Intelligent regulation and control system and method for compact siro fiber processing and medium

Abstract

The invention relates to an intelligent regulation and control system and method for compact siro fiber processing and a medium, and relates to the technical field of fiber processing regulation and control, and the intelligent regulation and control system comprises a data collection module which is used for collecting equipment operation parameters, workshop environment parameters and spinning product parameters in the compact siro fiber processing process; the data processing module is used for cleaning and associating equipment operation parameters, workshop environment parameters and spinning product parameters, and constructing a spinning quality parameter sequence; the data analysis module is used for constructing an abnormal quality judgment model according to historical spinning forming quality data in combination with a deep learning algorithm, monitoring a spinning quality parameter sequence, and identifying and positioning abnormal dimension parameters; the strategy matching module is used for analyzing and tracing the abnormal dimension parameters, determining a fault process section and matching a corresponding abnormal regulation and control rule; and the process regulation and control module is used for generating an exception correction instruction sequence according to the exception regulation and control rule, and distributing and controlling corresponding equipment in combination with the exception level.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of fiber processing regulation, in particular to an intelligent regulation system and method for compact siro fiber processing and a medium. BACKGROUND

[0002] Compact siro spinning is a new spinning process integrating siro spinning and compact spinning technology. Through the synergistic effect of double roving feeding and mesh ring gathering device, the fiber arrangement is compact, the hairiness is significantly reduced, and the yarn strength and fabric quality are effectively improved.

[0003] In the compact siro fiber processing process, the yarn is mainly manually sampled, and then the yarn feeding speed, negative pressure airflow and temperature and humidity environment parameters during processing are regulated according to the sampling results.

[0004] The existing patent discloses an intelligent spinning workshop intelligent management system, which comprises an information monitoring system, an information processing system and an information regulation system. The data of the entire spinning workshop is monitored in real time through the information monitoring system, then the data in the spinning workshop is processed through the information processing system to obtain the regulation parameters of the intelligent spinning workshop, and finally the operation of the equipment in the intelligent spinning workshop is regulated through the information regulation system. The intelligent spinning workshop intelligent management system provided by the above-mentioned application realizes the automatic production and transportation control from raw materials to finished products, and the online monitoring and control of product quality and yield through information network technology, fusion of sensor technology and frequency conversion speed regulation, etc., significantly improves product quality, production efficiency, reduces production cost, and has the advantages of modernization, scientization and standardization.

[0005] The existing technical solution in the above has the following defects: 1. The existing technology cannot regulate in real time according to the quality change of yarn forming, resulting in great randomness and hysteresis of the regulation of these parameters, uneven yarn forming quality and low stability. SUMMARY

[0006] In view of the deficiencies of the prior art, the purpose of the present application is to provide an intelligent regulation system, method and medium for compact siro fiber processing, which realizes the instant regulation of yarn feeding speed, negative pressure airflow, temperature and humidity and other parameters by real-time monitoring of yarn quality data and combining machine learning, solves the problem of manual sampling lag, and improves the stability and consistency of yarn forming quality.

[0007] The following technical solutions are adopted: In a first aspect, the application provides an intelligent regulation system for compact siro fiber processing, comprising: a data acquisition module for acquiring equipment operation parameters, workshop environment parameters and spinning product parameters in the compact siro fiber processing process; The data processing module is used to construct a sequence of spinning quality parameters by combining the operating parameters of cleaning and related equipment, workshop environmental parameters, and spinning product parameters. The data analysis module is used to construct an abnormal quality judgment model based on historical spinning and forming quality data and deep learning algorithms, and to monitor the spinning quality parameter sequence to identify and locate abnormal dimensional parameters. The strategy matching module is used to parse and trace the abnormal dimension parameters, determine the faulty process segment, and match the corresponding abnormal control rules. The process control module is used to generate anomaly correction instruction sequences based on anomaly control rules, and to allocate and control the corresponding equipment in combination with the anomaly level.

[0008] By adopting the above technical solution, the data acquisition module acquires multi-source production parameters in real time, and the data processing module cleans and correlates them to construct a quality sequence; the data analysis module uses deep learning algorithms such as LSTM to establish an anomaly detection model to achieve intelligent identification and location of parameter anomalies; the strategy matching module determines the faulty process segment based on the rule engine and causal tracing; and the process control module generates control instructions according to the anomaly level and drives the equipment to correct in real time, realizing full-chain, adaptive intelligent control of spinning quality, which significantly improves production stability, quality consistency and control efficiency.

[0009] Furthermore: the data acquisition module includes: The equipment parameter acquisition unit is used to detect the operating status of special equipment during the processing of compact Siro fibers and obtain the equipment operating parameters; The environmental parameter acquisition unit is used to monitor the environment of the processing and production workshop of compact Siro fiber and obtain the workshop environmental parameters; The process parameter acquisition unit is used to perform multi-dimensional detection of intermediate products in the processing of compact Siro fiber according to the spinning process, and obtain the spinning product parameters.

[0010] By adopting the above technical solution, deploying sensor networks and edge computing nodes, real-time collection of multi-source parameters of equipment, environment and process is achieved. By using data fusion and time series alignment algorithms to construct a unified data stream, real-time data collection of all elements and high synchronization is realized, which significantly improves the accuracy, completeness and timeliness of the data.

[0011] Furthermore, the data processing module includes: The parameter denoising unit is used to filter and denoise the product measurement data in the equipment operating parameters, workshop environment parameters and spinning product parameters according to the data type, so as to obtain noise-free operating parameters, noise-free environmental parameters and noise-free measurement data. The correction and denoising unit is used to perform distortion correction and filtering denoising on the product image data in the spinning product parameters to obtain a noise-free product image. The edge detection unit is used to detect edges in the noise-free product image, extract and mark the outline of the spun product; The target segmentation unit is used to perform semantic segmentation on the noiseless product image based on the outline of the spinning product to obtain the spinning product image. The product measurement unit is used to measure and convert images of spun yarn products to obtain product morphology data. The verification and filtering unit is used to perform morphological node verification on the noiseless measurement data based on the collection time period and product morphological data, and to filter and mark dissimilar node data. The parameter association unit is used to perform time-series association of noiseless operating parameters, noiseless environmental parameters, and dissimilar node data based on the spinning process to construct a spinning quality parameter sequence.

