Online production monitoring method and system of modified mineral cable and cloud platform

Abstract

The invention provides an online production monitoring method and system for a modified mineral cable and a cloud platform, and the method comprises the steps: obtaining temperature data and pressure data at an extrusion stage, and correcting the temperature data to obtain corrected temperature data; establishing a pressure drift model by adopting historical data of a pressure sensor, and correcting current pressure data in real time; according to the sampling frequency of the stress data, performing segmentation processing on the stress data to obtain the stress data after segmentation processing; performing dimension reduction processing on the multi-source data set by adopting a principal component analysis method, extracting key features, and generating a low-dimensional feature matrix; inputting the low-dimensional feature matrix subjected to real-time processing into a support vector machine model, training a classifier, and judging whether the current production state is in a normal range or not; and according to an output result of the classifier, a production state report is generated to guide adjustment and optimization of a subsequent production process. The monitoring precision of the modified mineral cable production process is effectively improved, and the production process is adjusted and optimized in real time.

Description

Technical Field

[0001] The present invention relates to the field of information technology, and in particular to an online production monitoring method, system and cloud platform for modified mineral cables. Background Art

[0002] Background: There is a key technical contradiction in the online production monitoring process of modified mineral cables. On the one hand, in order to comprehensively monitor all aspects of the production process, it is necessary to collect and fuse multi-source heterogeneous data, such as temperature, pressure, vibration, etc. These data come from different types of sensors with different sampling frequencies, precisions and formats. On the other hand, the fusion and processing of heterogeneous data face challenges such as data redundancy, noise interference and lack of real-time performance. How to efficiently integrate and utilize multi-source heterogeneous data while ensuring data quality and real-time performance is a technical problem that needs to be solved urgently. Specifically, in the extrusion stage of cable production, temperature and pressure are key factors affecting product quality. However, due to the high temperature and high pressure environment of the extrusion process, the measurement accuracy and stability of the sensor are easily affected, resulting in deviation and drift of the collected data. At the same time, in subsequent processes such as cooling and wrapping, factors such as vibration and stress will also affect the performance of the cable. How to achieve accurate measurement and real-time fusion of multiple parameters in a complex production environment is a technical challenge. In addition, the inconsistency and redundancy of heterogeneous data also increase the difficulty of data processing and analysis. How to design a high-precision data fusion algorithm, eliminate redundancy and noise, and extract key features is another technical challenge that needs to be overcome. Summary of the invention

[0003] The present invention provides an online production monitoring method for a modified mineral cable, which mainly comprises:

[0004] Acquire the temperature data and pressure data in the extrusion stage, and correct the temperature data through a preset temperature compensation model to obtain corrected temperature data;

[0005] The pressure drift model is established using the historical data of the pressure sensor, and the current pressure data is corrected in real time to obtain the corrected pressure data;

[0006] Obtain vibration data in the cooling stage and stress data in the wrapping stage, perform denoising on the vibration data through wavelet transform, extract effective features in the vibration signal, and obtain denoised vibration data;

[0007] According to the sampling frequency of stress data, the sliding window algorithm is used to process the data in sections, the mean and variance of each section of data are calculated, and the stress data after section processing is obtained;

[0008] Align the corrected temperature data, corrected pressure data, denoised vibration data, and segmented stress data according to the timestamp to generate a multi-source dataset in a unified format;

[0009] Use the principal component analysis method to perform dimensionality reduction on the multi-source dataset, extract key features, and generate a low-dimensional feature matrix;

[0010] Judge whether the processing time of the low-dimensional feature matrix meets the real-time requirement through a preset real-time threshold. If the processing time exceeds the real-time threshold, use the incremental principal component analysis method to dynamically update the data;

[0011] Input the low-dimensional feature matrix after real-time processing into the support vector machine model to train the classifier and judge whether the current production state is within the normal range;

[0012] Generate a production status report according to the output result of the classifier, and guide the adjustment and optimization of the subsequent production process according to the production status report.

