 4 to 10 to form a collimated airflow. Then, the two microphones of the pickup unit are placed on a one-dimensional linear guide rail. For a standard yarn sample, the system traverses the track in 0.1 mm increments within a 1 to 2 cm interval, and records the convergence rate γ of the adaptive noise cancellation algorithm at each position. Here, the convergence rate γ is defined as the number of iterations required for the algorithm's mean square error to decrease from the initial value to below the stable threshold. Finally, the interval that maximizes the γ value is selected as the solidified geometric parameter for this application scenario, thereby determining the physical layout and aerodynamic characteristics of the probe at an operating point that can improve the signal-to-noise ratio and processing speed. To convert the purified time-domain signal into a quantifiable and comparable feature, the processing unit is configured to first perform a Fast Fourier Transform on the target acoustic signal. The system generates a spectral fingerprint reflecting the resonance characteristics of the yarn. This spectral fingerprint is a feature vector composed of the frequencies and amplitudes of multiple resonance peaks. To classify this feature vector, the system has a built-in benchmark fingerprint library. This library is generated by repeatedly exciting and picking up standard yarn samples of various categories under a standard environment with constant temperature and humidity, and then statistically averaging and extracting features from the collected acoustic signals. In the recognition process, the processing unit calculates the cosine similarity between the spectral fingerprint feature vector of the yarn to be tested and each standard fingerprint vector stored in the benchmark fingerprint library. The calculated similarity value is compared with a preset matching threshold. This matching threshold, for example, 0.95, is a value that achieves a balance between preset accuracy and false positive rate, determined by performing receiver operation feature curve analysis on an offline verification sample set. When a certain cosine similarity is greater than this threshold, the processing unit outputs the yarn type corresponding to that standard fingerprint vector as the recognition result.

[0034] Furthermore, considering that the overall tension of the yarn package may change during consumption, causing a drift in the absolute value of the resonant frequency and affecting the stability of the identification, the processing unit is also configured to execute a harmonic ratio-based calibration logic to avoid this problem. Under this logic, the processing unit first automatically identifies the strongest fundamental frequency resonant peak from the spectral fingerprint, denoting its frequency as f1, and then sequentially identifies at least one harmonic resonant peak, denoting its frequency as f1. n+1 Subsequently, the processing unit calculates a set of frequency ratios R between the harmonic resonant peaks and the fundamental frequency resonant peaks. n =f n+1The dimensionless proportional fingerprint constructed by / f1 is used for database comparison. Since the proportional relationship between the harmonics of the vibrating body and the fundamental frequency is mainly determined by its geometry and material, and its sensitivity to tension changes is lower than that of the absolute value of the frequency, this mechanism ensures that the identification results remain consistent under different remaining amounts of yarn in the bobbin.

[0035] Furthermore, to avoid errors in frequency measurement caused by changes in air velocity due to fluctuations in workshop temperature and humidity, the processing unit is configured to execute a normalized calibration procedure based on acoustic transit time. At the beginning of each identification operation, the processing unit starts a high-precision timer the instant the excitation unit generates a pulsed airflow, and the pickup unit monitors the leading edge of the pulsed airflow sound pressure signal. When this leading edge is captured, the timer stops, thus obtaining a transit time t. current Based on the fixed physical distance d between the pickup unit and the excitation unit, the processing unit calculates according to formula c. current =d / t current The speed of sound in the current environment is calculated in real time; the apparent resonant frequency f is obtained through FFT. measured Then, the system uses a preset standard speed of sound c standard Compared with the current measured speed of sound c current Through formula f normalized =f measured ×(c standard / c current The entire spectral fingerprint is normalized and calibrated. This procedure utilizes the physical processes of the system's own operation to correct for the effects of changes in the acoustic medium, ensuring the comparability of identification results under different operating conditions. To extend the system's functionality from type identification to yarn quality evaluation, the processing unit is also configured to integrate multiple working modes. After completing the yarn type identification, the system can switch to a quality inspection mode. In this mode, the processing unit controls the excitation unit to apply a brief single-frequency energy excitation to the yarn at the resonant peak frequency with the strongest energy in the identified spectral fingerprint. After the excitation stops, the pickup unit captures the time-domain decay signal of the yarn's free vibration and calculates the decay rate of the signal envelope. The system generates data characterizing the physical integrity of the yarn by measuring the rate of vibration energy decay. A yarn damaged due to internal fiber breakage or uneven twist has greater internal damping, which leads to a faster rate of vibration energy decay. In addition, the system can also switch to a winding density evaluation mode. In this mode, the excitation unit is controlled to apply a non-resonant instantaneous acoustic pulse excitation to the yarn package, while the pickup unit captures the acoustic echo signal reflected by the entire yarn package. The processing unit determines the winding density of the yarn package by analyzing the time-domain decay characteristics of the echo signal. A tightly wound yarn package will produce an echo with slower decay, while a loosely wound yarn package will produce an echo with faster decay, thereby outputting quantitative data characterizing the winding density of the yarn package.