[0012] By adopting the above technical solution, multi-source parameters and images are preprocessed using algorithms such as filtering and noise reduction and distortion correction. Then, edge detection and semantic segmentation algorithms are used to extract product contour and morphological data. After morphological node verification and time series correlation algorithms are used for fusion processing, a high-quality and strongly correlated spinning quality parameter sequence is constructed. This achieves efficient and automated conversion from raw data to analysis-ready data, significantly improving the accuracy, consistency and interpretability of the data.

[0013] Furthermore, the data analysis module includes: The data receiving unit is used to receive the spinning quality parameter sequence in blocks according to the spinning cycle to obtain spinning process parameter blocks. The data collection unit is used to divide historical spinning and forming quality data according to data status labels to obtain training and validation sets; The model building unit is used to iteratively train the training set according to the deep learning algorithm, and update and optimize it in combination with the validation set to generate an abnormal quality judgment model. The data judgment unit is used to inspect and judge the parameter blocks of the spinning process according to the abnormal quality judgment model, and to determine the abnormal dimension parameters.

[0014] By adopting the above technical solution, processing parameters through time series block and data partitioning algorithms, and using deep learning algorithms to build and optimize the abnormal quality judgment model, real-time intelligent inspection and abnormal dimension identification of spinning process parameters are realized, which significantly improves the real-time performance, accuracy and early warning capability of quality monitoring.

[0015] This application further specifies that the model building unit includes: The feature extraction layer is used to extract features from the device operating parameters in the training set based on the first extraction branch using three 3*3 convolutional blocks and two average pooling blocks, to obtain the device operating feature matrix. Based on the second extraction branch, feature extraction is performed on the spinning product parameters in the training set using one 1*3 convolutional block, one 3*1 convolutional block, and one average pooling block to obtain the intermediate product feature matrix; Based on the third extraction branch, global features are extracted from the training set using 2 upsampling blocks, 2 downsampling blocks, and 1 max pooling block to obtain the spinning global feature matrix; The coupled training layer is used to perform coupled operations on the device operation feature matrix and intermediate product feature matrix using DCN network blocks based on the cross-coupled branches, so as to obtain the device product coupled feature matrix. Based on the first training branch, the coupling feature matrix of the equipment product is trained several times using two residual blocks, one Darknet block and the ReLU activation function to obtain the weight matrix of the equipment product. Based on the second training branch, the global feature matrix of spinning is trained several times using 3 inverted residual blocks, 2 CSP blocks, 2 SPP blocks and the softmax activation function to obtain the global feature weight matrix; The identification and verification layer is used to identify the verification set based on the first identification branch using two fully connected blocks and the cross-entropy loss function, combined with the global feature weight matrix, output abnormal quality labels, and match and verify them with the data state labels, and calculate the label matching degree. If the label matching degree is greater than the preset abnormal matching threshold, the validation set is re-identified according to the second identification branch using the Detect block combined with the device product weight matrix, the final abnormal label is output, and it is matched again with the data status label to locate and filter unidentified abnormal data and generate parameter correction coefficients. No, then according to the third identification branch, the global feature weight matrix and the device product weight matrix are extended and trained using FPN blocks, PAN blocks and spatiotemporal attention mechanism to generate a comprehensive weight matrix, and the validation set is extended to identify, locate and filter unidentified abnormal data, and generate identification correction coefficients. The update optimization layer is used to correct the preset iterative constraint parameters based on the parameter correction coefficients of the first optimization branch, obtain the corrected constraint parameters, and then train and match the global feature weight matrix and the equipment product weight matrix again until all abnormal data are identified and located, and determine that the current global feature weight matrix and the current equipment product weight matrix form an abnormal quality judgment model. The second optimization branch uses the identification correction coefficient to correct the cross-entropy loss function, resulting in the cross-entropy correction function. The comprehensive weight matrix is ​​then trained and matched again until all abnormal data is identified and located, and the current comprehensive weight matrix is ​​determined to be an abnormal quality judgment model.

[0016] By adopting the above technical solution, device, product, and global features are extracted through multi-branch convolutional networks, and features are coupled through DCN networks. Multiple rounds of iterative training are conducted using residual blocks, Darknet, CSP, and SPP structures. In the validation phase, anomaly identification is performed by combining fully connected layers and Detect blocks with cross-entropy loss. FPN / PAN and spatiotemporal attention mechanisms are introduced to optimize unidentified samples. Finally, the model is dynamically optimized based on parameter correction coefficients and loss function corrections, constructing a high-precision and robust anomaly quality judgment model. This achieves deep feature fusion and adaptive learning for complex production data, significantly improving the accuracy, generalization ability, and self-optimization efficiency of anomaly identification.

[0017] Furthermore, the strategy matching module includes: The anomaly analysis unit is used to deconstruct the anomaly dimension parameters based on the spinning process to determine the anomaly data source; The timing traceability unit is used to trace abnormal data sources based on timestamps, determine faulty process segments, and extract abnormal process timings. The anomaly location unit is used to locate and periodically extract the monitored spinning quality parameter sequence based on the abnormal process sequence to obtain the abnormal data sequence. The level determination unit is used to determine the danger level of abnormal data sequences based on the data source weight and generate a strategy matching priority. The strategy matching unit is used to perform strategy matching on abnormal data sequences based on strategy matching priority and a preset abnormal strategy library to determine abnormal control rules.

[0018] By adopting the above technical solution, the faulty process section is located through anomaly analysis and time-series tracing algorithms, the anomaly level and strategy priority are determined by risk assessment algorithms, and finally the optimal control rules are matched by the rule engine. This achieves rapid and accurate fault location and intelligent strategy recommendation, significantly improving the automation level and processing efficiency of anomaly response.

[0019] Furthermore, the rating determination unit includes: The data normalization layer is used to normalize the abnormal data sequence to obtain standard abnormal data; The weighted fusion layer is used to perform weighted calculations on standard outlier data of different dimensions based on preset data source weights to obtain an initial comprehensive risk value. The primary judgment layer is used to compare and judge the initial comprehensive risk value based on the preset static risk threshold to determine the initial hazard level; The secondary decision layer is used to perform time-domain analysis on standard outlier data, extracting slope features and duration. If the slope characteristic is in an upward trend and the duration reaches the preset time limit, the initial hazard level is upgraded and corrected to obtain the corrected hazard level. If the slope characteristic is decreasing and the duration reaches the time limit, the initial hazard level is downgraded to obtain the revised hazard level. No, then the current initial risk level will remain unchanged; The causal determination layer is used to establish causal relationships between standard abnormal data, calculate the failure rate, and compare it with a preset failure threshold. If the failure rate exceeds the failure threshold, the initial danger level is raised to the highest level. The dynamic sorting unit is used to calculate and sort the danger execution score based on the danger level, time decay function and policy matching degree, and generate policy matching priority.

