A method for intelligent sorting and processing quality control of cashmere raw materials

Abstract

This invention belongs to the field of cashmere raw material sorting and processing technology, specifically disclosing a method for intelligent sorting and processing quality control of cashmere raw materials. The method includes: acquiring multi-dimensional sensing data of the cashmere raw materials to be processed, including fiber scale structure images, surface gloss images, and component spectral data; performing feature-level fusion processing on the multi-dimensional sensing data to generate quantitative feature parameters; inputting the quantitative feature parameters into a machine learning hierarchical decision model to generate sorting instructions; controlling a nozzle array actuator that combines negative pressure adsorption and positive pressure jetting to perform non-contact automatic sorting according to the sorting instructions; and collecting real-time process quality data of subsequent combing and spinning processes through an IoT platform, and dynamically adjusting the sorting threshold and raw material ratio strategy based on this data. This invention achieves quantitative evaluation and closed-loop quality control of cashmere raw material quality, improving the accuracy of sorting, the adaptability of the production system, and the quality stability of the final product.

Description

Technical Field

[0001] This invention belongs to the field of cashmere raw material sorting and processing technology, and relates to a method for intelligent sorting and processing quality control of cashmere raw materials. Background Technology

[0002] As a precious natural fiber, the quality of cashmere raw materials directly determines the value and market competitiveness of the final textile products. In the cashmere processing industry chain, raw material sorting is a crucial first step, which aims to classify fibers according to indicators such as fiber length, fineness, color, and impurity content to ensure the smooth progress of subsequent processes such as combing and spinning, as well as the uniformity of the final product quality.

[0003] Currently, cashmere raw material sorting mainly relies on manual selection and traditional mechanical sorting. Manual selection depends entirely on the visual and tactile experience of sorting workers. This method is not only labor-intensive and inefficient, but also results in highly subjective and inconsistent sorting results, making it difficult to meet the high quality stability requirements of modern, large-scale production. While mechanical sorting equipment improves efficiency to some extent, its control system is usually based on simple photoelectric detection or image recognition, and can only perform rough classification on a single or a few macroscopic physical indicators such as color and length. Its control logic is fixed, and once the sorting threshold is set, it is difficult to change.

[0004] In existing technologies, whether manual or traditional mechanical sorting, the control systems are in an open-loop or isolated state. They cannot perceive the deep-seated quality characteristics of cashmere fibers, such as their microstructure and chemical composition, resulting in a one-sided sorting basis that fails to fully reflect the true spinnability of the raw materials. More importantly, there is a lack of effective data linkage and feedback mechanisms between the front-end sorting control system and the back-end carding and spinning production control systems. The quality of sorting cannot be verified and corrected through downstream actual production results, such as yarn quality and breakage rate, leading to a broken quality control chain in the entire production process. This prevents dynamic optimization of raw material classification and proportioning based on final product quality requirements, resulting in suboptimal utilization of raw material resources and potential fluctuations in product quality. Summary of the Invention

[0005] In view of this, in order to solve the problems mentioned in the background technology, a method for intelligent sorting and processing quality control of cashmere raw materials is proposed.

[0006] The objective of this invention can be achieved through the following technical solution: This invention provides a method for intelligent sorting and processing quality control of cashmere raw materials, comprising: acquiring multidimensional sensing data of the cashmere raw materials to be processed, wherein the multidimensional sensing data includes fiber scale structure images, fiber surface gloss images, and fiber composition spectral data.

[0007] The multidimensional sensing data is subjected to feature-level fusion processing to generate quantitative feature parameters related to cashmere quality.

[0008] The quantified feature parameters are input into a hierarchical decision model trained based on machine learning to generate sorting instructions corresponding to different quality levels.

[0009] According to the sorting instructions, the nozzle array actuator, which is composed of a negative pressure adsorption unit and a positive pressure spraying unit, is controlled to perform non-contact automatic sorting of cashmere raw materials, so that cashmere raw materials of different quality grades fall into the corresponding collection channels.

[0010] The IoT platform collects process quality data in real time for subsequent combing and spinning processes. The process quality data includes yield, breakage rate and yarn evenness. Based on the process quality data, the sorting threshold parameters and raw material ratio strategy are dynamically adjusted.

[0011] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. The present invention achieves accurate analysis of cashmere raw material quality through multi-dimensional perception and digital twin technology, gets rid of the limitations of traditional reliance on human senses or single physical indicators for evaluation, and improves the objectivity and consistency of quality assessment.

[0012] 2. This invention constructs a closed-loop quality control process, using the actual process quality data of downstream carding and spinning processes as the final evaluation index. This enables the upstream raw material sorting to be closely linked with the downstream production needs and actual performance, improving the self-adaptive and self-optimizing capabilities of the production system and enhancing the overall quality of yarn products and the stability of the production process from the source.

[0013] 3. This invention employs a non-contact nozzle array actuator, combined with its graded diagnostic and automatic recovery self-checking program, ensuring high efficiency, high precision, and high reliability in the sorting process. This guarantees the continuity and stability of the sorting operation, providing execution assurance for large-scale, high-quality automated production. Attached Figure Description

[0014] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0015] Figure 1 This is a schematic diagram of the method steps of the present invention.

[0016] Figure 2 This is a schematic diagram of the hierarchical decision-making model update process of the present invention.

[0017] Figure 3 This is a schematic diagram of the effective batch determination process of the present invention.

[0018] Figure 4 This is a schematic diagram of the non-contact automatic sorting process of the present invention.

[0019] Figure 5 This is a schematic diagram of the self-test procedure of the actuator of the present invention. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and 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.

[0021] Please see Figure 1 As shown, this invention provides a method for intelligent sorting and processing quality control of cashmere raw materials, the specific steps of which are as follows: Acquire multidimensional sensing data of the cashmere raw material to be processed, including fiber scale structure images, fiber surface gloss images, and fiber composition spectral data.

[0022] In a preferred embodiment of the present invention, a small-sample online calibration process is required before the fiber composition spectral data is acquired. The specific method is as follows: before each sorting task is started, no less than three reference samples with known protein content, fat content and moisture regain are retrieved from the standard cashmere sample library.

