Method and system for intelligent identification of process abnormalities and quality evaluation of cotton processing unmanned production line

Abstract

The application discloses a kind of process abnormal intelligent identification and quality evaluation method and system of cotton processing unmanned production line, belong to cotton processing intelligent manufacturing field.The method includes: synchronous acquisition production line multi-source heterogeneous data;Align data and extract features using timestamp interpolation method, construct multidimensional feature vector;Characteristic vector is input into integrated intelligent model of self-encoder and CNN-LSTM in parallel, and the recognition result for representing the stability state of process of production line is generated by dynamic threshold fusion;Establish multidimensional quality index system, combine AHP with fuzzy comprehensive evaluation model to calculate real-time comprehensive quality score and determine lint grade;With quality score deviation as feedback quantity, adaptive closed-loop control is carried out.The application realizes early warning of process abnormality, real-time quantitative evaluation of quality and adaptive optimization of production line parameters, and can effectively improve lint grade.

Description

Technical Field

[0001] This application relates to the field of intelligent manufacturing technology, specifically to a method and system for intelligent identification and quality assessment of process anomalies in unmanned cotton processing production lines. Background Technology

[0002] Cotton processing lines are typical continuous production processes with complex process parameters and numerous factors affecting quality. Current technologies primarily employ two methods for identifying process anomalies: manual judgment and sensor threshold alarms. In manual judgment, operators observe the production line's operating status through hearing and sight; by the time an anomaly is detected, a certain amount of quality loss has often already occurred. Sensor monitoring methods mostly use single, fixed threshold alarms, which have limited adaptability to different operating conditions and struggle to accurately distinguish between normal fluctuations and genuine process anomalies.

[0003] In terms of quality inspection, traditional methods employ sampling inspection, which has a long testing cycle and the results are susceptible to the subjective influence of operators. With offline testing, there is a significant lag in the feedback of test results to the production line, making real-time response to quality issues difficult. Furthermore, data collected in cotton processing production lines is typically stored in different systems, making it difficult to effectively utilize the correlation between process status and quality fluctuations.

[0004] In addition, the existing cotton quality evaluation system mainly focuses on spectral indicators such as fiber length and micronaire value, and does not give enough consideration to ginning quality and process stability, and the connection between quality evaluation and process control is not close enough. Summary of the Invention

[0005] To address the aforementioned problems, this invention provides a method and system for intelligent identification and quality assessment of process anomalies in unmanned cotton processing production lines.

[0006] In a first aspect, the present invention provides a method for intelligent identification and quality assessment of process anomalies in unmanned cotton processing production lines, comprising the following steps: S1. Synchronously collect multi-source heterogeneous data of the entire production line process. The multi-source heterogeneous data includes equipment operating condition time series data, machine vision image data, environmental parameters and fiber spectral quality data. S2. Use timestamp interpolation to align the multi-source heterogeneous data, extract time-domain statistics, frequency-domain power spectrum, image texture and LSTM trend features, and concatenate them to construct a multi-dimensional feature vector; S3. Input the multi-dimensional feature vectors into the comprehensive intelligent model in parallel, obtain the reconstruction error and fault probability, and then use dynamic threshold fusion to generate anomaly identification results to characterize the stability state of the production line process. S4. Establish a multi-dimensional quality index system that includes ginning quality, fiber quality and process stability. After determining the weights using AHP, combine it with a fuzzy comprehensive evaluation model. Use the fiber spectral quality data and the anomaly identification results output by S3 as inputs to calculate the real-time comprehensive quality score and determine the cotton grade. S5. Using the deviation between the comprehensive quality score and the target threshold as feedback, the key process parameters are adjusted through adaptive control, and the adjusted process parameters are sent to the production controller to achieve adaptive closed-loop optimization of the production line process parameters.

[0007] As a preferred embodiment of the technical solution of the present invention, step S2 includes: S21. Select the equipment operating condition time series data with the highest sampling frequency as the main time reference, and use the linear interpolation method to resample the timestamps of environmental parameters and fiber spectral quality data to the main time reference to achieve data spatiotemporal alignment. S22. Perform sliding window segmentation on the aligned equipment operating condition time series data, calculate the mean, variance, peak factor and kurtosis in each window to obtain time domain statistical characteristics; at the same time, perform fast Fourier transform on the time domain signal to extract the main frequency energy and spectral entropy in the spectrum to obtain frequency domain power spectrum characteristics. S23. Perform grayscale processing and Gaussian filtering to denoise the machine vision image data, and use the gray-level co-occurrence matrix algorithm to calculate the contrast, energy, entropy and homogeneity index of the image to obtain the image texture feature vector. S24. Arrange the aligned multi-source heterogeneous data into a time sequence according to a preset time window, input it into a pre-trained LSTM encoder, and extract the hidden layer output vector of the last time step as an LSTM trend feature to characterize the dynamic evolution of the production line. S25. The time-domain statistical features, frequency-domain power spectrum features, image texture feature vectors and LSTM trend features are respectively subjected to Z-score standardization. After eliminating the difference in dimensions, they are concatenated according to the channel dimension to obtain a multi-dimensional feature vector.

[0008] By selecting the data with the highest sampling frequency as the primary time reference and using linear interpolation to resample the timestamps of multi-source data, the data alignment problem caused by inconsistent sampling frequencies of different sensors is solved. This comprehensively characterizes the production line's operating status from multiple dimensions, improving the accuracy of anomaly identification.