[0020] By adopting the above technical solution, an initial comprehensive risk value is calculated through normalization algorithm and weighted fusion, and a preliminary judgment is made based on static threshold. Subsequently, the risk level is dynamically corrected using time-domain feature analysis (such as slope and duration), and a secondary calibration is performed by combining causal correlation analysis to calculate the failure rate. Finally, the risk execution score is dynamically calculated and sorted based on the time decay function and strategy matching degree to generate strategy matching priority. This achieves multi-level, adaptive intelligent risk assessment that integrates temporal and causal relationships, significantly improving the accuracy of risk level judgment, dynamic response capability, and rationality of strategy matching.

[0021] Secondly, this application also provides an intelligent control method for processing compact Siro fibers, employing the following technical solution: A smart control method for processing compact Siro fibers, applied to a smart control system, comprising: Collect equipment operating parameters, workshop environmental parameters, and spinning product parameters during the processing of compact Siro fibers; A sequence of spinning quality parameters is constructed by combining the operating parameters of cleaning and related equipment, workshop environmental parameters, and spinning product parameters. Based on historical spinning quality data and deep learning algorithms, an abnormal quality judgment model is constructed, and the spinning quality parameter sequence is monitored to identify and locate abnormal dimensional parameters. Analyze and trace the abnormal dimension parameters, identify the faulty process section, and match the corresponding abnormal control rules; An anomaly correction instruction sequence is generated based on the anomaly control rules, and the corresponding equipment is allocated and controlled in conjunction with the anomaly level.

[0022] By adopting the above technical solution, a quality sequence is constructed by real-time collection and fusion of multi-source parameters from equipment, environment, and products through sensor networks and data cleaning algorithms. Then, an anomaly judgment model is established using deep learning algorithms such as multi-branch convolutional networks, DCN coupling, and attention mechanisms to achieve accurate identification and location of abnormal parameters. Subsequently, the faulty process segment is determined based on time-series tracing and rule engine and matched with control rules. Finally, adaptive control instructions are generated according to the anomaly level and drive the equipment to correct in real time. This realizes intelligent monitoring, diagnosis, and closed-loop control of the entire spinning process, significantly improving the automation level of quality stability, production efficiency, and anomaly response.

[0023] Thirdly, this application also provides a computer storage medium, which adopts the following technical solution: A computer storage medium having a computer program stored thereon, which, when executed by a processor, implements the method described above.

[0024] In summary, the beneficial technical effects of this application are as follows: By monitoring yarn quality data in real time and combining it with machine learning, the system can instantly control parameters such as yarn feeding speed, negative pressure airflow, temperature and humidity, thus solving the problem of lag in manual sampling and improving the stability and consistency of yarn forming quality. By processing process parameters through time series block and data partitioning algorithms, and using deep learning algorithms to build and optimize anomaly quality judgment models, real-time intelligent inspection and anomaly dimension identification of spinning process parameters are achieved, significantly improving the real-time performance, accuracy and early warning capabilities of quality monitoring. Attached Figure Description

[0025] Figure 1 This is a schematic diagram of the intelligent control system in this application; Figure 2 This is a schematic diagram of the data analysis module in this application; Figure 3 This is a schematic diagram of the structure of the model building unit in this application; Figure 4 This is a schematic diagram of the intelligent control method in this application. Detailed Implementation

[0026] The present application will be further described in detail below with reference to the accompanying drawings.

[0027] Reference Figure 1 This application discloses an intelligent control system for processing compact Siro fibers, comprising: The data acquisition module is used to collect equipment operating parameters, workshop environmental parameters, and spinning product parameters during the processing of compact Siro fibers. The data processing module is used to construct a sequence of spinning quality parameters by combining the operating parameters of cleaning and related equipment, workshop environmental parameters, and spinning product parameters. The data analysis module is used to construct an abnormal quality judgment model based on historical spinning and forming quality data and deep learning algorithms, and to monitor the spinning quality parameter sequence to identify and locate abnormal dimensional parameters. The strategy matching module is used to parse and trace the abnormal dimension parameters, determine the faulty process segment, and match the corresponding abnormal control rules. The process control module is used to generate anomaly correction instruction sequences based on anomaly control rules, and to allocate and control the corresponding equipment in combination with the anomaly level.

[0028] The implementation principle of this embodiment is as follows: multi-source production data is collected in real time through a sensor network, and a quality parameter sequence is constructed based on filtering, image segmentation, and time series correlation algorithms. An anomaly judgment model is established using deep learning algorithms such as multi-branch convolutional networks, DCN coupling, and attention mechanisms. Then, the faulty process segment is located and matched with control rules through a rule engine and causal tracing. Finally, adaptive control instructions are generated based on the anomaly level and the equipment is driven to execute. This realizes full-process automation from data perception and intelligent diagnosis to closed-loop control, which significantly improves the stability, quality consistency, and anomaly response efficiency of the spinning process.

[0029] Preferably, the data acquisition module includes: The equipment parameter acquisition unit is used to detect the operating status of special equipment during the processing of compact Siro fibers and obtain the equipment operating parameters; The environmental parameter acquisition unit is used to monitor the environment of the processing and production workshop of compact Siro fiber and obtain the workshop environmental parameters; The process parameter acquisition unit is used to perform multi-dimensional detection of intermediate products in the processing of compact Siro fiber according to the spinning process, and obtain the spinning product parameters.

[0030] In this embodiment, equipment operating parameters refer to the process settings and real-time status that directly affect fiber forming. Opening / Cardding: Blanket weight, cotton layer thickness, licker-in / cylinder / doffer speed, spacing. Drawing: Draft ratio, roller pressure, sliver can weight, bell mouth diameter. Roving: Spindle speed, draft ratio, twist, winding density. Spinning (core process): Drafting system: Front, middle, and back roller speeds, roller pressure, nip spacing. Twisting and winding system: Spindle speed (spindle speed), ring rail lifting speed and stroke, air ring shape control parameters. Siro spinning specific: Spacing between the two roving feeds, yarn guide hook position, gathering air pressure (if it is compact Siro spinning).