[0023] It's important to explain why small-sample online calibration is necessary: ​​the detection accuracy of near-infrared spectrometers can be affected by environmental changes and equipment fluctuations, making direct data acquisition prone to bias. By retrieving benchmark samples from a standard cashmere sample library with known protein content, fat content, and moisture regain, acquiring their raw spectral signals, and processing them to generate a calibration dataset, a local linear regression mapping model can be constructed. This allows for real-time correction of the spectral data of the raw materials being tested, avoiding inaccurate component analysis due to detection bias. This ensures the accuracy of subsequent feature extraction, grading decisions, and sorting processes, laying a solid data foundation for the entire quality control process.

[0024] Each reference sample is sequentially sent into the detection area of ​​the near-infrared spectrometer to acquire the corresponding raw spectral signals.

[0025] The original spectral signal is subjected to noise filtering and band alignment to generate a calibration spectral dataset.

[0026] Based on the calibration spectral dataset and the physicochemical parameter labels of the corresponding benchmark samples, a local linear regression mapping model is constructed.

[0027] Furthermore, the local linear regression mapping model is the core tool for calibrating fiber composition spectral data, constructed based on standard cashmere benchmark samples. It takes a calibration spectral dataset that has undergone noise filtering and band alignment as input, and uses known physicochemical parameters of the benchmark samples, such as protein content, fat content, and moisture regain, as labels. Through a local linear regression algorithm, it establishes a precise mapping relationship between spectral signals and physicochemical indicators. The model is embedded in the spectral preprocessing module of the raw material to be tested, which can correct the raw data collected by the near-infrared spectrometer in real time, offsetting detection deviations caused by environmental changes and equipment fluctuations. This ensures that the subsequently extracted spectral feature vectors are accurate and reliable, providing precise data support for quantitative feature fusion, hierarchical decision-making, and sorting processing.

[0028] For example, 10 sets of standard cashmere reference samples with known physicochemical parameters were selected. Calibration spectral data of the target band were acquired using a near-infrared spectrometer, and the characteristic band signals were extracted after preprocessing as model input. The protein content, fat content, and moisture regain of each sample were used as physicochemical parameter labels. The Epanechnikov kernel function was used to assign neighborhood weights, and cross-validation was used to determine the optimal bandwidth. For each spectral feature point to be measured, the model performed local linear fitting only using the spectral-physicochemical parameter correspondence of neighboring reference samples. Finally, a quantitative mapping between the spectral feature vector and protein content, fat content, and moisture regain was established, providing data support for subsequent sorting decisions.

[0029] The local linear regression mapping model is embedded into the spectral preprocessing module of the cashmere raw material to be tested, so as to correct the collected spectral data in real time.

[0030] The standard cashmere sample library is regularly updated by offline laboratory testing results.

[0031] The multidimensional sensing data is subjected to feature-level fusion processing to generate quantitative feature parameters related to cashmere quality.

[0032] In a preferred embodiment of the present invention, the specific method for generating the quantized feature parameters is as follows: edge enhancement and texture analysis are performed on the image acquired by the polarized light imaging device to extract scale structure features, including scale density, consistency of arrangement direction and integrity index.

[0033] It should be noted that, regarding the microscopic physical morphology of the fibers, a polarized light imaging device is required to continuously photograph the cashmere raw material moving on the conveyor belt during operation, acquiring images that clearly show the outline of the scales on the fiber surface. The role of polarized light is to enhance the contrast between the scale edges and the fiber body, facilitating subsequent processing. After acquiring the images, the system will automatically execute an edge enhancement algorithm to sharpen the scale boundaries and use texture analysis technology to quantify the scale distribution pattern, thereby calculating three key scale structural features: the scale density represented by the number of scales per unit area, the consistency of the arrangement direction reflected by the regularity of the scale orientation, and the quantitative evaluation of whether the scale edges are damaged, namely the integrity index.

[0034] The surface smoothness index is obtained by statistically analyzing the grayscale distribution and calculating the reflectance of the gloss images acquired using an imaging device.

[0035] It should be noted that, in order to evaluate the optical properties of the fibers, a high-resolution camera needs to be used in parallel to capture images of the surface gloss of the same piece of cashmere raw material. The system will perform grayscale distribution statistics on the images, analyze the distribution of bright and dark pixels in the images, and combine them with a reflectance calculation model to convert the image information into a quantitative surface smoothness index. The higher the index value, the smoother the fiber surface and the stronger the gloss.

[0036] Baseline correction and principal component analysis were performed on the spectral data acquired using a near-infrared spectrometer to extract spectral feature vectors related to protein content, lipid content, and moisture regain.

[0037] Furthermore, select the top k principal components whose cumulative contribution rate reaches a preset percentage. For example, select the top 3 principal components whose cumulative contribution rate reaches 85% or more.

[0038] It should be noted that, to investigate the intrinsic chemical composition of the fiber, a near-infrared spectrometer is needed to scan the cashmere raw material and collect its absorption spectrum data in the near-infrared band. Since the raw spectral signal often contains background noise and baseline drift, baseline correction must be performed first to obtain a pure fiber absorption spectrum. Next, principal component analysis, a data dimensionality reduction technique, is used to extract the core information that best reflects changes in composition from the complex spectral curve, ultimately forming a spectral feature vector highly correlated with three key chemical indicators: protein content, fat content, and moisture regain.

[0039] After normalizing the above-mentioned scale structure features, surface smoothness indicators and spectral feature vectors, a weighted fusion algorithm is used to generate multidimensional feature vectors.