[0009] As a preferred embodiment of the technical solution of the present invention, the integrated intelligent model includes an autoencoder for unsupervised anomaly detection and a CNN-LSTM model for supervised fault classification. Step S3 includes: S31. The multidimensional feature vector is simultaneously input into the autoencoder and the CNN-LSTM model; wherein, the autoencoder performs encoding-decoding reconstruction on the input multidimensional feature vector and calculates the real-time reconstruction error between the original input vector and the reconstructed output vector; the CNN-LSTM model extracts the local spatial correlation between features through the CNN layer, then captures the temporal dependency through the LSTM layer, and outputs the fault probability vector of various known fault types through the Softmax layer. S32. Obtain the real-time reconstruction error output by the autoencoder and the fault probability vector output by the CNN-LSTM model; S33. Construct an upper limit of the dynamic confidence interval based on the historical reconstruction error. When the real-time reconstruction error exceeds the upper limit, it is judged as an anomaly. The maximum value in the fault probability vector is weighted and fused with the normalized real-time reconstruction error to obtain a comprehensive anomaly score. S34. Based on the comprehensive anomaly score, match the preset warning level to generate a three-level warning; determine the anomaly type by combining the category label corresponding to the maximum value in the fault probability vector, and determine the anomaly location by mapping the current timestamp; calculate the anomaly frequency based on the number of anomalies occurring within the preset time window, and calculate the defect density based on the cumulative number of defect-type anomalies; generate anomaly identification results that include anomaly type, location, three-level warning, anomaly frequency, and defect density.

[0010] The autoencoder detects unknown anomalies through reconstruction errors, while the CNN-LSTM model identifies known typical anomalies through fault classification; both processes are performed in parallel. A dynamic confidence interval is constructed based on historical reconstruction errors, and the reconstruction errors are weighted and fused with fault probabilities to obtain a comprehensive anomaly score. This avoids misjudgment under different operating conditions with fixed thresholds, thus improving the accuracy of anomaly identification.

[0011] As a preferred embodiment of the technical solution of the present invention, the method for constructing the upper limit of the dynamic confidence interval in S33 is as follows: Obtain the historical average reconfiguration error under the current production line operating conditions. with standard deviation ; Set the upper limit of the dynamic confidence interval to ,in This is a preset constant; Weighted fusion yields a comprehensive anomaly score The calculation formula is: ,in This refers to the normalized real-time reconstruction error. The maximum value in the failure probability vector. For dynamic fusion weights.

[0012] The upper limit of the dynamic confidence interval is constructed based on the mean and standard deviation of historical reconstruction errors, so that the anomaly judgment threshold can adaptively follow the changes in production line conditions, avoiding false alarms or missed alarms when fixed thresholds are used when operating conditions fluctuate.

[0013] As a preferred embodiment of the technical solution of the present invention, step S4 includes: S41. Establish a hierarchical model including a target layer, a criterion layer, and an indicator layer; the criterion layer includes fiber quality, rolling quality, and process stability; the indicator layer includes length, micronaire value, strength, and impurity content corresponding to fiber spectral quality data, as well as defect density, abnormal frequency, and warning level corresponding to the abnormal identification results output by S3. S42. Construct pairwise comparison judgment matrices for each level of indicators, calculate the largest eigenvalue and its corresponding eigenvector, obtain the initial weights of each indicator after normalization, and determine them as the final weight vectors after passing the consistency test. S43. Set up a set of evaluation criteria for cotton grades, and use the membership function to convert the real-time collected fiber spectral quality data into membership degrees for each grade. At the same time, based on the anomaly frequency and warning level in the anomaly identification results output by S3, dynamically correct the membership degrees of the ginning quality and process stability dimensions, and generate the corrected fuzzy relation matrix. S44. Use the final weight vector to perform a weighted synthesis operation on the modified fuzzy relation matrix to obtain a comprehensive evaluation vector; determine the current grade of the cotton lint according to the principle of maximum membership, and calculate the real-time comprehensive quality score using the weighted average method.

[0014] A three-tiered indicator system, comprising a target layer, a criterion layer, and an indicator layer, was established, incorporating intelligent extended indicators to achieve quantifiable quality assessment. The weights of each indicator were determined using Analytic Hierarchy Processing (AHP), and the fuzziness and uncertainty of the indicators were addressed using a fuzzy comprehensive evaluation model. The system outputs a comprehensive quality score of 0-100 points and four levels: Excellent, Good, Average, and Poor. The scoring results are correlated with the cotton market pricing mechanism.

[0015] The membership degree of the rolling quality and process stability dimensions is dynamically corrected based on the frequency of anomalies and the warning level, so that the quality assessment can perceive the abnormal status of the production line in real time and realize early warning and source tracing of quality problems.

[0016] As a preferred embodiment of the technical solution of the present invention, the dynamic correction in S43 includes: The abnormal frequency and warning level are converted into correction coefficients, and the original membership degree of the rolling quality and process stability dimensions is multiplicatively reduced; where the higher the abnormal frequency or the higher the warning level, the smaller the correction coefficient and the lower the corrected membership degree. The formula for calculating the correction factor is: , This is an abnormal frequency. This is the warning level coefficient. and This is the preset sensitivity coefficient.

[0017] abnormal frequency and warning level coefficient This is converted into a correction coefficient. The higher the anomaly frequency or the higher the warning level, the smaller the correction coefficient, and the lower the membership degree after correction. The degree to which rolling quality and process stability are affected by anomalies differs; this can be addressed by setting a sensitivity coefficient. and To achieve differentiated correction, the evaluation results are made more consistent with the actual technological rules of cotton processing.

[0018] As a preferred embodiment of the technical solution of the present invention, the weighted synthesis operation in S44 includes: for each evaluation level, multiplying the membership degree of each indicator under that level with the corresponding indicator weight and then summing the results to obtain the comprehensive membership degree of that level, and the comprehensive membership degrees of all levels constitute a comprehensive evaluation vector. The method for determining the grade of cotton lint is as follows: the grade corresponding to the maximum value in the comprehensive evaluation vector is taken as the grade of the current cotton lint; this avoids the information loss that may be caused by the main factor-determined operator, and the evaluation result is more accurate.

[0019] The calculation method for the real-time comprehensive quality score is as follows: multiply the comprehensive membership degree of each level by the representative score value of the corresponding level, sum the results, and then divide by the sum of the comprehensive membership degrees to obtain the comprehensive quality score. This overcomes the deficiency that simple level determination cannot reflect subtle changes.