[0031] Workshop environmental parameters: These refer to the macroscopic conditions affecting fiber physical properties and equipment stability. Temperature and humidity: Temperature and relative humidity in each process workshop and key machine areas. Cleanliness: Air dust concentration and short fiber dispersion. Vibration: Vibration intensity of critical equipment (such as the main motor and spindles of the spinning machine).

[0032] Spinning product parameters: These are indicators used to directly evaluate the quality of the final yarn.

[0033] Online / offline inspection indicators: yarn evenness (CV%, -50% fineness, +50% coarseness, +200% neps), hairiness index (H value), breaking strength, twist, and defects (defect grading data per 100,000 meters of yarn). Appearance quality: yarn spots, color difference, and poor forming, detected by the aforementioned image system.

[0034] Equipment operating parameter collection: Step 1: Protocol Integration. Collaborate with the equipment supplier or automation department to obtain the PLC's communication protocol and address mapping table, and determine the memory addresses of the required parameters (such as speed, pressure, and setpoints).

[0035] Step 2: Deploy the data acquisition gateway. Install an industrial gateway in the electrical cabinet of each critical piece of equipment (such as a spinning machine), connect it to the PLC via network cable / fieldbus, and configure the data point table.

[0036] Step 3: Data Extraction and Upload. The gateway reads data from the PLC at a set frequency (e.g., once per second), performs protocol conversion, and uploads it to the data platform via the factory network.

[0037] Workshop environmental parameter collection: Step 1: Site Layout Planning. Based on the workshop layout and airflow organization, scientifically deploy a network of temperature, humidity, and dust sensors in key process areas (such as the front and rear of the spinning machine and air conditioning vents).

[0038] Step 2: Installation and Networking. Fix the sensors and connect them to the IoT gateway via wired or wireless means to form a self-organizing sensor network.

[0039] Step 3: Calibration and Configuration. Perform initial calibration of the sensor, and configure the sampling frequency (e.g., 1 minute / time) and alarm threshold.

[0040] Data collection of spinning product parameters: Step 1: Online Data Integration. Connect the online testing instrument (such as USTER® QUANTUM3) to the factory network, configure its data export interface (such as OPCUA), and stream the real-time yarn indicators to the data platform.

[0041] Step 2: Offline Data Association. Establish a laboratory information management system to manually or automatically import data obtained manually from experimental instruments (such as strength testers and twist testers) into the platform after associating them with production data through work order numbers and batch numbers.

[0042] Step 3: Visual inspection data fusion. The defect categories, locations, quantities, and image features identified by the aforementioned image defect detection system are associated with process parameter data using a unified timestamp and spindle / bobbin number.

[0043] Preferably, the data processing module includes: The parameter denoising unit is used to filter and denoise the product measurement data in the equipment operating parameters, workshop environment parameters and spinning product parameters according to the data type, so as to obtain noise-free operating parameters, noise-free environmental parameters and noise-free measurement data. The correction and denoising unit is used to perform distortion correction and filtering denoising on the product image data in the spinning product parameters to obtain a noise-free product image. The edge detection unit is used to detect edges in the noise-free product image, extract and mark the outline of the spun product; The target segmentation unit is used to perform semantic segmentation on the noiseless product image based on the outline of the spinning product to obtain the spinning product image. The product measurement unit is used to measure and convert images of spun yarn products to obtain product morphology data. The verification and filtering unit is used to perform morphological node verification on the noiseless measurement data based on the collection time period and product morphological data, and to filter and mark dissimilar node data. The parameter association unit is used to perform time-series association of noiseless operating parameters, noiseless environmental parameters, and dissimilar node data based on the spinning process to construct a spinning quality parameter sequence.

[0044] In this embodiment, a unified network time protocol service is deployed for all data acquisition devices in the plant to ensure that the timestamps of PLCs, sensors, cameras, and detectors are highly synchronized (with errors in the millisecond range).

[0045] Each yarn segment and each yarn bobbin is given a unique "digital identity," recording information such as its production start and end times, machine number, spindle number, operator, and the batch of roving used. This is the key that links "parameters" with "product quality."

[0046] Handle missing values ​​caused by communication interruptions and abrupt value jumps caused by sensor malfunctions.

[0047] Use time-series databases to store high-frequency equipment and environmental data, and use relational databases to store structured data such as orders, batches, and quality inspection reports.

[0048] In the data platform, data records from different sources are linked by "timestamps and association keys" to form complete data records that can be analyzed.

[0049] Reference Figure 2 Preferably, the data analysis module includes: The data receiving unit is used to receive the spinning quality parameter sequence in blocks according to the spinning cycle to obtain spinning process parameter blocks. The data collection unit is used to divide historical spinning and forming quality data according to data status labels to obtain training and validation sets; The model building unit is used to iteratively train the training set according to the deep learning algorithm, and update and optimize it in combination with the validation set to generate an abnormal quality judgment model. The data judgment unit is used to inspect and judge the parameter blocks of the spinning process according to the abnormal quality judgment model, and to determine the abnormal dimension parameters.

[0050] In this embodiment, the equipment operation parameter sequence and workshop environment parameter sequence collected in the aforementioned steps are precisely correlated with the final spinning product quality parameters on a timeline.

[0051] Using the previously established "digital identity" and unified timestamp, create a master data table that covers both "cause" (process parameters) and "effect" (quality indicators). For example, associate all process parameters within a specific time period (such as a doffing cycle) with the laboratory test data (such as CV%, strength) of the produced yarn segment.

[0052] If the historical data already has a clear quality level (such as excellent, first-class, qualified, unqualified), it can be directly used as the classification label.

[0053] If no explicit label is provided, SPC methods (such as control charts) can be used to analyze historical quality parameters (such as stripe CV%), marking data points that exceed the control limits as "abnormal" and the rest as "normal".

[0054] The labels must correspond to process data for a specific time period. For example, for yarn labeled "abnormal," all process parameter sequences within the corresponding production period are considered "abnormal samples."

[0055] The continuous process parameter time series is sliced ​​into segments of fixed length (e.g., 30 minutes, 1 hour) or by production unit (e.g., one doffing). Each segment corresponds to a quality label (normal / abnormal).