[0040] For example, normalization is performed using three features: assuming the extracted scale density is 80-120 scales / μm. 2The surface smoothness index is 0.3-0.7, and the protein correlation component in the spectral feature vector is 1200-1800. The min-max normalization method is used, according to the formula... The calculations show that a scale density of 80 corresponds to 0, and 120 corresponds to 1. If a sample has a scale density of 100, it is normalized to 0.5. Surface smoothness of 0.3 corresponds to 0, and 0.7 corresponds to 1. If a sample has a surface smoothness of 0.5, it is normalized to 0.5. Spectral feature components of 1200 correspond to 0, and 1800 correspond to 1. If a sample has a spectral feature component of 1500, it is normalized to 0.5. After processing, all three types of features are mapped to the [0,1] interval, eliminating dimensional differences and laying the foundation for weighted fusion.

[0041] It's important to note that the core of this step lies in achieving a deep and comprehensive characterization of cashmere raw material quality through the integration of multimodal sensing technologies, surpassing the limitations of traditional methods that rely solely on single macroscopic indicators such as fiber fineness and length. The technological advantage lies in its ability to construct a far more accurate and comprehensive quality evaluation model than single-dimensional detection by comprehensively analyzing the microscopic scale structure, macroscopic optical luster, and internal chemical composition of the fiber. This comprehensive quantitative characteristic parameter can more profoundly reveal potential factors affecting downstream spinning performance, such as scale integrity affecting cohesion, fat content and moisture regain affecting static electricity and breakage, and smoothness affecting the feel and appearance of the finished product.

[0042] Furthermore, this method provides a highly reliable and precise decision-making basis for subsequent intelligent sorting, improving the accuracy of raw material grading from the source. This lays a solid foundation for optimizing the process parameters of subsequent carding and spinning processes, improving the quality stability and yield of the final yarn products, and realizing feedforward quality control of the entire cashmere processing chain.

[0043] In a preferred embodiment of the present invention, a weighted fusion algorithm is used to generate a multi-dimensional feature vector, wherein the weight coefficients of each feature dimension are dynamically determined through a periodic multi-objective optimization mechanism, specifically as follows: an optimization objective function is defined, which includes three sub-objectives: maximizing sorting accuracy, maximizing downstream production yield, and minimizing decapitation rate.

[0044] It should be noted that in this embodiment, considering the long computation time of the genetic algorithm, in order to avoid affecting the real-time response of the sorting line, the optimization process is set to run once in the background during non-production periods or after processing M batches.

[0045] It should be noted that the purpose of this function is to simultaneously pursue three core production benefits: maximizing sorting accuracy, maximizing the improvement of downstream process yield, and minimizing the reduction of breakage rate.

[0046] Based on the correlation between multidimensional feature vectors and corresponding process quality data in historical sorting batches, an input-output mapping database for the objective function is constructed.

[0047] It's important to explain that, to drive optimization, the system needs to establish an input-output mapping database. Specifically, this involves continuously collecting and organizing data from historical production batches. The normalized multidimensional feature vectors generated during the sorting of each batch are paired one-to-one with the actual process quality data fed back from subsequent combing and spinning processes, such as yield and breakage rate, to form a data association. With this database in place, a non-dominated sorting genetic algorithm is then used to iteratively optimize the combination of weight coefficients.

[0048] It should be noted that the non-dominated sorting genetic algorithm described in this invention performs iterative optimization of the weight coefficient combination on a daily or batch-by-batch basis, which is an offline optimization.

[0049] A non-dominated sorting genetic algorithm is used to iteratively search for the Pareto front solution set that optimizes the overall performance of the objective function by combining weight coefficients.

[0050] The weights that satisfy the preset constraints are selected from the Pareto front solution set as the fusion weights for the current period.

[0051] The selected weighting coefficients are applied to the weighted fusion process of the normalized scale structure features, surface smoothness index and spectral feature vector to generate a multidimensional feature vector for hierarchical decision-making.

[0052] The constraints include the sum of all weight coefficients being 1 and the value range of each weight coefficient being limited to a preset threshold.

[0053] It should be noted that this step establishes a closed-loop feedback and self-optimization mechanism from raw material characteristics to the final process result, abandoning static or empirical weight setting methods and instead allowing data-driven decision-making. Through a multi-objective optimization algorithm, this method can find the optimal balance among multiple interdependent objectives such as sorting accuracy, production efficiency, and product quality, rather than sacrificing one for another. Its technical effect lies in improving the intelligence and adaptability of sorting decisions. Because different batches of cashmere raw materials have different inherent properties and defects, and the operating conditions of downstream production equipment also change over time, a fixed weight allocation scheme is difficult to maintain optimally in all cases.

[0054] Furthermore, this method dynamically determines weighting coefficients, periodically adjusting the focus on different dimensions of features such as scales, luster, and composition based on the latest production data feedback. This allows the generated multi-dimensional feature vector to more accurately capture the key factors currently affecting the final yarn quality. This adaptive optimization capability ensures that the sorting system can continuously operate in the optimal manner, thereby more effectively improving the utilization rate of high-quality raw materials (i.e., yield) and significantly reducing yarn breakage caused by raw material issues, ultimately achieving a synergistic improvement in production efficiency and product quality.

[0055] The quantified feature parameters are input into a hierarchical decision model trained based on machine learning to generate sorting instructions corresponding to different quality levels.

[0056] For a preferred embodiment of the present invention, please refer to Figure 2 As shown, the hierarchical decision model is implemented as follows: after each sorting operation, the correspondence between the multidimensional feature vector output of the current batch of cashmere raw materials and the actual sorting action is recorded.

[0057] It should be explained that after each sorting operation is completed, the control system will immediately record and save the multi-dimensional feature vector output result of the current batch of cashmere raw materials. At the same time, it will establish a clear correspondence between this result and the actual sorting action performed by the actuator, that is, which quality grade channel the fiber is assigned to.

[0058] Upon receiving process quality data from subsequent processes, the process quality data is correlated and mapped with the multidimensional feature vector of the current batch.