[0036] To further enhance the reliability of decision-making in specific applications, the system may also include an electrostatic probe positioned near the nozzle of the excitation unit, and the processing unit is configured to execute a decision arbitration process. After initially determining the yarn type through spectral fingerprinting, the processing unit controls the excitation unit to continuously rub the yarn with airflow. Utilizing the independent physicochemical property that different yarn materials acquire different charges due to triboelectric effect, the electrostatic probe senses the potential characteristics of the yarn surface. The processing unit compares the sensed potential characteristics with an internally stored material-triboelectric correspondence table. If the acoustic recognition result matches the electrostatic verification result, the recognition result is finally confirmed; if they conflict, an uncertainty alarm is output. This mechanism reduces the risk of misjudgment that may arise from a single principle by introducing an independent verification channel. For specific applications requiring the highest recognition reliability, the system can be equipped with a decision arbitration module. This module is activated after the acoustic fingerprint comparison yields the preliminary recognition result and controls the electrostatic probe to continuously rub the yarn with airflow for 2 seconds to collect a stable triboelectric potential V. ss Subsequently, the system queries a preset list containing various standard materials and their corresponding potential ranges [V]. min V max The calibration lookup table is used, and the following arbitration logic is executed: if the preliminary identification result is material X and the measured V... ss In its corresponding [V minX V maxX If the value is within the specified range, then output a high-confidence confirmation result; if V ss Falling into the range of another material Y [V minY V maxY If the result is not found to be in conflict, an uncertainty alarm will be output. This independent physical dimension verification mechanism, by introducing deterministic quantitative comparison logic, provides an independent verification mechanism for the final output of the identification system.

[0037] Example 1: In an automated textile workshop producing high-end blended fabrics, two parallel production lines run two types of yarn that are visually indistinguishable: one is a blend of 70% cotton and 30% Tencel, and the other is a blend of 65% cotton and 35% Tencel. The two yarns not only share the same color and basic texture, but also attract airborne dust due to electrostatic attraction during high-speed flow. Furthermore, differences in humidity from previous processes cause subtle changes in their surface optical properties that are difficult for the human eye and even high-resolution cameras to consistently distinguish. Any identification method relying on analyzing two-dimensional optical image information degrades to a level close to random guessing, thus leading to the greatest impact of material mixing. The quality of the final fabric; to address this working condition, a yarn type identification system was deployed on-site. This system did not attempt to use image algorithms to deal with the ever-changing optical appearance, but instead applied high-frequency pulsed airflow excitation to the yarn via its excitation unit. This shifted the physical basis for identification from optical reflection characteristics to acoustic resonance characteristics determined by the yarn's own density and elastic modulus. Furthermore, when picking up the response signal, the system's pickup unit, which includes two MEMS microphones, and the adaptive noise cancellation algorithm executed by the processing unit worked together to separate the target acoustic signal, which had suppressed environmental and structural conducted noise, from the noise of the workshop equipment. This provided a data foundation for subsequent spectrum analysis.