[0056] Feature Construction: In addition to the original parameters (such as spindle speed and humidity), statistical features within each time window need to be calculated: Time-domain features: mean, variance, kurtosis, skewness, range. Trend features: moving average, linear regression slope. Frequency-domain features (optional): Fourier transform the vibration and rotational speed signals to extract the dominant frequency energy. Dimension Alignment: Ensure that the dimensions of each time window sample are fixed, such as [time step, number of parameters].

[0057] The training, validation, and test sets are divided chronologically (to avoid data leakage caused by random shuffling). For classification models: cross-entropy loss is used. For autoencoders: mean squared error is used as the reconstruction loss.

[0058] Connect to the real-time data pipeline and continuously generate the latest multi-parameter time series segments in the same format as during training using a sliding window. The model outputs a value between 0 and 1, representing the risk score of a quality anomaly occurring in the current time period. An error sequence of the same length as the input sequence is obtained, with high error points indicating the time when the anomaly occurred.

[0059] Determine an optimal threshold on the validation set. Trigger an alarm when the anomaly score exceeds this threshold. Multiple alarm levels (such as warnings and alerts) can also be configured, and false alarms caused by momentary fluctuations can be prevented (e.g., by introducing logic that only alarms are triggered when the score consistently exceeds the threshold).

[0060] Calculate the gradient of the model output (anomaly probability) with respect to each data point in the input sequence. Locations with large absolute gradient values ​​indicate that even small changes at that point can significantly impact the model's judgment, meaning the parameter at that location is exceptionally important. For autoencoders, analyze the time point with the largest reconstruction error and the corresponding parameter dimension. In the input sequence, slightly perturb each parameter and observe the change in the anomaly score. The parameter with the most drastic change is the key anomaly dimension.

[0061] Accuracy, recall, F1 score, and ROC-AUC are calculated on the test set. More importantly, the results are compared with actual quality incident records to verify the accuracy and lead time of the model's alarms.

[0062] Reference Figure 3 Preferably, the model building unit includes: The feature extraction layer is used to extract features from the device operating parameters in the training set based on the first extraction branch using three 3*3 convolutional blocks and two average pooling blocks, to obtain the device operating feature matrix. Based on the second extraction branch, feature extraction is performed on the spinning product parameters in the training set using one 1*3 convolutional block, one 3*1 convolutional block, and one average pooling block to obtain the intermediate product feature matrix; Based on the third extraction branch, global features are extracted from the training set using 2 upsampling blocks, 2 downsampling blocks, and 1 max pooling block to obtain the spinning global feature matrix; The coupled training layer is used to perform coupled operations on the device operation feature matrix and intermediate product feature matrix using DCN network blocks based on the cross-coupled branches, so as to obtain the device product coupled feature matrix. Based on the first training branch, the coupling feature matrix of the equipment product is trained several times using two residual blocks, one Darknet block and the ReLU activation function to obtain the weight matrix of the equipment product. Based on the second training branch, the global feature matrix of spinning is trained several times using 3 inverted residual blocks, 2 CSP blocks, 2 SPP blocks and the softmax activation function to obtain the global feature weight matrix; The identification and verification layer is used to identify the verification set based on the first identification branch using two fully connected blocks and the cross-entropy loss function, combined with the global feature weight matrix, output abnormal quality labels, and match and verify them with the data state labels, and calculate the label matching degree. If the label matching degree is greater than the preset abnormal matching threshold, the validation set is re-identified according to the second identification branch using the Detect block combined with the device product weight matrix, the final abnormal label is output, and it is matched again with the data status label to locate and filter unidentified abnormal data and generate parameter correction coefficients. No, then according to the third identification branch, the global feature weight matrix and the device product weight matrix are extended and trained using FPN blocks, PAN blocks and spatiotemporal attention mechanism to generate a comprehensive weight matrix, and the validation set is extended to identify, locate and filter unidentified abnormal data, and generate identification correction coefficients. The update optimization layer is used to correct the preset iterative constraint parameters based on the parameter correction coefficients of the first optimization branch, obtain the corrected constraint parameters, and then train and match the global feature weight matrix and the equipment product weight matrix again until all abnormal data are identified and located, and determine that the current global feature weight matrix and the current equipment product weight matrix form an abnormal quality judgment model. The second optimization branch uses the identification correction coefficient to correct the cross-entropy loss function, resulting in the cross-entropy correction function. The comprehensive weight matrix is ​​then trained and matched again until all abnormal data is identified and located, and the current comprehensive weight matrix is ​​determined to be an abnormal quality judgment model.

[0063] In this embodiment, the time-series data of equipment operation parameters in the training set includes the operating status parameters of the spinning equipment at various time points, such as spindle speed, draft ratio, and roller pressure. The data is organized in a multi-dimensional time series format. The data is first processed through three 3×3 convolutional blocks, each containing convolution operations, batch normalization, and activation functions to extract local temporal patterns. Then, it passes through two average pooling blocks to downsample the feature map, compressing the temporal dimension while enhancing feature robustness. The equipment operation feature matrix is ​​a compact representation after deep feature extraction, preserving the key temporal patterns of the equipment's operating state. The dimensionality is significantly reduced compared to the original input, but the information density is significantly increased.

[0064] The training set contains spinning product parameter data, including quality indicators such as yarn evenness, hairiness index, and breaking strength. The data is typically organized as two-dimensional or multi-dimensional arrays. The data is processed sequentially through 1×3 and 3×1 convolutional blocks to extract lateral and longitudinal feature relationships, respectively. This asymmetric convolutional design captures feature correlations in different directions. Subsequent average pooling blocks are used for feature compression. The intermediate product feature matrix is ​​a comprehensive representation integrating multi-dimensional quality features, preserving the key spatial distribution characteristics of the product parameters.

[0065] The complete training set data includes integrated data on equipment operating parameters, environmental parameters, and product parameters, covering multimodal information throughout the entire spinning process. The data is alternately processed through upsampling and downsampling blocks twice each. Upsampling expands the feature resolution to capture details, while downsampling compresses features to extract abstract patterns. Finally, max pooling blocks are used to extract the most salient features. The spinning global feature matrix is ​​a highly abstract feature vector encompassing the entire process, integrating the complex interactions between equipment, environment, and products during spinning.