[0059] For example, during the sorting of a batch of cashmere, the generated multidimensional feature vector is [scale density 0.6, smoothness 0.7, protein content 0.85, fat content 0.4, moisture regain 0.5]. Subsequent spinning processes provide feedback on process quality data as follows: yield 88%, breakage rate 28 times / 1000 spindles·hour, and yarn evenness CV value 12%. By establishing a data association table, the feature vector of this batch is bound to the corresponding quality data, and the mapping relationship between feature vector and quality result is marked. If the breakage rate exceeds the qualified threshold of 25 times / 1000 spindles·hour, the correlation logic between low scale density, high fat content, and excessive breakage rate is further marked, providing a precise mapping basis for model retraining and sorting parameter adjustment.

[0060] If the process quality data exceeds the preset acceptable range, the model retraining process is triggered, and the multidimensional feature vector of the current batch and its corresponding process quality deviation are used as new training samples.

[0061] It should be noted that the preset acceptable range needs to be determined comprehensively based on the production line's historical best data, downstream process requirements, and industry standards. For example, statistical analysis of process quality data such as yield, breakage rate, and yarn evenness from past stable production batches is used, and their 95% confidence interval is taken as the basic range. Equipment parameters such as the fiber damage tolerance threshold of carding machines and the breakage rate warning threshold of spinning machines are referenced, as well as relevant industry standards for cashmere yarn, to define upper and lower limits for indicators such as yield, breakage rate, and CV value of yarn evenness. Finally, the preset acceptable range is formed and entered into the system as the basis for judging whether the quality data is abnormal and whether model retraining is triggered.

[0062] A sliding window mechanism is used to maintain the training sample set, retaining the data from the most recent N valid batches and removing historical samples that are outside the time window.

[0063] The weight parameters of the hierarchical decision model are fine-tuned using the updated training sample set to generate an optimized hierarchical decision model suitable for the next cycle of sorting tasks.

[0064] It should be noted that fine-tuning must be carried out based on the updated effective training sample set. For example, first fix the core network structure of the model and adjust only the output layer weight parameters; use the mini-batch gradient descent algorithm with the loss function being to minimize the deviation between the predicted sorting grade and the actual quality grade and maximize the compliance rate of process quality data; set a small learning rate to avoid drastic fluctuations in weights; iterate training until the loss function converges or reaches the preset number of iterations, and finally output the optimized graded decision model to ensure that the model adapts to the latest production conditions and raw material characteristics, and improves the sorting accuracy in the next cycle.

[0065] It should be noted that when the source of raw materials changes or the operating conditions of downstream equipment drift, the model can quickly learn and adjust to adapt to new production conditions, rather than relying on manual intervention for recalibration. This continuous self-optimization capability allows the accuracy of sorting to improve continuously as production progresses, more effectively separating raw materials that may cause production problems in advance. This effectively suppresses fluctuations in downstream process quality, ensuring the long-term stability and continuous improvement of the final yarn product quality, and realizing an intelligent upgrade of quality control from passive monitoring to proactive prediction and self-optimization.

[0066] For a preferred embodiment of the present invention, please refer to Figure 3 As shown, the specific method for determining the validity of a batch is as follows: After each sorting operation is completed, check whether the nozzle array actuator has completed all the command actions and returned a confirmation signal. If any action confirmation is missing, the batch is marked as invalid.

[0067] Simultaneously evaluate the signal-to-noise ratio and feature extraction success rate of multidimensional sensing data. If the feature loss rate of any modality exceeds a preset threshold or the image blur is higher than a preset threshold, the batch of sensing data is deemed unqualified.

[0068] It should be added that the preset thresholds need to be determined in combination with equipment performance, data quality requirements and production practice: First, statistically analyze multiple batches of normally collected multidimensional sensing data to determine the usual success rate of feature extraction for each modality, the reasonable range of signal-to-noise ratio, and the benchmark for image clarity; then, based on the accuracy requirements of downstream feature fusion and hierarchical decision-making, set the feature missing rate threshold and the image blur threshold to ensure that the data meets the requirements of subsequent processing; after small-batch trial production verification and optimization, the thresholds are solidified and used as the standard for judging whether the sensing data is qualified.

[0069] When receiving subsequent process quality data uploaded by the IoT platform, verify whether its timestamp is within the preset time window after sorting is completed, and whether the data fluctuation range does not exceed the normal operating range of the equipment; otherwise, it is considered unreliable feedback.

[0070] It should be noted that the preset time window needs to be scientifically defined in conjunction with the entire production process cycle: the historical average time from the completion of cashmere sorting to the output of corresponding process quality data in downstream combing and spinning processes is statistically analyzed, and the conventional time spent on equipment data collection, transmission and processing is added to determine the basic time; then a reasonable buffer period is reserved, and finally the preset time window after the completion of sorting is defined; at the same time, it is clear that the data within the window must be consistent with the fluctuation range of quality data when the equipment is running normally under the same working conditions, to ensure that the quality data of subsequent processes are accurately correlated with the current sorting batch, and to avoid unreliable feedback caused by time misalignment.

[0071] The batch of data is included in the effective sample set of the sliding window only if all three validations pass.

[0072] The sliding window retains the data from the most recent N valid batches, where N is a preset integer.

[0073] It should be noted that this step establishes a strict data quality firewall for the self-learning process of the hierarchical decision-making model, adhering to the fundamental principle in machine learning that high-quality input determines high-quality output. Through physical verification of the executed actions, quality assessment of the perceived data source, and credibility verification of the feedback results, a full-chain credibility review mechanism is constructed, from behavior to data to result. Its technical effect is to enhance the stability and effectiveness of model training. By filtering out invalid and erroneous samples caused by equipment failure, abnormal collection, or data transmission delays, it effectively prevents the model from being misled by contaminated data and developing incorrect learning directions.

[0074] According to the sorting instructions, the nozzle array actuator, which is composed of a negative pressure adsorption unit and a positive pressure spraying unit, is controlled to perform non-contact automatic sorting of cashmere raw materials, so that cashmere raw materials of different quality grades fall into the corresponding collection channels.

[0075] For a preferred embodiment of the present invention, please refer to Figure 4 As shown, the non-contact automatic sorting of cashmere raw materials is carried out in the following manner: the real-time position of the cashmere raw materials on the conveying path is determined according to the sorting instructions, and the collection channel position corresponding to the target quality grade is matched.