[0020] As a preferred embodiment of the technical solution of the present invention, S5 includes the following steps: S51. Obtain the real-time comprehensive quality score output by S4, calculate the score deviation between it and the preset target quality threshold, and construct a quality feedback vector with the score deviation and its rate of change as input. S52. Based on a pre-established nonlinear coupling model of process parameters and quality indicators, calculate the sensitivity coefficient of the quality feedback vector to key process parameters. The sensitivity coefficient is the partial derivative of the comprehensive quality score with respect to the key process parameters. The key process parameters include at least the ginning machine drum speed, drying temperature, and cotton feeding density. S53. Input the scoring deviation and the sensitivity coefficient into the adaptive controller, and use the gradient descent method or fuzzy PID algorithm to update the control gain online, and calculate the parameter adjustment amount of each key process parameter. S54. The parameter adjustment amount is superimposed on the current process parameter setting value to generate the optimized target process parameters; the target process parameters are sent to the production line PLC controller to perform physical adjustment.

[0021] By calculating the sensitivity coefficient of quality score to process parameters using a nonlinear coupling model, the controller can accurately identify the process parameters that have the greatest impact on quality, thus avoiding efficiency losses caused by blind adjustments.

[0022] The optimized target process parameters are sent to the PLC controller for physical adjustment, realizing adaptive optimization of production line process parameters and improving production efficiency while maintaining stable quality.

[0023] As a preferred embodiment of the technical solution of the present invention, in S53, when the nonlinear coupling model is differentiable, the gradient descent method is used to solve for the parameter adjustment amount: ,in For learning rate, The partial derivative of the overall quality score with respect to key process parameters is the sensitivity coefficient. When the model is inaccurate, fuzzy PID control is used. The PID gain adjustment amount is determined by looking up the fuzzy rule table based on the scoring deviation and its rate of change.

[0024] Secondly, the present invention also provides an intelligent identification and quality assessment system for process anomalies in unmanned cotton processing production lines, used to implement the method described in the first aspect, including: Multi-source data acquisition module: configured to synchronously acquire multi-source heterogeneous data from the entire production line, including equipment operating condition time-series data, machine vision image data, environmental parameters, and fiber spectral quality data; Feature engineering module: connected to the multi-source data acquisition module, configured to align the multi-source heterogeneous data using timestamp interpolation, extract time-domain statistical features, frequency-domain power spectrum features, image texture features and LSTM trend features, and concatenate them to construct a multi-dimensional feature vector; Anomaly identification module: connected to the feature engineering module, configured to input multi-dimensional feature vectors into the comprehensive intelligent model in parallel, obtain reconstruction error and fault probability, and then generate anomaly identification results to characterize the stability state of the production line process through dynamic threshold fusion; Quality assessment module: connected to the anomaly identification module, configured to establish a multi-dimensional quality index system including ginning quality, fiber quality and process stability. After determining the weights using AHP, it combines the fuzzy comprehensive evaluation model, using the fiber spectral quality data and the anomaly identification results output by the anomaly identification module as input, to calculate the real-time comprehensive quality score and determine the cotton grade. Adaptive control module: Connected to the quality assessment module, configured to use the deviation between the comprehensive quality score and the target threshold as feedback quantity, adjust key process parameters through adaptive control, and send the adjusted process parameters to the production controller to realize adaptive closed-loop optimization of production line process parameters.

[0025] As can be seen from the above technical solutions, this application has the following advantages: By synchronously collecting equipment operating time-series data, machine vision image data, environmental parameters, and fiber spectral quality data, a full-process multi-source sensing network is constructed, overcoming the shortcomings of incomplete information from a single data source. Using the anomaly identification results as input to the quality assessment model enables the quality assessment to perceive the health status of the production line process, achieve early warning of anomalies, and improve the stability and improvement of cotton lint quality.

[0026] Using the deviation between the comprehensive quality score and the target threshold as feedback, key process parameters are adjusted through adaptive control to form a fully closed-loop optimization system, thereby maximizing production efficiency while ensuring stable quality. Attached Figure Description

[0027] To more clearly illustrate the technical solution of this application, the accompanying drawings used in the description will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0028] Figure 1 This is a flowchart illustrating the method provided in an embodiment of the present invention.

[0029] Figure 2 A block diagram of a system provided in an embodiment of the present invention. Detailed Implementation

[0030] To make the purpose, features, and advantages of this application more apparent and understandable, specific embodiments and accompanying drawings will be used to clearly and completely describe the technical solution protected by this application. Obviously, the embodiments described below are only some embodiments of this application, and not all embodiments. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0031] Unless otherwise defined, all technical and scientific terms used in this application have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used in this application and in the specification of this invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention.

[0032] like Figure 1 As shown in the figure, this invention provides a method for intelligent identification and quality assessment of process anomalies in an unmanned cotton processing production line, including the following steps: S1. Synchronously collect multi-source heterogeneous data of the entire production line process. The multi-source heterogeneous data includes equipment operating condition time series data, machine vision image data, environmental parameters and fiber spectral quality data. This step enables the synchronous acquisition of multi-source heterogeneous data throughout the entire production line. Specifically, various types of sensors and detection equipment are deployed in key processes such as ginning, cleaning, and packaging in the cotton processing production line.

[0033] The acquisition of equipment operating condition timing data includes: installing vibration sensors on key parts such as bearing seats and housings of equipment such as cotton gins, cotton cleaning machines, and fans, with a sampling frequency set to 1kHz, to pick up vibration signals during equipment operation; installing current transformers in the motor drive circuit to collect current signals; installing speed sensors on the main shaft of the equipment to collect speed signals; installing temperature sensors on the bearings and motor windings of the equipment to collect temperature signals; and installing pressure sensors in the air pipeline to collect air pressure signals.

[0034] Machine vision image data acquisition includes: installing a line scan camera above the cotton layer conveying channel, using white, red, and green multispectral light sources, to acquire images of foreign fibers, defects, voids, and cotton knots; and installing a high-definition industrial camera at the cotton feed inlet to acquire images of cotton layer uniformity and feed rate fluctuations. The camera's sampling frame rate is set according to the production line speed, typically 30 frames per second.