[0038] The processing unit receives the noise-suppressed signal, performs a Fast Fourier Transform, and identifies the distinguishable fundamental frequency resonant peak f1 and harmonic resonant peak f2 from it. n+1 At this point, the system does not directly use these absolute frequency values ​​that drift with changes in yarn tension, but further calculates their frequency ratio R. n =f n+1 The / f1 function generates a dimensionless proportional fingerprint that is insensitive to tension changes. This series of sequential processing steps, from noise suppression to frequency feature extraction and then to fingerprint dimensionless generation, enables the system to maintain the consistency of recognition results while adapting to changing working conditions. Finally, by comparing the dimensionless proportional fingerprint generated on-site with the benchmark fingerprint database, the system continuously distinguishes between yarns of two blended ratios and sends the recognition results to the central control system of the production line. The incidence of related material mixing events is significantly reduced, the material accuracy of the entire production process is improved, and the consistency of the final fabric quality is maintained.

[0039] Example 2: This example provides an experiment to verify the effectiveness of the aforementioned technical solution. The test platform consists of a yarn type identification system and is placed in an acoustic test chamber with a background noise level below 30 decibels. To simulate an industrial environment, a broadband noise generator is also installed in the chamber to generate a sound pressure level of 85 decibels and a noise frequency characteristic referencing a typical textile workshop environment. Four types of yarn samples were selected for this experiment: 100% pure cotton yarn, 100% polyester yarn, 70% cotton and 30% Tencel blended yarn, and 65% cotton and 35% Tencel blended yarn. Samples C1 and C2 are indistinguishable visually. The system's baseline fingerprint database has pre-recorded standard fingerprints of four samples under standard conditions. The cosine similarity matching threshold used in the experiment needs to balance the recall and precision of the identification. The specific value of this threshold is determined by an independent calibration process. This process performs receiver operating characteristic curve analysis on a validation set containing hundreds of samples and selects the inflection point on the curve that simultaneously optimizes the true positive rate and false positive rate as the threshold. For the sample set used in this experiment, the matching threshold obtained by this calibration process is 0.95.

[0040] The experiment was divided into two groups. The first group was conducted in a quiet environment in the acoustic test chamber, while the second group was conducted in a simulated industrial noise environment of 85 dB using a broadband noise generator. Each sample under each group condition was repeatedly executed 100 times for independent identification, and its harmonic ratio R1 = f2 / f1 and cosine similarity with the corresponding benchmark fingerprint were recorded. In the quiet environment, all samples were accurately identified. Although the harmonic ratio R1 of samples C1 and C2 were close in value, their distribution ranges did not overlap, showing distinguishable feature differences. When switched to a noisy environment, the original pickup signal contained significant noise components. However, after processing by the system's adaptive noise cancellation algorithm, the values ​​of the key feature parameters of the spectral fingerprint generated by the output target acoustic signal, as shown in Table 1, did not deviate significantly from the results in the quiet environment.

[0041] Table 1: Key Data Recording Table for Yarn Cylindrical Identification Test

[0042]

[0043] Referring to Table 1, the decrease in cosine similarity values ​​under noisy conditions compared to those under quiet conditions is within the allowable fluctuation range of the set 0.95 matching threshold. This result indicates that the system's adaptive noise cancellation algorithm can suppress environmental interference and preserve signal fidelity for subsequent feature extraction. Meanwhile, the harmonic ratio R1 values ​​of samples C1 and C2 remain stable and separable under both environments, indicating that the identification mechanism based on dimensionless proportional fingerprints can distinguish materials with subtle differences in physical properties.

[0044] Example 3: This example combines Figures 1 to 3 This section describes a machine vision-based yarn type identification system and method, such as... Figure 1 As shown in the diagram, the process begins with applying excitation to the yarn bobbin under test. The excitation unit applies high-frequency pulsed airflow excitation to the yarn. Subsequently, the pickup unit synchronously acquires acoustic signals containing the target signal and noise through dual microphones. The adaptive noise cancellation module extracts the target acoustic signal from these signals. Then, a raw spectral fingerprint is generated through Fast Fourier Transform (FFT). This fingerprint enters a multi-dimensional normalization calibration module, which contains two sub-functions: one is to generate a dimensionless proportional fingerprint to eliminate the influence of yarn tension changes, and the other is to perform sound velocity transit time calibration to avoid interference from environmental temperature and humidity fluctuations. After calibration, the fingerprint is finally sent to the fingerprint comparison and recognition module, where cosine similarity calculation is performed with a benchmark fingerprint library to output the yarn bobbin type identification result. In addition, the system also has a multi-modal quality assessment module, which can obtain the physical integrity data of the yarn by analyzing the time-domain attenuation signal after identification, or determine the winding density of the yarn bobbin by analyzing the acoustic echo signal.