[0076] The negative pressure adsorption unit is activated simultaneously to form a local negative pressure area above the cashmere raw material conveying path, causing the target fiber bundle to detach from the main material flow.

[0077] After the negative pressure adsorption action is completed, the positive pressure jet unit corresponding to the landing point is immediately activated, and the separated fiber bundle is pushed to the designated collection channel by directional airflow.

[0078] It should be noted that the nozzle array actuator structure described in this invention is a combination of multiple positive pressure nozzles nested inside a negative pressure suction port. During operation, it is transported by an adsorption belt, and different quality grades of cashmere raw materials are pushed to different collection channels using different nozzle structures.

[0079] It should be noted that after the command is issued, the system will simultaneously activate the negative pressure adsorption unit located above the cashmere raw material conveying path, forming a local negative pressure zone directly below it. The airflow suction will precisely lift and detach the identified target fiber bundles from the main material flow. After the negative pressure adsorption unit completes its operation, the system will immediately activate the positive pressure jet unit, precisely corresponding to the predetermined landing point. A directional, controlled airflow will powerfully push the adsorbed and separated fiber bundles into the designated collection channel.

[0080] It should be noted that the principle of this non-contact sorting method lies in utilizing pneumatics to achieve precise control over individual or small bundles of fibers. Through the coordinated action of a nozzle array actuator composed of a negative pressure adsorption unit and a positive pressure spray unit, the cashmere raw material is gently, quickly, and without damage. Its technical advantages lie in avoiding the physical damage that mechanical contact might cause to precious cashmere fibers, thus protecting the integrity and natural quality of the fibers to the greatest extent possible. Secondly, the seamless integration of negative pressure adsorption and positive pressure spray achieves high sorting speed and accuracy, ensuring that raw materials of different qualities are cleanly and efficiently separated into their respective channels, avoiding mixing and cross-contamination.

[0081] The system monitors the operating status of each unit in the nozzle array, including air pressure, response delay time, and airflow stability. If any parameter deviates from a preset threshold, the actuator self-test program is triggered.

[0082] It should be noted that the preset threshold setting needs to be combined with the performance parameters of the nozzle array actuator, the sorting accuracy requirements, and production practice data: First, the normal fluctuation range of air pressure value, response delay time, and airflow stability during normal equipment operation is statistically analyzed, and its ±3σ is used as the initial benchmark; then, the air pressure value threshold, response delay time threshold, and airflow stability threshold are adjusted with reference to the accuracy requirements of fiber separation during sorting; after multiple batches of trial production verification and elimination of extreme working condition data, the final threshold is optimized and determined to ensure accurate identification of actuator abnormalities, while avoiding false triggering of the self-test program.

[0083] After completing a single sorting operation, the nozzle array is cleaned and purged.

[0084] It should be noted that the built-in real-time monitoring and self-inspection cleaning mechanism ensures the long-term stability and reliability of the actuator, reduces sorting errors and production interruptions caused by equipment failure, and provides a solid physical execution guarantee for the stable and efficient operation of the entire intelligent quality control system.

[0085] For a preferred embodiment of the present invention, please refer to Figure 5 As shown, the specific method of the self-test procedure of the actuator is as follows: read the abnormal parameter type and deviation level of the deviation degree greater than the preset threshold, and determine whether it belongs to a minor fault that can be automatically compensated.

[0086] If the air pressure value is abnormal, the opening of the air supply valve will be automatically adjusted and the zero point of the pressure sensor will be recalibrated.

[0087] It should be added that the determination of abnormal air pressure value should be based on the preset normal operating air pressure range of the actuator: First, the preset air pressure standard value and allowable deviation threshold of the system should be identified. This threshold is determined based on equipment performance, sorting accuracy requirements and production practice data, such as standard air pressure ±0.05MPa; The actual working air pressure of the negative pressure adsorption unit or positive pressure injection unit in the nozzle array should be collected in real time, and the difference between it and the standard air pressure should be calculated; If the difference is lower than the lower limit of the standard air pressure, but does not exceed the preset deviation threshold, and does not affect the basic execution of a single sorting action, it is determined to be a minor fault that can be automatically compensated. At this time, it can be repaired by automatically adjusting the opening of the air supply valve and recalibrating the zero point of the pressure sensor.

[0088] If the response delay time exceeds the limit, the nozzle solenoid valve response speed test sequence will be activated to identify whether there is mechanical jamming.

[0089] It should be noted that determining whether the response delay time exceeds the limit requires using a preset normal response delay threshold for the nozzle solenoid valve as a benchmark: First, based on equipment performance parameters, sorting accuracy requirements, and normal production data from multiple batches, determine the acceptable upper limit of the response delay time as the preset threshold; Real-time acquisition of the time interval between the nozzle solenoid valve receiving the sorting command and the actual start action, i.e., the actual response delay time; Compare this actual value with the preset threshold. If the actual response delay time exceeds the preset acceptable upper limit, and after excluding non-mechanical factors such as signal transmission delay, it still meets the exceeding condition, then it is determined that the response delay time exceeds the limit, and a response speed test sequence is initiated to investigate mechanical jamming issues.

[0090] If the airflow stability is abnormal, switch to the backup air path and record the main path fault code.

[0091] It should be noted that determining abnormal airflow stability requires using a preset airflow stability threshold as a benchmark. The specific method is as follows: First, based on equipment performance parameters, sorting accuracy requirements, and normal production data from multiple batches, determine the acceptable range of airflow velocity fluctuation as the preset threshold; then, collect the actual airflow velocity data of the positive pressure injection unit or negative pressure adsorption unit of the nozzle array in real time using airflow sensors, continuously monitor and calculate the velocity fluctuation amplitude per unit time; if the fluctuation amplitude exceeds the preset acceptable threshold, and remains abnormal even after excluding external interference factors such as air source pressure fluctuations, it is determined to be an abnormal airflow stability, and the operation of switching to the backup air path channel and recording the fault code of the main channel is immediately executed.