[0035] The collection of environmental parameters includes: installing temperature and humidity sensors in the production line workshop to collect ambient temperature and humidity; installing dust concentration sensors in the dust removal ducts to collect dust concentration; and installing air pressure sensors in the air circuit system to collect air pressure parameters.

[0036] The acquisition of fiber spectral quality data includes: installing an online near-infrared spectrometer in the cotton flow channel to detect quality indicators such as fiber length, strength, micronaire value, and impurity content in real time. The sampling frequency of the near-infrared spectrometer is typically 1 Hz.

[0037] All the aforementioned sensors and detection devices are synchronized in time using a high-precision clock synchronization protocol to ensure the spatiotemporal consistency of the collected data.

[0038] S2. Align the multi-source heterogeneous data using timestamp interpolation, extract time-domain statistics, frequency-domain power spectrum, image texture, and LSTM trend features, and concatenate them to construct a multi-dimensional feature vector; this step specifically includes: S21. Select the equipment operating condition time series data with the highest sampling frequency as the main time reference, and use the linear interpolation method to resample the timestamps of environmental parameters and fiber spectral quality data to the main time reference to achieve data spatiotemporal alignment. S22. Perform sliding window segmentation on the aligned equipment operating condition time series data, calculate the mean, variance, peak factor and kurtosis in each window to obtain time domain statistical characteristics; at the same time, perform fast Fourier transform on the time domain signal to extract the main frequency energy and spectral entropy in the spectrum to obtain frequency domain power spectrum characteristics. After alignment, the equipment operating condition time-series data is segmented using a sliding window. Data is sampled within a preset window length (e.g., 128 time steps), and time-domain statistical features such as mean, variance, peak factor, and kurtosis are calculated within each window. The peak factor reflects the presence of impact components in the signal, while kurtosis is particularly sensitive to early bearing failures. Simultaneously, a Fast Fourier Transform is performed on the time-domain signal within the window to obtain the frequency-domain power spectrum. The dominant frequency energy and spectral entropy are extracted as frequency-domain power spectrum features. The dominant frequency energy reflects the vibration intensity at the equipment's main operating frequency, while the spectral entropy reflects the complexity of the frequency components.

[0039] S23. Perform grayscale processing and Gaussian filtering to denoise the machine vision image data, and use the gray-level co-occurrence matrix algorithm to calculate the contrast, energy, entropy and homogeneity index of the image to obtain the image texture feature vector. The following processing is performed on the machine vision image data, namely the cotton layer images acquired by the line scan camera: The original color image was converted to a grayscale image using a weighted average method: Gray = 0.299R + 0.587G + 0.114B.

[0040] Noise reduction was achieved by using a Gaussian filter (kernel size 3×3, standard deviation σ=0.8) to suppress noise caused by cotton lint and dust.

[0041] The grayscale levels were quantized and compressed from 256 levels to 16 levels to reduce computation. The grayscale co-occurrence matrix was calculated in four directions: pixel-to-pixel distance d=1 and direction θ = 0°, 45°, 90°, and 135°. The average value of the matrices in the four directions was then taken to obtain rotation invariance.

[0042] Based on the normalized GLCM, four texture feature indices are calculated: contrast, which reflects the gray-level difference between the cotton knot and the surrounding fibers; energy, which reflects the uniformity of the texture; entropy, which reflects the complexity of the texture; and homogeneity, which reflects the uniformity of local gray levels. These four features are used to construct an image texture feature vector for subsequent multimodal feature fusion.

[0043] Contrast: Measures the difference in grayscale between a knot and surrounding fibers, reflecting the clarity of the texture. The calculation formula is:

[0044] Energy: Measures the uniformity and regularity of the texture, reflecting the evenness of the cotton layer distribution. The calculation formula is:

[0045] Entropy: Measures the complexity and information content of a texture. The formula for calculation is:

[0046] Homogeneity: Measures the local uniformity of texture. The calculation formula is:

[0047] Where L is the gray quantization level. The first in the normalized gray-level co-occurrence matrix Line number The element values ​​of the column. The above four features constitute the image texture feature vector, which is used for subsequent multimodal fusion with equipment condition data, spectral data, etc. This texture feature extraction method is a well-known technology in the field of image processing, and its basic principles will not be elaborated in this embodiment.

[0048] S24. Arrange the aligned multi-source heterogeneous data into a time sequence according to a preset time window, input it into a pre-trained LSTM encoder, and extract the hidden layer output vector of the last time step as an LSTM trend feature to characterize the dynamic evolution of the production line. The training process for the LSTM encoder is as follows: (1) Collect historical data from the target cotton processing production line for at least 7 consecutive days of operation, including: equipment operating time series data (vibration, current, speed, temperature, pressure), machine vision image features (cotton layer uniformity, cotton feeding amount), environmental parameters (temperature and humidity, dust concentration), and fiber spectral quality data (length, strength, micronaire value, impurity content). (2) Remove abnormal operating condition segments such as equipment shutdown and sensor failure; have quality inspection experts mark the data segments as normal or various abnormalities; use the normally marked data segments as training samples for the LSTM encoder.

[0049] (3) Perform sliding window sampling on the aligned multi-source data with a fixed time window length L and a step size of L / 4 to generate a training sample set, where each sample has the following dimensions: [window length, feature dimension]; The LSTM encoder employs a sequence-to-vector (Seq2Vec) architecture. Since the goal of the LSTM encoder is to extract temporal patterns under normal operating conditions, this embodiment uses an autoencoder for pre-training. (4) Construct an LSTM autoencoder and an LSTM encoder to output a 32-dimensional trend feature vector z; and an LSTM decoder to reconstruct the original input sequence with z as the initial state.

[0050] (5) The loss function is the reconstruction error, using the mean squared error: in For the original input, To reconstruct the output of the decoder, This represents the length of the time window.