[0045] like Figure 2 As shown in the figure, the horizontal axis represents time in milliseconds (ms), and the vertical axis represents normalized amplitude. The figure contains three curves, representing healthy samples, defective samples, and a judgment threshold for classification. The healthy sample, represented by the solid line, has a slower vibration energy decay and a time constant of τ = 450 ms. The defective sample, represented by the dotted dashed line, has a faster energy decay due to its greater internal damping and a time constant of τ = 280 ms. The judgment threshold, represented by the long dashed line, is set at τ = 350 ms. The system determines whether the physical integrity of the sample is up to standard by comparing the measured decay time constant with this threshold.

[0046] like Figure 3 As shown, the system initially enters a standby / ready state. When the yarn package is detected to be in place, it enters the recognition execution state, and performs environmental calibration processes such as sound velocity normalization calibration during this process. After obtaining a valid fingerprint, the system enters the category recognition completion state and finally outputs the result. After that, the system can enter different functional modes according to the received external instructions. If a new material learning instruction is received, the system switches to the new material learning mode. After learning is completed and the database is updated, it returns to the standby / ready state. If a quality flaw detection instruction is received after outputting the result, the system will enter the quality flaw detection mode and return to the standby / ready state after the flaw detection is completed, thus forming a complete, multi-functional closed-loop workflow.

[0047] Example 4: To construct and maintain the benchmark fingerprint database built into the aforementioned system, before the system is put into use, the operator selects N known types of standard yarn bobbins. For each type of standard yarn bobbin, M independent yarn bobbins with no physical differences are selected. The system sequentially executes the acoustic recognition process described in the aforementioned specific implementation for these N×M yarn bobbins, collecting and processing their acoustic signals to generate M independent dimensionless proportional fingerprint feature vectors for each type of standard yarn bobbin. Subsequently, the processing unit performs statistical analysis on the M feature vectors belonging to the same type, calculating their geometric representation in the multidimensional feature space. The center is the mean vector, which is used as the final standard fingerprint of this type of standard yarn and stored in the benchmark fingerprint database. At the same time, in order to define the acceptable fluctuation range of this type of sample, the processing unit also calculates the maximum Euclidean distance between the M feature vectors and the mean vector, and uses this distance value as the confidence radius of the match. In the subsequent online recognition process, when the feature vector of a yarn to be tested not only has a cosine similarity greater than 0.95 with a certain standard fingerprint, but its Euclidean distance with the standard fingerprint also falls within the set confidence radius, the recognition result is marked as a high confidence match.

[0048] When a new type of yarn package not included in the baseline fingerprint library is introduced into the production line, the system can enter a new material learning mode. The operator places the standard yarn package of this new type at the detection position and triggers the learning process. The system will automatically execute the aforementioned M independent measurement and statistical analysis processes to generate the standard fingerprint and confidence radius of the new material, adding it as a new entry to the baseline fingerprint library, thus expanding the library on-site. Furthermore, in the quantitative calibration of carbon fiber yarn package quality inspection, offline calibration is first performed. A batch of carbon fiber yarn packages that pass physical performance testing are selected as healthy samples, and another batch of yarn packages confirmed to have internal microcracks through microstructural scanning are selected as defective samples. The system performs a quality inspection process on these two batches of samples one by one. That is, after identifying their main resonant frequency, a brief single-frequency energy excitation is applied, and the time-domain decay signal after the excitation stops is captured. The processing unit performs exponential fitting on the amplitude envelope of the time-domain decay signal to obtain a time constant τ characterizing its energy decay rate. Healthy samples, due to their small internal damping, have a lower time constant τ. good The values ​​are distributed within a numerical range, while the defective sample experiences increased vibrational energy dissipation due to internal damage, resulting in a time constant τ. bad The values ​​are then distributed in another, lower range. By analyzing the statistical distribution of these two sets of data, a threshold τ can be set to distinguish between healthy and defective states. threshold Its value is set at the midpoint of two distribution intervals to maximize the classification interval; after offline calibration, the system can perform online physical integrity quantitative assessment. When a roll of carbon fiber yarn passes by on the production line, the system automatically enters the quality flaw detection mode after completing the type identification and calculates its vibration decay time constant τ.measured The processing unit compares this measured value with the calibrated threshold τ. threshold Compare, if τ measured Less than τ threshold If the system outputs a warning signal, it indicates that the yarn bobbin is at risk of internal physical damage.