[0066] The equipment operation feature matrix and intermediate product feature matrix are derived from the first and second extraction branches, respectively. The two feature matrices are coupled using a deformable convolutional network (DCN) block. By learning the spatial offset of feature points, the DCN enables the convolutional kernel to adaptively adjust its receptive field, better aligning equipment operation features with product quality features. The equipment-product coupled feature matrix is ​​a feature representation that deeply integrates the correlation between equipment state and product quality, reflecting the complex mapping relationship between "cause" (equipment operation) and "effect" (product quality).

[0067] The device-product coupling feature matrix is ​​derived from the output of the cross-coupling branch. Deep features are learned through two residual blocks, with residual connections mitigating the vanishing gradient problem. Further feature extraction is then performed using a Darknet block, which combines multiple convolutional layers and skip connections. Finally, a ReLU activation function is used to introduce non-linearity. The device-product weight matrix is ​​a set of weight parameters obtained through multiple iterations of training, specifically designed to identify anomalous patterns from the device-product coupling features.

[0068] The global feature matrix for spinning is derived from the output of the third extraction branch. Features are extracted using three inverse residual blocks, which are first increased in dimensionality and then decreased to balance computational efficiency and feature representation capability. Next, two CSP (Cross-Stage Partial Connectivity) blocks enhance feature fusion capabilities. Then, two SPP (Spatial Pyramid Pooling) blocks capture multi-scale features. Finally, a Softmax activation function is used for probability normalization. The global feature weight matrix is ​​a set of weight parameters optimized for global spinning features, capable of identifying abnormal patterns throughout the entire process.

[0069] Validation set data and a global feature weight matrix. The validation set includes data such as device operating parameters and output parameters that were not used in training. The validation set features, after processing the global feature weight matrix, are classified using two fully connected blocks. The cross-entropy loss function is used to measure the difference between the predicted labels and the true labels. The matching degree between the predicted abnormal quality labels and the true state labels is calculated. If the label matching degree is greater than a preset abnormal matching threshold, it indicates that the model's initial recognition effect is good, and it proceeds to the second recognition branch for refined recognition. Otherwise, it indicates that the model's recognition ability is insufficient, and it proceeds to the third recognition branch for extended training.

[0070] The validation set data and the device product weight matrix are combined with the device product weight matrix using a Detect block to re-identify the validation set. The Detect block typically contains multiple detectors, enabling simultaneous target classification and localization. After outputting the final anomaly labels, they are matched again with the true labels to identify incorrectly classified anomalies. The unidentified anomalies are located and filtered, along with the generated parameter correction coefficients. These parameter correction coefficients reflect the direction and magnitude of adjustment required for the current weight matrix.

[0071] The global feature weight matrix and the device output weight matrix are fused and enhanced through FPN (Feature Pyramid Network) and PAN (Path Aggregation Network) blocks combined with a spatiotemporal attention mechanism. FPN achieves top-down feature enhancement, PAN achieves bottom-up feature aggregation, and the spatiotemporal attention mechanism focuses on important features in the time and spatial dimensions.

[0072] The combined weight matrix, along with the unidentified anomaly data identified after expanding and filtering the validation set, generates identification correction coefficients. These coefficients reflect the direction in which the cross-entropy loss function needs adjustment.

[0073] The preset iterative constraint parameters (such as learning rate and regularization coefficient) are adjusted using parameter correction coefficients to obtain the corrected constraint parameters. The global feature weight matrix and the device product weight matrix are then retrained using these corrected constraint parameters, and matched and validated on the validation set. This training and validation process is repeated until all outliers can be correctly identified and located. The optimized global feature weight matrix and the device product weight matrix together constitute the final anomaly quality judgment model.

[0074] The cross-entropy loss function is modified using an identification correction coefficient, resulting in a modified cross-entropy function. This modification may include adjusting class weights, adding regularization terms, or modifying the loss calculation method. The modified loss function is then used to retrain and validate the integrated weight matrix. This training and validation process is repeated until all outliers can be correctly identified and located. The optimized integrated weight matrix serves as the final outlier quality assessment model.

[0075] Preferably, the strategy matching module includes: The anomaly analysis unit is used to deconstruct the anomaly dimension parameters based on the spinning process to determine the anomaly data source; The timing traceability unit is used to trace abnormal data sources based on timestamps, determine faulty process segments, and extract abnormal process timings. The anomaly location unit is used to locate and periodically extract the monitored spinning quality parameter sequence based on the abnormal process sequence to obtain the abnormal data sequence. The level determination unit is used to determine the danger level of abnormal data sequences based on the data source weight and generate a strategy matching priority. The strategy matching unit is used to perform strategy matching on abnormal data sequences based on strategy matching priority and a preset abnormal strategy library to determine abnormal control rules.

[0076] The implementation principle of this embodiment is as follows: based on the process deconstruction algorithm, abnormal parameters are traced and analyzed. The faulty process section is located and the abnormal sequence is extracted using the timestamp tracking and periodic extraction algorithm. Then, the hazard level and strategy matching order are determined by the weighted risk assessment and dynamic priority ranking algorithm. Finally, the optimal control rules are intelligently matched from the strategy library with the help of the rule engine. This realizes full automation and accurate decision-making from anomaly location to strategy recommendation, which significantly improves the response speed of fault handling and the pertinence of control measures.

[0077] Preferably, the grade determination unit includes: The data normalization layer is used to normalize the abnormal data sequence to obtain standard abnormal data; The weighted fusion layer is used to perform weighted calculations on standard outlier data of different dimensions based on preset data source weights to obtain an initial comprehensive risk value. The primary judgment layer is used to compare and judge the initial comprehensive risk value based on the preset static risk threshold to determine the initial hazard level; The secondary decision layer is used to perform time-domain analysis on standard outlier data, extracting slope features and duration. If the slope characteristic is in an upward trend and the duration reaches the preset time limit, the initial hazard level is upgraded and corrected to obtain the corrected hazard level. If the slope characteristic is decreasing and the duration reaches the time limit, the initial hazard level is downgraded to obtain the revised hazard level. No, then the current initial risk level will remain unchanged; The causal determination layer is used to establish causal relationships between standard abnormal data, calculate the failure rate, and compare it with a preset failure threshold. If the failure rate exceeds the failure threshold, the initial danger level is raised to the highest level. The dynamic sorting unit is used to calculate and sort the danger execution score based on the danger level, time decay function and policy matching degree, and generate policy matching priority.