[0051] (6) Perform training. After training is complete, discard the decoder and keep only the encoder for extracting LSTM trend features.

[0052] (7) Calculate the average reconstruction error on the validation set, save the model parameters with the lowest validation error as the final LSTM encoder.

[0053] S25. The time-domain statistical features, frequency-domain power spectrum features, image texture feature vectors and LSTM trend features are respectively subjected to Z-score standardization. After eliminating the difference in dimensions, they are concatenated according to the channel dimension to obtain a multi-dimensional feature vector.

[0054] S3. Input the multi-dimensional feature vectors into the integrated intelligent model in parallel, obtain the reconstruction error and fault probability, and then fuse them through dynamic thresholding to generate anomaly identification results for characterizing the stability state of the production line process; the integrated intelligent model includes an autoencoder for unsupervised anomaly detection and a CNN-LSTM model for supervised fault classification. This step specifically includes: S31. The multidimensional feature vector is simultaneously input into the autoencoder and the CNN-LSTM model; wherein, the autoencoder performs encoding-decoding reconstruction on the input multidimensional feature vector and calculates the real-time reconstruction error between the original input vector and the reconstructed output vector; the CNN-LSTM model extracts the local spatial correlation between features through the CNN layer, then captures the temporal dependency through the LSTM layer, and outputs the fault probability vector of various known fault types through the Softmax layer. S32. Obtain the real-time reconstruction error output by the autoencoder and the fault probability vector output by the CNN-LSTM model; S33. Construct an upper limit for the dynamic confidence interval based on historical reconstruction errors. When the real-time reconstruction error exceeds the upper limit, it is determined to be an anomaly. The maximum value in the fault probability vector is weighted and fused with the normalized real-time reconstruction error to obtain a comprehensive anomaly score. The method for constructing the upper limit of the dynamic confidence interval is as follows: Obtain the historical average reconfiguration error under the current production line operating conditions. with standard deviation ; Set the upper limit of the dynamic confidence interval to ,in This is a preset constant; Weighted fusion yields a comprehensive anomaly score The calculation formula is: ,in This refers to the normalized real-time reconstruction error. The maximum value in the failure probability vector. For dynamic fusion weights.

[0055] In this embodiment, a historical reconstruction error sliding window of length N=1000 is maintained, and the mean is calculated in real time. with standard deviation The upper limit of the dynamic confidence interval is set to UCL= +3 , where constant =3 corresponds to a 99.7% confidence level. It should be noted that when the historical reconstruction error is insufficient, an initial threshold based on prior experience is used.

[0056] When the real-time reconstruction error REc satisfies REc>UCL, the current state is determined to be abnormal; otherwise, it is determined to be normal and subsequent fusion is not triggered.

[0057] After identifying the anomaly, the maximum value is extracted from the fault probability vector P output by the CNN-LSTM model as the supervised confidence score. Simultaneously, the real-time reconstruction error is normalized using Min-Max: .

[0058] The overall anomaly score is calculated using a weighted fusion formula: ,in This refers to the normalized real-time reconstruction error. The maximum value in the fault probability vector represents the dynamic fusion weight. Fusion weight The value is dynamically adjusted according to the production line conditions: 0.5 for the stable period, 0.3 for the period with frequent known faults, and 0.7 for the period with frequent unknown abnormalities. The value range is [0,1], and the higher the value, the more severe the abnormality.

[0059] S34. Based on the comprehensive anomaly score, match the preset warning level to generate a three-level warning; S35. Determine the anomaly type by combining the category label corresponding to the maximum value in the fault probability vector, and determine the anomaly location by mapping the current timestamp; count the anomaly frequency based on the number of anomalies occurring within the preset time window, and count the defect density based on the cumulative number of defect-type anomalies; generate anomaly identification results that include anomaly type, location, three-level warning, anomaly frequency, and defect density.

[0060] S4. Establish a multi-dimensional quality index system encompassing ginning quality, fiber quality, and process stability. After determining the weights using AHP, combine this system with a fuzzy comprehensive evaluation model. Using the fiber spectral quality data and the anomaly identification results output in S3 as input, calculate the real-time comprehensive quality score and determine the cotton lint grade. Specifically, this includes: S41. Establish a hierarchical model including a target layer, a criterion layer, and an indicator layer; the criterion layer includes fiber quality, rolling quality, and process stability; the indicator layer includes length, micronaire value, strength, and impurity content corresponding to fiber spectral quality data, as well as defect density, abnormal frequency, and warning level corresponding to the abnormal identification results output by S3. S42. Construct pairwise comparison judgment matrices for each level of indicators, calculate the largest eigenvalue and its corresponding eigenvector, obtain the initial weights of each indicator after normalization, and determine them as the final weight vectors after passing the consistency test. In this step, pairwise comparison judgment matrices for each level of indicators are constructed. Domain experts use a 1-9 scaling method to compare the importance of each indicator pairwise. The largest eigenvalue of the judgment matrix and its corresponding eigenvector are calculated. After normalizing the eigenvectors, the initial weights of each indicator are obtained. The consistency ratio is calculated for verification. When the consistency ratio is less than 0.1, the judgment matrix is ​​considered to meet the consistency requirements, and the initial weights are determined as the final weight vector.

[0061] S43. Set up a set of evaluation criteria for cotton grades, and use the membership function to convert the real-time collected fiber spectral quality data into membership degrees for each grade. At the same time, based on the anomaly frequency and warning level in the anomaly identification results output by S3, dynamically correct the membership degrees of the ginning quality and process stability dimensions, and generate the corrected fuzzy relation matrix. In this step, a set of evaluation criteria for cotton lint grades is established, typically including four levels: Excellent, Good, Medium, and Poor. Membership functions are then used to convert the real-time collected fiber spectral quality data into membership degrees for each grade. For quantitative indicators, trapezoidal or triangular membership functions are used; for qualitative indicators, membership degrees are set based on expert experience.