[0049] Example 5: Before deploying the system in an industrial environment with wide temperature variations, the integrated sensor probe, which includes the excitation unit and the pickup unit, needs to undergo a preliminary hardware calibration. This procedure involves placing the probe in a temperature and humidity-controlled environmental test chamber and fixing a standard reference target with stable physical properties and a known resonant frequency in front of it. Then, within the entire preset operating temperature range of the system, the system is driven to continuously measure the standard reference target after stabilizing at each temperature point for 30 minutes in 5-degree Celsius increments.

[0050] At each stable temperature point, the processing unit records the apparent resonant frequency and response amplitude offset of the measured reference target material, and simultaneously records the current temperature value measured by a temperature sensor integrated on the probe. By performing polynomial fitting on the offset data and temperature data across the entire temperature range, the system generates a set of hardware calibration coefficients that characterize the sensor probe response as a function of temperature. These calibration coefficients are stored in the non-volatile memory of the processing unit. In subsequent field operations, the processing unit uses the corresponding calibration coefficients based on the real-time measured ambient temperature to compensate for the original signal acquired by the pickup unit, thereby correcting the measurement error introduced by hardware thermal drift and maintaining the consistency of the recognition results under different ambient temperatures.

[0051] Example 6: When the system is adapted to a new type of yarn package or production line station, the geometric layout of its sensor probe is determined through a pre-calibration procedure. This procedure first confirms that there is a yarn package in place at the station to be tested using an infrared through-beam sensor. If no object is detected, the subsequent process is stopped. After confirming that the object is in place, a sample of the new standard yarn package of this type is fixed on a three-axis movable platform. While keeping the angle between the excitation nozzle and the yarn surface unchanged, the vertical distance between the nozzle and the yarn surface and the vertical distance between the pickup unit and the yarn surface are adjusted. Under each geometric position combination, an acoustic signal excitation and acquisition process is executed once, and the signal-to-noise ratio of the target acoustic signal obtained after processing is calculated. By traversing the preset distance range, a geometric position combination that maximizes the signal-to-noise ratio of the target acoustic signal can be determined. This combination is determined as the geometric layout parameters for this application scenario and is fixed.

[0052] After determining and fixing the geometric layout of the sensor probe, the pulse airflow intensity of the excitation unit is calibrated. The processing unit controls the miniature air pump of the excitation unit, so that the output airflow pressure starts from an initial low value and gradually increases in fixed pressure steps. At each pressure step, the system performs a measurement and records the vibration amplitude of the acquired target acoustic signal. At the same time, an external machine vision module is used to monitor whether the yarn surface is loosened or structurally disturbed under the impact of airflow. By analyzing the recorded pressure-vibration amplitude relationship curve, a saturation region and a linear region of amplitude response can be determined. The calibration procedure selects the end of the linear region, that is, the pressure value when the amplitude growth begins to slow down and is lower than the pressure value when any physical structural disturbance occurs, as the excitation intensity of this type of yarn. The finally determined geometric layout and excitation intensity parameters are stored in the processing unit for subsequent online identification tasks of this type of yarn.

[0053] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0054] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A machine vision-based cone yarn variety identification system, characterized by, The system comprises: an excitation unit; a pickup unit, the pickup unit comprising two microphones arranged at a preset interval; a processing unit, the processing unit being connected with the excitation unit and the pickup unit; wherein the processing unit is configured to: control the excitation unit to apply a high-frequency pulsed airflow excitation to the yarn of the cheese, so that the yarn generates forced vibration; control the two microphones of the pickup unit to synchronously collect two acoustic signals containing acoustic signals generated by the forced vibration of the yarn and environmental noise; perform an adaptive noise cancellation algorithm based on the difference information between the two acoustic signals to generate a target acoustic signal in which the structure-borne noise and the environmental noise are suppressed; based on the target acoustic signal, generate a frequency spectrum fingerprint reflecting the resonance characteristics of the yarn through fast Fourier transform; and compare the frequency spectrum fingerprint with a reference fingerprint library to output an identification result representing the type of cheese; identifying a fundamental resonance peak and at least one harmonic resonance peak from the spectral fingerprint; and generating a dimensionless ratio fingerprint, the dimensionless ratio fingerprint comprising a frequency ratio of the harmonic resonance peak to the fundamental resonance peak comprising is the frequency of the fundamental resonance peak, is the frequency of the nth harmonic resonance peak, the processing unit specifically bases the comparison on the dimensionless ratio fingerprint.