[0078] In this embodiment, abnormal data sequences from different dimensions (equipment operating parameters, spinning product parameters, and workshop environmental parameters) are standardized to eliminate the influence of dimensions. Then, based on a preset data source weight table (e.g., equipment operating parameters weight 0.5, spinning product quality parameters weight 0.3, and workshop environmental parameters weight 0.2), the outliers (e.g., deviations exceeding thresholds) of each parameter sequence are weighted and calculated to obtain the initial comprehensive risk value for each abnormal event.

[0079] The initial comprehensive risk value is compared with preset static risk thresholds (high, medium, and low) to obtain the initial hazard level. The temporal characteristics of the abnormal sequence are analyzed (e.g., rising slope, duration). If the risk value, although not reaching the high-level threshold, shows a rapid upward trend or a prolonged duration, a trend amplification coefficient is triggered, increasing its hazard level. Using causal graphs or knowledge graphs, it is analyzed whether the current anomaly may trigger a chain reaction of failures in related equipment or lead to a serious quality accident. If significant associated risks exist, its hazard level is directly raised to the highest level.

[0080] A base score is assigned based on the assessed hazard level (e.g., high=3, medium=2, low=1). A time decay function is introduced, so that the more recently an anomaly occurred, the higher its urgency score. Considering the status of currently available processing resources (e.g., maintenance teams, debugging engineers), anomaly tasks requiring similar resources are prioritized, with anomalies where resources are ready being processed first. A preliminary matching of a pre-built processing strategy library is performed, prioritizing anomalies with existing, mature, and efficient strategies. Anomalies with unclear strategies or requiring the development of new strategies may have their execution priority appropriately delayed.

[0081] Output a sorted list of exception handling priorities, with each item containing the exception ID, danger level, recommended handling strategy index, priority score, and key judgment criteria.

[0082] The implementation principle of this embodiment is as follows: Real-time multi-source data from equipment, environment, and products are collected via sensor networks and edge computing. Data cleaning and feature extraction are performed using algorithms such as filtering and denoising, distortion correction, edge detection, and semantic segmentation to construct a spinning quality parameter sequence. Then, an abnormal quality judgment model is built and trained using multi-branch convolutional networks, DCN coupling, attention mechanisms, and dynamic optimization algorithms to achieve intelligent identification and location of parameter anomalies. Subsequently, faulty process segments and hazard levels are determined based on process deconstruction, time-series tracing, and weighted risk assessment algorithms. Optimal control rules are matched using a rule engine. Finally, adaptive control instructions are generated based on the anomaly level, driving real-time equipment correction. This forms a closed-loop intelligent control system covering the entire process from perception, diagnosis, decision-making to control, significantly improving production stability, quality consistency, and the automation level of anomaly response.

[0083] Reference Figure 4 A smart control method for processing compact Siro fibers, applied to a smart control system for processing compact Siro fibers, comprising: S1: Collect equipment operating parameters, workshop environmental parameters, and spinning product parameters during the processing of compact Siro fibers; S2: Construct a sequence of spinning quality parameters by combining the operating parameters of cleaning and related equipment, workshop environmental parameters, and spinning product parameters; S3: Based on historical spinning and forming quality data and deep learning algorithms, construct an abnormal quality judgment model, monitor the spinning quality parameter sequence, and identify and locate abnormal dimension parameters; S4: Analyze and trace the abnormal dimension parameters, determine the faulty process section, and match the corresponding abnormal control rules; S5: Generate an anomaly correction instruction sequence based on the anomaly control rules, and combine the anomaly level to allocate and control the corresponding equipment.

[0084] The implementation principle of this embodiment is as follows: multi-source production data is collected in real time through a sensor network, and a quality sequence is constructed through cleaning and correlation algorithms; an anomaly judgment model is established using deep learning algorithms such as multi-branch convolution, DCN coupling and attention mechanism to realize parameter anomaly identification and location; then, control rules are matched and adjusted through time-series tracing and rule engine; finally, adaptive control instructions are generated according to the anomaly level and the equipment is driven to correct in real time, realizing closed-loop intelligent control from perception, diagnosis to control, which significantly improves production stability and quality consistency.

[0085] A computer storage medium having a computer program stored thereon, which, when executed by a processor, implements the method described above.

[0086] The embodiments described in this specific implementation are preferred embodiments of this application and are not intended to limit the scope of protection of this application. Therefore, all equivalent changes made in accordance with the structure, shape and principle of this application should be covered within the scope of protection of this application.

Claims

1. An intelligent control system for processing compact Siro fibers, comprising: The data acquisition module is used to collect equipment operating parameters, workshop environmental parameters, and spinning product parameters during the processing of compact Siro fibers. The data processing module is used to clean and associate the equipment operating parameters, the workshop environment parameters, and the spinning product parameters to construct a spinning quality parameter sequence; The data analysis module is used to construct an abnormal quality judgment model based on historical spinning and forming quality data and deep learning algorithms, and to monitor the spinning quality parameter sequence to identify and locate abnormal dimension parameters. The strategy matching module is used to parse and trace the anomaly dimension parameters, determine the faulty process segment, and match the corresponding anomaly control rules. The process control module is used to generate an anomaly correction instruction sequence according to the anomaly control rules, and to allocate and control the corresponding equipment in combination with the anomaly level.

2. The intelligent control system for processing compact Siro fibers according to claim 1, characterized in that, The data acquisition module includes: The equipment parameter acquisition unit is used to detect the operating status of special equipment during the processing of compact Siro fibers and obtain the equipment operating parameters; The environmental parameter acquisition unit is used to monitor the environment of the processing and production workshop of compact Siro fiber and obtain the workshop environmental parameters; The process parameter acquisition unit is used to perform multi-dimensional detection of intermediate products in the processing of compact Siro fiber according to the spinning process, and obtain the spinning product parameters.