[0062] The dynamic correction includes: The abnormal frequency and warning level are converted into correction coefficients, which are then used to multiply and decrease the original membership degree of the rolling quality and process stability dimensions. Specifically, the higher the abnormal frequency or the higher the warning level, the smaller the correction coefficient and the lower the corrected membership degree. The formula for calculating the correction coefficient is as follows: , This is an abnormal frequency. This is the warning level coefficient. and This is the preset sensitivity coefficient.

[0063] It should be noted that the impact of anomaly frequency on product quality follows the law of diminishing marginal returns. The impact on product quality is most significant when an anomaly occurs for the first time on the production line; as the anomaly frequency continues to increase, the additional impact on quality from each new anomaly gradually decreases. The function exhibits a monotonically decreasing characteristic as the independent variable increases, with the rate of decrease gradually slowing down, which perfectly describes this engineering pattern. As the frequency of anomalies increases, the marginal impact of each new anomaly on the correction factor gradually decreases, consistent with engineering practice.

[0064] Warning level coefficient The value is determined according to the following rules: When the warning level is Level 1 (normal), =0; When the warning level is Level II (warning), =0.5; when the warning level is level three (alarm), =1. The advantages of this linear function are: simple calculation, suitable for real-time processing of edge computing units; clear physical meaning, the correction coefficient decreases proportionally with the increase of the warning level, intuitively reflecting the degree of impact of the warning level on quality; and the ability to preset sensitivity coefficients. The impact of the warning level on the membership degree can be flexibly controlled.

[0065] The impact of anomaly frequency and warning level on product quality exhibits a coupled amplification effect. When both the anomaly frequency and warning level are high, the two anomaly factors superimpose and reinforce each other, resulting in a combined impact on quality greater than the simple sum of their individual effects. The multiplicative form effectively reflects this coupled amplification effect, making the correction coefficient decrease more significantly under dual anomaly conditions. The smaller correction coefficient obtained through multiplicative synthesis corresponds to a greater decrease in membership degree, better reflecting the engineering judgment that quality should be further reduced under dual anomaly conditions.

[0066] Sensitivity coefficient and The following methods can be used to determine the specific production line conditions: Historical quality data under both normal and abnormal operating conditions are collected. The goal is to maximize the correlation between the corrected quality score and the actual quality grade, and the optimal parameters are obtained through least squares or grid search methods. Alternatively, calibration is performed by cotton processing experts based on their technological experience. The value range is from 0.3 to 0.8. The value range is from 0.2 to 0.5.

[0067] It should be further noted that the correction coefficient formula satisfies the following boundary conditions: when no anomalies occur ( =0, =0), =1, no membership degree correction; when the anomaly frequency approaches infinity and the warning level is the highest ( →∞, =1), →0, the membership degree is decayed to zero, indicating extremely poor quality; when and When taking the median value, The value varies continuously between 0 and 1, which can precisely characterize the impact of anomalies of different degrees on quality.

[0068] S44. The final weight vector is used to perform a weighted synthesis operation on the corrected fuzzy relation matrix to obtain a comprehensive evaluation vector; the current grade of the cotton is determined according to the maximum membership principle, and a weighted average method is used to calculate the real-time comprehensive quality score. The weighted synthesis operation uses a weighted average type fuzzy operator: ,in For the first The final weight value of each indicator, For the corrected fuzzy relation matrix, the first... The first indicator for the first Membership degree of each evaluation level; Overall quality score The calculation method is as follows: ,in For the comprehensive evaluation vector, the first... Membership degree at each level, For the first The representative score value corresponding to each level is calculated by weighting each component of the comprehensive evaluation vector with the representative score value of the corresponding level to obtain a continuous score within the range of 0-100.

[0069] S5. Using the deviation between the comprehensive quality score and the target threshold as feedback, key process parameters are adjusted through adaptive control, and the adjusted process parameters are sent to the production controller to achieve adaptive closed-loop optimization of production line process parameters. Specific steps: S51. Obtain the real-time comprehensive quality score output by S4, calculate the score deviation between it and the preset target quality threshold, and construct a quality feedback vector with the score deviation and its rate of change as input. S52. Based on a pre-established nonlinear coupling model of process parameters and quality indicators, calculate the sensitivity coefficient of the quality feedback vector to key process parameters. The sensitivity coefficient is the partial derivative of the comprehensive quality score with respect to the key process parameters. The key process parameters include at least the ginning machine drum speed, drying temperature, and cotton feeding density. The nonlinear coupling model can be in the form of a neural network model, response surface model, or mechanistic model, and the model parameters can be obtained by training with historical data.

[0070] S53. Input the scoring deviation and the sensitivity coefficient into the adaptive controller, and use the gradient descent method or fuzzy PID algorithm to update the control gain online, and calculate the parameter adjustment amount of each key process parameter. When the nonlinear coupling model is differentiable, the gradient descent method is used to solve for the parameter adjustment. The adjustment is equal to the learning rate multiplied by the sensitivity coefficient. The parameters are adjusted along the gradient ascent direction to maximize the overall quality score. That is, the parameter adjustment. ,in For learning rate, The partial derivative of the comprehensive quality score with respect to key process parameters is the sensitivity coefficient. When the model is inaccurate, fuzzy PID control is used, and the PID gain adjustment is determined by looking up the fuzzy rule table based on the score deviation and its rate of change.

[0071] S54. The parameter adjustment amount is added to the current process parameter setting value to generate the optimized target process parameters; the target process parameters are sent to the production line PLC controller to perform physical adjustment. The PLC controller drives the frequency converter or actuator to perform physical adjustment and correct the production line operating status.

[0072] Steps S51 to S54 are repeated to form a continuous control loop. In each loop, the physical state of the production line changes after parameter adjustment. The multi-source data acquisition in the next cycle will reflect this change, and the new quality score will be recalculated and used for the next round of control decisions, thereby achieving adaptive closed-loop optimization of the production line process parameters.