[0013] The present invention provides an on-line production monitoring system for modified mineral cables, mainly including:

[0014] A temperature data acquisition and correction module for acquiring temperature data and pressure data in the extrusion stage, and correcting the temperature data through a preset temperature compensation model to obtain corrected temperature data;

[0015] A pressure data acquisition and correction module for establishing a pressure drift model using the historical data of the pressure sensor and performing real-time correction on the current pressure data to obtain corrected pressure data;

[0016] A vibration data denoising and feature extraction module for acquiring vibration data in the cooling stage and stress data in the winding stage, performing denoising processing on the vibration data through wavelet transform, extracting effective features in the vibration signal, and obtaining denoised vibration data;

[0017] A stress data segmentation processing module for segmenting the data using a sliding window algorithm according to the sampling frequency of the stress data, calculating the mean and variance of each segment of data, and obtaining segmented stress data;

[0018] A multi-source dataset alignment module for aligning the corrected temperature data, corrected pressure data, denoised vibration data, and segmented stress data according to the timestamp to generate a multi-source dataset in a unified format;

[0019] A data dimensionality reduction processing module for performing dimensionality reduction on the multi-source dataset using the principal component analysis method, extracting key features, and generating a low-dimensional feature matrix;

[0020] A real-time judgment and dynamic update module, which is used to judge whether the processing time of the low-dimensional feature matrix meets the real-time requirement through a preset real-time threshold. If the processing time exceeds the real-time threshold, incremental principal component analysis is used to dynamically update the data;

[0021] A classifier training module, which is used to input the low-dimensional feature matrix after real-time processing into a support vector machine model to train the classifier and judge whether the current production status is within the normal range;

[0022] A production status report generation module, which is used to generate a production status report according to the output result of the classifier and guide the adjustment and optimization of the subsequent production process based on the production status report.

[0023] The present invention provides a cloud platform, including the on-line production monitoring system of the modified mineral cable described in any one of the above.

[0024] The technical solution provided by the embodiment of the present invention may include the following beneficial effects:

[0025] The present invention discloses a real-time monitoring method for the production process of a modified mineral cable. Aiming at the problems of acquisition and processing of multi-source data such as temperature, pressure, vibration and stress in the production process, a comprehensive solution is proposed. The present invention first corrects the temperature and pressure data in the extrusion stage to eliminate the influence of the environment and drift. Then, the vibration and stress data in the cooling and winding stages are denoised and segmented to extract effective features. Next, the multi-source data are aligned and dimension-reduced to generate a low-dimensional feature matrix. Finally, a support vector machine model is used to classify and monitor the production status in real time. Through the fusion analysis and real-time processing of multi-source data, the present invention effectively improves the monitoring accuracy of the production process of the modified mineral cable and provides a reliable basis for the adjustment and optimization of the production process. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 It is a flowchart of an on-line production monitoring method for a modified mineral cable of the present invention.

[0027] Figure 2 It is a schematic diagram of an on-line production monitoring method, system and cloud platform for a modified mineral cable of the present invention.

[0028] Figure 3 It is another schematic diagram of an on-line production monitoring method, system and cloud platform for a modified mineral cable of the present invention.

[0029] Figure 4 It is a structural schematic diagram of an on-line production monitoring method, system and cloud platform for a modified mineral cable of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0030] The technical solution of the present invention will be clearly and completely described below in conjunction with the embodiments. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative work belong to the scope of protection of the present invention.

[0031] As Figures 1-4 , an on-line production monitoring method, system and cloud platform for a modified mineral cable in this embodiment may specifically include:

[0032] Step S101, obtain the temperature data and pressure data in the extrusion stage, and correct the temperature data through a preset temperature compensation model to obtain the corrected temperature data.

[0033] Obtain the original temperature data and pressure data in the extrusion stage, and use the original temperature data and pressure data as the input of the temperature compensation model. According to the preset temperature compensation model, perform correction calculation on the original temperature data to obtain the corrected temperature data.