3. The intelligent control system for processing compact Siro fibers according to claim 1, characterized in that, The data processing module includes: The parameter denoising unit is used to filter and denoise the product measurement data in the equipment operating parameters, workshop environment parameters and spinning product parameters according to the data type, so as to obtain noise-free operating parameters, noise-free environmental parameters and noise-free measurement data. The correction and denoising unit is used to perform distortion correction and filtering denoising on the product image data in the spinning product parameters to obtain a noise-free product image. An edge detection unit is used to perform edge detection on the noise-free product image, extract and mark the outline of the spun product; The target segmentation unit is used to perform semantic segmentation on the noise-free product image based on the outline of the spinning product to obtain the spinning product image. The product measurement unit is used to measure and convert the image of the spun product to obtain product morphology data. The verification and filtering unit is used to perform morphological node verification on the noiseless measurement data based on the collection time period and the product morphological data, and to filter and mark dissimilar node data. The parameter association unit is used to perform time-series association of the noiseless operating parameters, the noiseless environment parameters, and the dissimilar node data according to the spinning process, and construct a spinning quality parameter sequence.

4. The intelligent control system for processing compact Siro fibers according to claim 1, characterized in that: The data analysis module includes: The data receiving unit is used to receive the spinning quality parameter sequence in blocks according to the spinning cycle to obtain spinning process parameter blocks. The data collection unit is used to divide historical spinning and forming quality data according to data status labels to obtain training and validation sets; The model building unit is used to iteratively train the training set according to the deep learning algorithm, and update and optimize it in combination with the validation set to generate an abnormal quality judgment model. The data judgment unit is used to inspect and judge the spinning process parameter block according to the abnormal quality judgment model, and determine the abnormal dimension parameters.

5. The intelligent control system for processing compact Siro fibers according to claim 4, characterized in that: The model building unit includes: The feature extraction layer is used to extract features from the device operating parameters in the training set based on the first extraction branch, and obtain the device operating feature matrix. Based on the second extraction branch, feature extraction is performed on the spinning product parameters in the training set to obtain the intermediate product feature matrix; Global feature extraction is performed on the training set according to the third extraction branch to obtain the spinning global feature matrix; The coupled training layer is used to perform coupling operations on the device operation feature matrix and the intermediate product feature matrix based on the cross-coupling branches to obtain the device product coupled feature matrix. The equipment product coupling feature matrix is ​​trained iteratively several times based on the first training branch to obtain the equipment product weight matrix. The global feature matrix of spinning is trained iteratively several times according to the second training branch to obtain the global feature weight matrix.

6. The intelligent control system for processing compact Siro fibers according to claim 5, characterized in that: The model building unit also includes: The identification and verification layer is used to identify the verification set based on the first identification branch using two fully connected blocks and the cross-entropy loss function, combined with the global feature weight matrix, output abnormal quality labels, and match and verify them with the data state labels, and calculate the label matching degree. If the label matching degree is greater than the preset abnormal matching threshold, the verification set is re-identified according to the second identification branch using the Detect block combined with the device product weight matrix, the final abnormal label is output, and it is matched again with the data status label to locate and filter unidentified abnormal data and generate parameter correction coefficients. No, then according to the third identification branch, the global feature weight matrix and the device product weight matrix are extended and trained using FPN blocks, PAN blocks and spatiotemporal attention mechanism to generate a comprehensive weight matrix, and the validation set is extended and identified to locate and filter unidentified abnormal data and generate identification correction coefficients. The update optimization layer is used to correct the preset iterative constraint parameters according to the first optimization branch using the parameter correction coefficient to obtain the corrected constraint parameters, and to train and match the global feature weight matrix and the equipment product weight matrix again until all abnormal data are identified and located, and to determine that the current global feature weight matrix and the current equipment product weight matrix constitute an abnormal quality judgment model. The cross-entropy loss function is corrected using the identification correction coefficient according to the second optimization branch to obtain the cross-entropy correction function. The comprehensive weight matrix is ​​then trained and matched again until all abnormal data is identified and located, and the current comprehensive weight matrix is ​​determined to be the abnormal quality judgment model.

7. The intelligent control system for processing compact Siro fibers according to claim 1, characterized in that: The strategy matching module includes: The anomaly analysis unit is used to deconstruct the anomaly dimension parameters based on the spinning process to determine the anomaly data source; The timing traceability unit is used to trace the abnormal data source according to the timestamp, determine the faulty process segment, and extract the abnormal process timing. An anomaly location unit is used to locate and periodically extract the monitored spinning quality parameter sequence according to the abnormal process sequence to obtain an abnormal data sequence. The level determination unit is used to determine the danger level of the abnormal data sequence based on the data source weight and generate a strategy matching priority. The strategy matching unit is used to perform strategy matching on the abnormal data sequence according to the strategy matching priority and in combination with a preset abnormal strategy library, and to determine the abnormal control rules.

8. The intelligent control system for processing compact Siro fibers according to claim 7, characterized in that: The level determination unit includes: The data normalization layer is used to normalize the abnormal data sequence to obtain standard abnormal data; The weighted fusion layer is used to perform weighted calculations on standard outlier data of different dimensions based on preset data source weights to obtain an initial comprehensive risk value. The primary judgment layer is used to compare and judge the initial comprehensive risk value based on a preset static risk threshold to determine the initial danger level; The secondary decision layer is used to perform time-domain analysis on the standard anomaly data and extract slope features and duration. If the slope feature is increasing and the duration reaches a preset time limit, the initial danger level is upgraded and corrected to obtain a corrected danger level. If the slope feature is decreasing and the duration reaches the time limit value, then the initial hazard level is downgraded to obtain the corrected hazard level. No, then the current initial risk level will remain unchanged; The causal determination layer is used to perform causal correlation on the standard abnormal data, calculate the failure rate, and compare it with a preset failure threshold. If the failure rate is greater than the failure threshold, the initial danger level is raised to the highest level. The dynamic sorting unit is used to calculate and sort the danger execution score based on the danger level, time decay function and policy matching degree, and generate policy matching priority.

9. A smart control method for processing compact Siro fibers, applied to the system as described in any one of claims 1-8, characterized in that, include: Collect equipment operating parameters, workshop environmental parameters, and spinning product parameters during the processing of compact Siro fibers; Clean and associate the equipment operating parameters, the workshop environment parameters, and the spinning product parameters to construct a spinning quality parameter sequence; Based on historical spinning and forming quality data and deep learning algorithms, an abnormal quality judgment model is constructed, and the spinning quality parameter sequence is monitored to identify and locate abnormal dimension parameters. The abnormal dimension parameters are analyzed and traced to determine the faulty process segment and match the corresponding abnormal control rules. An anomaly correction instruction sequence is generated based on the anomaly control rules, and the corresponding devices are allocated and controlled in conjunction with the anomaly level.

10. A computer storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method as described in claim 9.

Images