[0073] like Figure 2 As shown, this embodiment of the invention also provides an intelligent identification and quality assessment system for process anomalies in an unmanned cotton processing production line, used to implement the methods described in the above embodiments, including: Multi-source data acquisition module: configured to synchronously acquire multi-source heterogeneous data from the entire production line, including equipment operating condition time-series data, machine vision image data, environmental parameters, and fiber spectral quality data; Feature engineering module: connected to the multi-source data acquisition module, configured to align the multi-source heterogeneous data using timestamp interpolation, extract time-domain statistical features, frequency-domain power spectrum features, image texture features and LSTM trend features, and concatenate them to construct a multi-dimensional feature vector; Anomaly identification module: connected to the feature engineering module, configured to input multi-dimensional feature vectors into the comprehensive intelligent model in parallel, obtain reconstruction error and fault probability, and then generate anomaly identification results to characterize the stability state of the production line process through dynamic threshold fusion; Quality assessment module: connected to the anomaly identification module, configured to establish a multi-dimensional quality index system including ginning quality, fiber quality and process stability. After determining the weights using AHP, it combines the fuzzy comprehensive evaluation model, using the fiber spectral quality data and the anomaly identification results output by the anomaly identification module as input, to calculate the real-time comprehensive quality score and determine the cotton grade. Adaptive control module: Connected to the quality assessment module, configured to use the deviation between the comprehensive quality score and the target threshold as feedback quantity, adjust key process parameters through adaptive control, and send the adjusted process parameters to the production controller to realize adaptive closed-loop optimization of production line process parameters.

[0074] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for intelligent identification and quality assessment of process anomalies in an unmanned cotton processing production line, characterized in that, Includes the following steps: S1. Synchronously collect multi-source heterogeneous data of the entire production line process. The multi-source heterogeneous data includes equipment operating condition time series data, machine vision image data, environmental parameters and fiber spectral quality data. S2. Use timestamp interpolation to align the multi-source heterogeneous data, extract time-domain statistics, frequency-domain power spectrum, image texture and LSTM trend features, and concatenate them to construct a multi-dimensional feature vector; S3. Input the multi-dimensional feature vectors into the comprehensive intelligent model in parallel, obtain the reconstruction error and fault probability, and then use dynamic threshold fusion to generate anomaly identification results to characterize the stability state of the production line process. S4. Establish a multi-dimensional quality index system that includes ginning quality, fiber quality and process stability. After determining the weights using AHP, combine it with a fuzzy comprehensive evaluation model. Use the fiber spectral quality data and the anomaly identification results output by S3 as inputs to calculate the real-time comprehensive quality score and determine the cotton grade. S5. Using the deviation between the comprehensive quality score and the target threshold as feedback, the key process parameters are adjusted through adaptive control, and the adjusted process parameters are sent to the production controller to achieve adaptive closed-loop optimization of the production line process parameters.

2. The method for intelligent identification and quality assessment of process anomalies in unmanned cotton processing production lines according to claim 1, characterized in that, The steps in S2 include: S21. Select the equipment operating condition time series data with the highest sampling frequency as the main time reference, and use the linear interpolation method to resample the timestamps of environmental parameters and fiber spectral quality data to the main time reference to achieve data spatiotemporal alignment. S22. Perform sliding window segmentation on the aligned equipment operating condition time series data, calculate the mean, variance, peak factor and kurtosis in each window to obtain time domain statistical characteristics; at the same time, perform fast Fourier transform on the time domain signal to extract the main frequency energy and spectral entropy in the spectrum to obtain frequency domain power spectrum characteristics. S23. Perform grayscale processing and Gaussian filtering to denoise the machine vision image data, and use the gray-level co-occurrence matrix algorithm to calculate the contrast, energy, entropy and homogeneity index of the image to obtain the image texture feature vector. S24. Arrange the aligned multi-source heterogeneous data into a time sequence according to a preset time window, input it into a pre-trained LSTM encoder, and extract the hidden layer output vector of the last time step as an LSTM trend feature to characterize the dynamic evolution of the production line. S25. The time-domain statistical features, frequency-domain power spectrum features, image texture feature vectors and LSTM trend features are respectively subjected to Z-score standardization. After eliminating the difference in dimensions, they are concatenated according to the channel dimension to obtain a multi-dimensional feature vector.

3. The method for intelligent identification and quality assessment of process anomalies in unmanned cotton processing production lines according to claim 2, characterized in that, The integrated intelligent model includes an autoencoder for unsupervised anomaly detection and a CNN-LSTM model for supervised fault classification. Steps in S3 include: S31. The multidimensional feature vector is simultaneously input into the autoencoder and the CNN-LSTM model; wherein, the autoencoder performs encoding-decoding reconstruction on the input multidimensional feature vector and calculates the real-time reconstruction error between the original input vector and the reconstructed output vector; the CNN-LSTM model extracts the local spatial correlation between features through the CNN layer, then captures the temporal dependency through the LSTM layer, and outputs the fault probability vector of various known fault types through the Softmax layer. S32. Obtain the real-time reconstruction error output by the autoencoder and the fault probability vector output by the CNN-LSTM model; S33. Construct an upper limit of the dynamic confidence interval based on the historical reconstruction error. When the real-time reconstruction error exceeds the upper limit, it is judged as an anomaly. The maximum value in the fault probability vector is weighted and fused with the normalized real-time reconstruction error to obtain a comprehensive anomaly score. S34. Based on the comprehensive anomaly score, match the preset warning level to generate a three-level warning; determine the anomaly type by combining the category label corresponding to the maximum value in the fault probability vector, and determine the anomaly location by mapping the current timestamp; calculate the anomaly frequency based on the number of anomalies occurring within the preset time window, and calculate the defect density based on the cumulative number of defect-type anomalies; generate anomaly identification results that include anomaly type, location, three-level warning, anomaly frequency, and defect density.