[0092] After completing automatic compensation or channel switching, perform a simulated sorting action, collect the output status of the actuator, and compare it with the standard response curve.

[0093] It should be noted that the comparison must first clarify that the standard response curve is generated based on the optimal sorting data under normal operating conditions of the equipment, and includes a standard time-series curve of key parameters such as air pressure change, response time, and airflow velocity. When performing simulated sorting, the output status data of the actuator is collected in real time to form the actual response curve. The two curves are aligned along the same time axis, and the deviation values ​​of the corresponding parameters are compared point by point. If the deviation of all parameters is within the preset allowable range, the comparison is deemed qualified, the fault alarm is cleared, and sorting is resumed. If there is a deviation that exceeds the allowable range, the comparison is deemed unqualified, sorting is suspended, and a maintenance request is reported.

[0094] If the simulation results meet the accuracy requirements, clear the fault alarm and resume the sorting operation.

[0095] If deviations still exist, the sorting process will be paused and a maintenance request will be submitted. At the same time, the self-inspection log will be stored in the equipment health record for subsequent predictive maintenance analysis.

[0096] It should be noted that this step embeds an intelligent, hierarchical fault diagnosis and self-repair mechanism into the high-precision nozzle array actuator, giving it basic self-healing capabilities. This elevates equipment management from simple passive alarms to proactive intervention, enabling online identification, assessment, and attempts to resolve common performance drift issues. The technical effects are significant, transforming equipment maintenance from reactive, post-event response to proactive intervention and pre-event intelligent early warning. Through online automatic compensation and repair of faults, production interruptions caused by actuator performance drift or minor malfunctions are reduced, ensuring the continuity and stability of the sorting process.

[0097] Furthermore, this self-diagnostic and verification capability ensures that the accuracy of the sorting process remains at its optimal level, avoiding quality issues caused by execution deviations. More importantly, by meticulously recording the fault handling process and results and storing them in the equipment health record, the foundation is laid for predictive maintenance based on big data analysis. This upgrades equipment management from traditional periodic maintenance to a more scientific and economical condition-based maintenance and predictive maintenance model, thereby comprehensively improving the reliability and overall operational efficiency of the entire intelligent sorting system.

[0098] The IoT platform collects process quality data in real time for subsequent combing and spinning processes. The process quality data includes yield, breakage rate and yarn evenness. Based on the process quality data, the sorting threshold parameters and raw material ratio strategy are dynamically adjusted.

[0099] In a preferred embodiment of the present invention, the specific method of dynamically adjusting the sorting threshold parameter and the raw material ratio strategy is as follows: analyzing the process quality data from the carding and spinning processes to identify the key influencing factors related to the quality of the raw materials.

[0100] It should be explained that the key influencing factors are core indicators directly related to the quality of cashmere raw materials and that dominate the quality performance of downstream production, extracted from the process quality data of the carding and spinning processes. These factors originate from the multidimensional characteristics of the raw materials, such as scale integrity and fat content affecting breakage rate, fiber surface smoothness related to yarn evenness, and protein content and moisture regain affecting yield. They are the core basis for locating the root cause of quality abnormalities and adjusting sorting thresholds and raw material ratios.

[0101] Furthermore, from the process quality data of the combing and spinning processes collected by the IoT platform, abnormal indicators that exceed the preset qualified range are screened out. Through statistical methods such as correlation analysis and variance analysis, process fluctuation items strongly correlated with the abnormal indicators are extracted. At the same time, combined with historical production data, the quality data differences between abnormal batches and qualified batches are compared to identify the core process correlation items that play a leading role in quality abnormalities, which are the key influencing factors, such as insufficient fiber cohesion and fiber agglomeration, which are strongly correlated with the breakage rate.

[0102] By cross-referencing key influencing factors with the multidimensional feature vector outputs in historical sorting records, the characteristic dimensions of raw materials that cause quality abnormalities can be located.

[0103] Furthermore, the multidimensional feature vector of the abnormal batch of cashmere is retrieved, including scale structure, surface smoothness, and component spectrum-related feature parameters. The identified key influencing factors are cross-compared with this multidimensional feature vector. Combining the correlation logic between the quality characteristics of cashmere raw materials and downstream processes, such as scale integrity affecting cohesion and fat content affecting fiber agglomeration, the raw material feature dimensions directly corresponding to the key influencing factors are located, and finally the specific feature items that cause quality abnormalities are determined.

[0104] For example, after the raw fiber of the target batch GR-202405 was processed, the IoT platform collected process quality data from the spinning process, showing an average breakage rate as high as 32 times / thousand spindles·hour, exceeding the target threshold of 25 times / thousand spindles·hour. The system triggered a dynamic adjustment process. Through cross-comparison analysis, it was found that the main reason for the high breakage rate was that some fibers with low scale integrity but high fat content were mixed in with the raw material. These fibers were incorrectly classified as grade 2 instead of grade 3 during sorting.

[0105] It should be noted that through this data mining, the system can accurately pinpoint which one or more characteristic dimensions of the raw material are causing the downstream quality abnormalities, such as insufficient scale integrity or excessive fat content.

[0106] Based on a multi-objective optimization algorithm, the sorting boundary thresholds between each quality grade are recalculated to make the sorting results more suitable for the process window of the downstream process.

[0107] It needs to be explained that the sorting boundary threshold is a quantitative judgment standard for classifying cashmere raw materials into different quality grades. It is a critical value of characteristic parameters determined by a multi-objective optimization algorithm based on the joint process window of downstream carding and spinning processes, such as the upper limit of fiber feed amount in the carding machine and the draft ratio range of the spinning machine. It uses multi-dimensional feature vectors, including features such as scale structure, surface smoothness, and component spectrum, to clearly define the boundaries between different quality grades. For example, it sets a scale integrity index of 0.6 and a fat content of 0.4 as critical values, classifying those with characteristic parameters higher than these values ​​as Grade 1 and those lower as Grade 2.

[0108] Based on the optimized sorting boundary threshold, a new raw material ratio scheme is generated, which specifies the mixing ratio of raw materials of different quality grades in the blending process.