2. A machine vision based cone yarn variety identification system as claimed in claim 1 wherein, The excitation unit comprises a micro air pump and a precision nozzle, and the processing unit controls the excitation unit to generate a high-frequency pulsed airflow excitation, which is a pulsed airflow for frequency sweeping in the frequency range of 1 kHz to 20 kHz.

3. A machine vision based cone yarn variety identification system as claimed in claim 1 wherein, The processing unit is further configured to: measure the transit time from the generation of the high-frequency pulsed airflow excitation at the excitation unit to the capture of the acoustic signal by the pickup unit; calculate the sound speed of the current environment based on the transit time and the fixed distance between the excitation unit and the pickup unit; and normalize and calibrate the frequency spectrum fingerprint using the sound speed of the current environment and the standard sound speed.

4. A machine vision based cone yarn variety identification system as claimed in claim 1 wherein, The processing unit is further configured to: control the excitation unit to apply a non-resonant transient acoustic pulse excitation to the cheese; control the pickup unit to capture the acoustic echo signal reflected by the cheese; and determine the winding density of the cheese by analyzing the time-domain decay characteristics of the acoustic echo signal.

5. A machine vision based cone yarn variety identification system as claimed in claim 1 wherein, The processing unit is further configured to: after outputting the identification result, control the excitation unit to apply an energy excitation at the frequency of the resonance peak with the strongest energy in the frequency spectrum fingerprint; after the energy excitation stops, control the pickup unit to capture the time-domain decay signal of the yarn vibration; and based on the decay rate of the time-domain decay signal, generate a data representing the physical integrity of the yarn.

6. A machine vision based cone yarn variety identification system as claimed in claim 1 wherein, The system further comprises an electrostatic probe; and the processing unit is further configured to: after preliminarily determining the type of cheese through the frequency spectrum fingerprint, control the excitation unit to rub the yarn with airflow, and confirm or modify the preliminarily determined type of cheese according to the friction electrification potential characteristics sensed by the electrostatic probe.

7. A machine vision based cone yarn variety identification system as claimed in claim 1 wherein, The system further comprises an electrostatic probe; and the processing unit is further configured to: after preliminarily determining the type of cheese through the frequency spectrum fingerprint, control the excitation unit to rub the yarn with airflow, and confirm or modify the preliminarily determined type of cheese according to the friction electrification potential characteristics sensed by the electrostatic probe.

8. A machine vision based cone yarn variety identification system as claimed in claim 1 wherein, The reference fingerprint library comprises acoustic fingerprints of multiple standard cheese materials, and the acoustic fingerprints are generated by performing multiple excitation and pickup operations on standard cheese samples in a standard environment with constant temperature and humidity, and statistically averaging and extracting features from the collected multiple acoustic signals.

9. A kind of machine vision-based cone yarn kind identification method, based on the machine vision-based cone yarn kind identification system of any one of claims 1~8, it is characterized by: The method comprises: applying a high-frequency pulsed airflow excitation to the yarn of the cheese, so that the yarn generates forced vibration; Step b, two microphones arranged at a preset interval are used to synchronously collect two acoustic signals containing acoustic signals generated by yarn due to forced vibration and environmental noise; Step c, an adaptive noise cancellation algorithm is executed based on the differential information between the two acoustic signals to generate a target acoustic signal in which structural transmission noise and environmental noise have been suppressed; Step d, based on the target acoustic signal, a frequency spectrum fingerprint reflecting the resonance characteristics of the yarn is generated through fast Fourier transform; Step e, the frequency spectrum fingerprint is compared with a reference fingerprint library to output an identification result representing the type of the cheese yarn.

Images