4. The method for intelligent identification and quality assessment of process anomalies in unmanned cotton processing production lines according to claim 3, characterized in that, The method for constructing the upper limit of the dynamic confidence interval described in S33 is as follows: Obtain the historical average reconfiguration error under the current production line operating conditions. with standard deviation ; Set the upper limit of the dynamic confidence interval to ,in This is a preset constant; Weighted fusion yields a comprehensive anomaly score The calculation formula is: ,in This refers to the normalized real-time reconstruction error. The maximum value in the failure probability vector. For dynamic fusion weights.

5. The method for intelligent identification and quality assessment of process anomalies in unmanned cotton processing production lines according to claim 4, characterized in that, The steps in S4 include: S41. Establish a hierarchical model including a target layer, a criterion layer, and an indicator layer; the criterion layer includes fiber quality, rolling quality, and process stability; the indicator layer includes length, micronaire value, strength, and impurity content corresponding to fiber spectral quality data, as well as defect density, abnormal frequency, and warning level corresponding to the abnormal identification results output by S3. S42. Construct pairwise comparison judgment matrices for each level of indicators, calculate the largest eigenvalue and its corresponding eigenvector, obtain the initial weights of each indicator after normalization, and determine them as the final weight vectors after passing the consistency test. S43. Set up a set of evaluation criteria for cotton grades, and use the membership function to convert the real-time collected fiber spectral quality data into membership degrees for each grade. At the same time, based on the anomaly frequency and warning level in the anomaly identification results output by S3, dynamically correct the membership degrees of the ginning quality and process stability dimensions, and generate the corrected fuzzy relation matrix. S44. Use the final weight vector to perform a weighted synthesis operation on the modified fuzzy relation matrix to obtain a comprehensive evaluation vector; determine the current grade of the cotton lint according to the principle of maximum membership, and calculate the real-time comprehensive quality score using the weighted average method.

6. The method for intelligent identification and quality assessment of process anomalies in unmanned cotton processing production lines according to claim 5, characterized in that, The dynamic corrections described in S43 include: The abnormal frequency and warning level are converted into correction coefficients, and the original membership degree of the rolling quality and process stability dimensions is multiplicatively reduced; where the higher the abnormal frequency or the higher the warning level, the smaller the correction coefficient and the lower the corrected membership degree. The formula for calculating the correction factor is: , This is an abnormal frequency. This is the warning level coefficient. and This is the preset sensitivity coefficient.

7. The method for intelligent identification and quality assessment of process anomalies in unmanned cotton processing production lines according to claim 6, characterized in that, In S44, the weighted composite operation includes: for each evaluation level, multiplying the membership degree of each indicator under that level with the corresponding indicator weight and then summing the results to obtain the comprehensive membership degree of that level. The comprehensive membership degrees of all levels constitute the comprehensive evaluation vector. The method for determining the grade of cotton lint is as follows: take the grade corresponding to the maximum value in the comprehensive evaluation vector as the grade of the current cotton lint. The calculation method for the real-time comprehensive quality score is as follows: multiply the comprehensive membership degree of each level by the corresponding level's representative score, sum the results, and then divide by the sum of the comprehensive membership degrees to obtain the comprehensive quality score.

8. The method for intelligent identification and quality assessment of process anomalies in unmanned cotton processing production lines according to claim 7, characterized in that, S5 includes the following steps: S51. Obtain the real-time comprehensive quality score output by S4, calculate the score deviation between it and the preset target quality threshold, and construct a quality feedback vector with the score deviation and its rate of change as input. S52. Based on a pre-established nonlinear coupling model of process parameters and quality indicators, calculate the sensitivity coefficient of the quality feedback vector to key process parameters. The sensitivity coefficient is the partial derivative of the comprehensive quality score with respect to the key process parameters. The key process parameters include at least the ginning machine drum speed, drying temperature, and cotton feeding density. S53. Input the scoring deviation and the sensitivity coefficient into the adaptive controller, and use the gradient descent method or fuzzy PID algorithm to update the control gain online, and calculate the parameter adjustment amount of each key process parameter. S54. The parameter adjustment amount is superimposed on the current process parameter setting value to generate the optimized target process parameters; the target process parameters are sent to the production line PLC controller to perform physical adjustment.

9. The method for intelligent identification and quality assessment of process anomalies in unmanned cotton processing production lines according to claim 8, characterized in that, In S53, when the nonlinear coupling model is differentiable, the gradient descent method is used to solve for the parameter adjustment. ,in For learning rate, The partial derivative of the overall quality score with respect to key process parameters is the sensitivity coefficient. When the model is inaccurate, fuzzy PID control is used. The PID gain adjustment amount is determined by looking up the fuzzy rule table based on the scoring deviation and its rate of change.

10. A process anomaly intelligent identification and quality assessment system for an unmanned cotton processing production line, used to implement the method described in any one of claims 1-9, characterized in that, include: Multi-source data acquisition module: configured to synchronously acquire multi-source heterogeneous data from the entire production line, including equipment operating condition time-series data, machine vision image data, environmental parameters, and fiber spectral quality data; Feature engineering module: connected to the multi-source data acquisition module, configured to align the multi-source heterogeneous data using timestamp interpolation, extract time-domain statistical features, frequency-domain power spectrum features, image texture features and LSTM trend features, and concatenate them to construct a multi-dimensional feature vector; Anomaly identification module: connected to the feature engineering module, configured to input multi-dimensional feature vectors into the comprehensive intelligent model in parallel, obtain reconstruction error and fault probability, and then generate anomaly identification results to characterize the stability state of the production line process through dynamic threshold fusion; Quality assessment module: connected to the anomaly identification module, configured to establish a multi-dimensional quality index system including ginning quality, fiber quality and process stability. After determining the weights using AHP, it combines the fuzzy comprehensive evaluation model, using the fiber spectral quality data and the anomaly identification results output by the anomaly identification module as input, to calculate the real-time comprehensive quality score and determine the cotton grade. Adaptive control module: Connected to the quality assessment module, configured to use the deviation between the comprehensive quality score and the target threshold as feedback quantity, adjust key process parameters through adaptive control, and send the adjusted process parameters to the production controller to realize adaptive closed-loop optimization of production line process parameters.

Images