[0034] The temperature compensation model adopts the multiple linear regression algorithm. Taking the original temperature data, ambient temperature, etc. as independent variables and the corrected temperature data as the dependent variable, establish a regression equation, and correct it in combination with the influence of high ambient temperature on temperature measurement. Judge whether the corrected temperature data is within the preset normal range. If the corrected temperature data exceeds the normal range, trigger a temperature anomaly alarm. Perform correlation analysis on the corrected temperature data and the original pressure data, calculate the correlation coefficient between the corrected temperature data and the original pressure data through the Pearson correlation coefficient, and determine the degree of influence of temperature change on pressure. According to the correlation coefficient, dynamically adjust the parameters of the PID controller, such as the proportional coefficient, integral time, differential time, etc., to achieve precise control of temperature and pressure. Apply the adjusted control parameters to the control of the extrusion process. Record and store the corrected temperature data, the original pressure data, and the adjusted control parameters to form an extrusion stage database. Regularly evaluate and optimize the temperature compensation model, introduce new influencing factors such as material characteristics and equipment status, and adopt optimization algorithms such as gradient descent and regularization to improve the accuracy and reliability of temperature compensation.

[0035] Specifically, during the extrusion process, the original temperature data and pressure data are collected by sensors. The sampling frequency of the temperature data is 10 Hz, and the sampling frequency of the pressure data is 5 Hz. The collected original temperature data and pressure data are input into a preset temperature compensation model. This model uses the multiple linear regression algorithm, comprehensively considering the influence of factors such as ambient temperature and humidity on temperature measurement, and fitting out the temperature compensation equation by the least squares method: T_corrected = 98 * T_raw + 02 * H + 5, where T_corrected is the corrected temperature data, T_raw is the original temperature data, and H is the ambient humidity. The original temperature data is corrected according to the temperature compensation equation to obtain the corrected temperature data. It is judged whether the corrected temperature data is within the normal range (180°C - 220°C). If the temperature data exceeds the normal range, a temperature anomaly alarm is triggered, and the system automatically sends an alarm message to the operator to prompt an inspection. The corrected temperature data and the original pressure data are subjected to correlation analysis. The Pearson correlation coefficient method is used to calculate the correlation coefficient between temperature and pressure, and the correlation coefficient r = 85 is obtained, indicating that there is a strong positive correlation between temperature change and pressure change.

[0036]

[0037] r represents the Pearson correlation coefficient, T_i and P_i represent the i-th temperature and pressure data points respectively, T and P represent the average values of temperature and pressure respectively, and n represents the total number of data points. According to the correlation coefficient between temperature and pressure, the control parameters of the extrusion process are dynamically adjusted. The PID control algorithm is used, the temperature target value is set to 200°C, and the pressure target value is set to 10 MPa. The precise control of temperature and pressure is achieved by adjusting the heating power and extrusion speed to ensure the stability of the extrusion quality. The information such as the corrected temperature data, the original pressure data, and the adjusted control parameters is recorded and stored in chronological order. A record is generated every 10 minutes to form a complete extrusion stage database, providing data support for subsequent data analysis and process optimization. The temperature compensation model is evaluated and optimized regularly. The model is retrained once a month. By introducing new influencing factors (such as equipment vibration) and algorithms (such as support vector machines), the accuracy and reliability of temperature compensation are continuously improved. After 3 months of optimization, the error of the temperature compensation model is reduced by 20%, ensuring the authenticity and accuracy of the measurement data in the extrusion stage.

[0038] Step S102, establish a pressure drift model using the historical data of the pressure sensor to perform real-time correction on the current pressure data to obtain the corrected pressure data.