[0109] The updated sorting boundary thresholds and raw material proportioning schemes are simultaneously sent to the sorting control system and the batching scheduling system to ensure the consistency and executability of closed-loop control commands.

[0110] It should be noted that this step establishes a dynamic closed-loop control system that traces and optimizes the quality of downstream final products back to upstream raw material sorting standards. It treats the entire production process as an interconnected organic whole. Its technical effect lies in achieving a profound transformation in cashmere processing quality control, shifting from a previous open-loop model to an intelligent closed-loop model. This method ensures that sorting standards are no longer static but can self-correct based on actual production results, thereby enhancing the production line's adaptability to fluctuations in raw materials and changes in operating conditions.

[0111] In a preferred embodiment of the present invention, the specific optimization method of the sorting boundary threshold is as follows: obtain the upper limit of fiber feed amount, the adjustment range of combing intensity and the fiber damage tolerance threshold of the current combing machine from the combing process control system.

[0112] The feasible range of draft ratio, twist control accuracy, and breakage rate warning threshold of the spinning machine are obtained from the spinning process control system.

[0113] The above parameters are transformed into constraints on the quality characteristics of raw materials, forming a joint process window for downstream processes.

[0114] In the multi-objective optimization model, the sorting boundary threshold is used as the decision variable, and the optimization objectives are to maximize the production rate, minimize the breakage rate, and minimize the variance of strip uniformity.

[0115] During the solution process, the constraint condition is that the feature vector of the sorted raw material set falls within the window of the combined process and is not less than a preset percentage threshold.

[0116] Constraint processing techniques are used to penalize solutions that violate the process window, guiding the algorithm to converge to a feasible solution that satisfies both the quality objective and the equipment capability, and finally outputting a sorting boundary threshold that is adapted to the current production line status.

[0117] It's important to note that the principle behind this step lies in transforming raw material sorting decisions from an isolated quality rating process into a systems engineering project deeply embedded in the entire production chain and strictly constrained by the actual downstream processing capabilities. By establishing a joint process window that includes the physical limitations of downstream carding and spinning equipment, the sorting problem is transformed into a constrained multi-objective optimization problem. Essentially, this establishes a rigid mathematical constraint link between raw material sorting and downstream processing. The technical effect is unprecedented production synergy and predictability, fundamentally avoiding the predicament of a mismatch between the sorted raw material grade and the downstream equipment's processing capacity.

[0118] The above content is merely an example and illustration of the concept of the present invention. Those skilled in the art can make various modifications or additions to the specific embodiments described, or use similar methods to replace them, as long as they do not deviate from the concept of the invention or exceed the scope defined by the present invention, and all such modifications and additions should fall within the protection scope of the present invention.

Claims

1. A method for intelligent sorting and processing quality control of cashmere raw materials, characterized in that: include: Acquire multidimensional sensing data of the cashmere raw material to be processed, including fiber scale structure image, fiber surface gloss image and fiber composition spectral data; The multidimensional sensing data is subjected to feature-level fusion processing to generate quantitative feature parameters related to cashmere quality; The quantified feature parameters are input into a hierarchical decision model trained based on machine learning to generate sorting instructions corresponding to different quality levels. According to the sorting instructions, the nozzle array actuator, which is composed of a negative pressure adsorption unit and a positive pressure spraying unit, is controlled to perform non-contact automatic sorting of cashmere raw materials, so that cashmere raw materials of different quality grades fall into the corresponding collection channels respectively. The IoT platform collects process quality data in real time for subsequent combing and spinning processes. The process quality data includes yield, breakage rate and yarn evenness. Based on the process quality data, the sorting threshold parameters and raw material ratio strategy are dynamically adjusted.

2. The intelligent sorting and processing quality control method for cashmere raw materials according to claim 1, characterized in that: The specific method for generating the quantized feature parameters is as follows: Edge enhancement and texture analysis were performed on images acquired using a polarized light imaging device to extract scale structure features, including scale density, consistency of arrangement direction, and integrity index. The surface smoothness index is obtained by statistically analyzing the grayscale distribution and calculating the reflectance of the gloss images acquired using the imaging device. Baseline correction and principal component analysis were performed on the spectral data acquired using a near-infrared spectrometer to extract spectral feature vectors related to protein content, lipid content, and moisture regain. After normalizing the above-mentioned scale structure features, surface smoothness indicators and spectral feature vectors, a weighted fusion algorithm is used to generate multidimensional feature vectors.

3. The intelligent sorting and processing quality control method for cashmere raw materials according to claim 2, characterized in that: The multi-dimensional feature vector is generated using a weighted fusion algorithm, where the weight coefficients of each feature dimension are dynamically determined through a periodic multi-objective optimization mechanism, as follows: Define an optimization objective function, which includes three sub-objectives: maximizing sorting accuracy, maximizing downstream production yield, and minimizing breakage rate. Based on the correlation between multidimensional feature vectors and corresponding process quality data in historical sorting batches, an input-output mapping database for the objective function is constructed. A non-dominated sorting genetic algorithm is used to iteratively search the combination of weight coefficients to find the Pareto front solution set that optimizes the overall performance of the objective function. Select the combination of weight coefficients that satisfy the preset constraints from the Pareto front solution set as the fusion weight for the current period; The selected weighting coefficients are applied to the weighted fusion process of the normalized scale structure features, surface smoothness index and spectral feature vector to generate a multi-dimensional feature vector for hierarchical decision-making. The constraints include the sum of all weight coefficients being 1 and the value range of each weight coefficient being limited to a preset threshold.

4. The intelligent sorting and processing quality control method for cashmere raw materials according to claim 1, characterized in that: The hierarchical decision-making model is implemented as follows: After each sorting operation is completed, record the correspondence between the multidimensional feature vector output of the current batch of cashmere raw materials and the actual sorting action; After receiving process quality data from subsequent processes, the process quality data is associated and mapped with the multidimensional feature vector of the current batch. If the process quality data exceeds the preset acceptable range, the model retraining process is triggered, and the multidimensional feature vector of the current batch and its corresponding process quality deviation are used as new training samples. A sliding window mechanism is used to maintain the training sample set, retaining the data of the most recent N valid batches and removing historical samples that are outside the time window. The weight parameters of the hierarchical decision model are fine-tuned using the updated training sample set to generate an optimized hierarchical decision model suitable for the next cycle of sorting tasks.

5. The intelligent sorting and processing quality control method for cashmere raw materials according to claim 4, characterized in that: The specific method for determining the valid batch is as follows: After each sorting operation is completed, check whether the nozzle array actuator has completed all the command actions and returned a confirmation signal. If any action confirmation is missing, the batch is marked as invalid. Simultaneously evaluate the signal-to-noise ratio and feature extraction success rate of multidimensional sensing data. If the feature loss rate of any modality data exceeds the preset threshold or the image blur is higher than the preset threshold, the batch of sensing data is deemed unqualified. When receiving subsequent process quality data uploaded by the IoT platform, verify whether its timestamp is within the preset time window after sorting is completed, and whether the data fluctuation range does not exceed the normal operating range of the equipment; otherwise, it is considered unreliable feedback. The batch of data is included in the effective sample set of the sliding window only if all three validations pass; The sliding window retains the data from the most recent N valid batches, where N is a preset integer.

6. The intelligent sorting and processing quality control method for cashmere raw materials according to claim 1, characterized in that: The specific method for non-contact automatic sorting of cashmere raw materials is as follows: The real-time position of the cashmere raw material on the conveying path is determined according to the sorting instructions, and the collection channel position corresponding to the target quality grade is matched. The negative pressure adsorption unit is activated simultaneously to form a local negative pressure area above the cashmere raw material conveying path, causing the target fiber bundle to detach from the main material flow; After the negative pressure adsorption action is completed, the positive pressure jet unit corresponding to the landing point is immediately activated, and the separated fiber bundle is pushed to the designated collection channel by directional airflow. Monitor the working status of each unit in the nozzle array, including air pressure, response delay time and airflow stability. If the deviation of any parameter is greater than the preset threshold, the actuator self-test program is triggered. After completing a single sorting operation, the nozzle array is cleaned and purged.

7. The intelligent sorting and processing quality control method for cashmere raw materials according to claim 6, characterized in that: The specific method of the self-inspection procedure for the actuator is as follows: Read the abnormal parameter type and deviation level when the deviation is greater than a preset threshold, and determine whether it belongs to a minor fault that can be automatically compensated; If the air pressure value is abnormal, the opening of the air supply valve will be automatically adjusted and the zero point of the pressure sensor will be recalibrated. If the response delay time exceeds the limit, the nozzle solenoid valve response speed test sequence will be activated to identify whether there is mechanical jamming. If the airflow stability is abnormal, switch to the backup air path and record the main path fault code; After completing automatic compensation or channel switching, perform a simulated sorting action, collect the output status of the actuator and compare it with the standard response curve; If the simulation results meet the accuracy requirements, clear the fault alarm and resume the sorting operation; If deviations still exist, the sorting process will be paused and a maintenance request will be submitted. At the same time, the self-inspection log will be stored in the equipment health record for subsequent predictive maintenance analysis.

8. The intelligent sorting and processing quality control method for cashmere raw materials according to claim 1, characterized in that: The specific method for dynamically adjusting the sorting threshold parameter and raw material ratio is as follows: Analyze the process quality data from the carding and spinning processes to identify key influencing factors related to raw material quality; By cross-comparing key influencing factors with the multi-dimensional feature vector outputs in historical sorting records, the raw material feature dimensions that cause quality abnormalities can be located. Based on a multi-objective optimization algorithm, the sorting boundary thresholds between each quality grade are recalculated to make the sorting results more compatible with the process window of the downstream process. Based on the optimized sorting boundary threshold, a new raw material ratio scheme is generated, which specifies the mixing ratio of raw materials of different quality grades in the blending process; The updated sorting boundary thresholds and raw material ratio schemes will be simultaneously sent to the sorting control system and the batching scheduling system.

9. The intelligent sorting and processing quality control method for cashmere raw materials according to claim 8, characterized in that: The specific optimization method for the sorting boundary threshold is as follows: The upper limit of fiber feed, the range of carding intensity adjustment, and the fiber damage tolerance threshold of the current carding machine are obtained from the carding process control system. The feasible range of draft ratio, twist control accuracy, and breakage rate warning threshold of the spinning frame are obtained from the spinning process control system. The above parameters are transformed into constraints on the quality characteristics of raw materials, forming a joint process window for downstream processes. In the multi-objective optimization model, the sorting boundary threshold is used as the decision variable, and the optimization objectives are to maximize the production rate, minimize the breakage rate, and minimize the variance of strip uniformity. During the solution process, the constraint condition is that the feature vector of the sorted raw material set falls within the window of the combined process and is not less than a preset percentage threshold. Constraint processing techniques are used to penalize solutions that violate the process window, guiding the algorithm to converge to a feasible solution that satisfies both the quality objective and the equipment capability, and finally outputting a sorting boundary threshold that is adapted to the current production line status.

10. The intelligent sorting and processing quality control method for cashmere raw materials according to claim 1, characterized in that: Before acquiring the fiber composition spectral data, a small-sample online calibration process is required, the specific method of which is as follows: Before each sorting task is started, no less than three reference samples with known protein content, fat content and moisture regain are retrieved from the standard cashmere sample library. Each reference sample is sequentially sent into the detection area of ​​the near-infrared spectrometer to acquire the corresponding raw spectral signals; The original spectral signal is subjected to noise filtering and band alignment to generate a calibration spectral dataset; Based on the physicochemical parameter labels of the calibration spectral dataset and the corresponding benchmark samples, a local linear regression mapping model is constructed. The local linear regression mapping model is embedded into the spectral preprocessing module of the cashmere raw material to be tested, so as to correct the collected spectral data in real time. The standard cashmere sample library is regularly updated by offline laboratory testing results.

Images