[0039] Obtain the historical data of the pressure sensor and store it in the database for subsequent processing. Read the historical data of the pressure sensor from the database, preprocess the historical data, and obtain the preprocessed historical data. Among them, the preprocessing process includes removing the outliers and noise data therein, which can be achieved by setting thresholds, median filtering and other methods; it is not limited in the embodiments of the present invention. According to the preprocessed historical data, use the LSTM neural network algorithm to construct a pressure drift prediction model. Divide the preprocessed historical data into a training set and a test set for model training and verification. During the training process of the pressure drift prediction model, adopt the K-fold cross-validation method, divide the training set data into K parts, successively take K-1 of them as training data, and the remaining 1 part as verification data, and conduct K times of training and verification to obtain K models. At the same time, adopt the grid search method to traverse different hyperparameter combinations, such as the number of hidden layers, the number of neurons in the hidden layer, the learning rate, etc., and select the hyperparameters with the best verification effect from them. Package the trained pressure drift prediction model as an API interface and deploy it to the real-time data processing system. The real-time system inputs the real-time data collected by the pressure sensor into the model by calling this API to obtain the pressure drift prediction result. According to the pressure drift prediction result, obtain the drift value and judge whether the current pressure sensor drifts. If the drift value exceeds the preset threshold, it is considered that the real-time pressure data needs to be corrected. During correction, subtract the predicted drift value from the real-time pressure value to obtain the corrected pressure data.

[0040] Furthermore, perform a moving average smoothing process on the corrected pressure data to smooth the short-term fluctuations of the pressure. The specific method is to take the arithmetic average of the nearest N pressure measurement values. N can be set to 1 to 2 times the sampling frequency to obtain the smoothed pressure measurement value. Output and display the smoothed pressure measurement value, or store it in the database for subsequent business applications. Add the smoothed pressure measurement value to the historical data set for the iterative update training of the subsequent pressure drift prediction model.

[0041] Specifically, first, historical data is obtained from the pressure sensor and the data is preprocessed. The 3-σ criterion is used to eliminate outliers, that is, data points that exceed the mean ±3 times the standard deviation are considered outliers. Then, wavelet transform is used to reduce the noise of the data, and a suitable wavelet basis function such as db4 is selected to decompose the data into 5 layers to remove high-frequency noise components. Then, the preprocessed historical data is divided into a training set and a test set with a ratio of 8:2. The support vector regression (SVR) algorithm is selected to construct a pressure drift prediction model. The radial basis kernel function is used. The model hyperparameters, such as the penalty coefficient C and the kernel function parameter γ, are optimized through 5-fold cross validation and grid search. The root mean square error (RMSE) and determination coefficient (R2) of the model are evaluated on the test set. The optimized model RMSE is 0.5 and R2 is 98. The trained SVR model is deployed in the real-time data processing system to predict the real-time data collected by the pressure sensor every second. According to the prediction results of the SVR model, the real-time pressure data is corrected. The correction formula is: P_corrected = P_measured-P_dr i ft, where P_corrected is the corrected pressure value, P_measured is the sensor measurement value, and P_dr i ft is the drift value predicted by the SVR model. Kalman filtering is performed on the corrected pressure data, and the process noise covariance is set to 01, the measurement noise covariance is set to 1, the initial estimate is the first corrected pressure value, and the initial estimate covariance is 1. State prediction, Kalman gain calculation and state update are performed recursively to obtain the smoothed pressure estimate. Finally, the corrected and smoothed pressure measurement value is output to the data visualization system in JSON format, and the control system is triggered to make corresponding adjustments according to the pressure value. Predicting pressure drift through machine learning algorithms, combined with real-time correction and Kalman filtering, can effectively improve the measurement accuracy and stability of pressure sensors and provide high-quality pressure data for downstream systems.

[0042] Step S103, obtaining vibration data in the cooling stage and stress data in the wrapping stage, denoising the vibration data through wavelet transform, extracting effective features in the vibration signal, and obtaining denoised vibration data.

[0043] According to the vibration sensor in the cooling stage and the stress sensor in the wrapping stage, the vibration data in the cooling stage and the stress data in the wrapping stage are collected in real time. According to the timestamp information contained in the vibration data and the stress data, the data are synchronized and aligned in time, and the data are divided into two parts, the cooling stage and the wrapping stage, according to the preset stage division rules. The vibration data in the cooling stage is subjected to wavelet denoising, and Daubech i es wavelet is selected as the wavelet basis function, and the number of decomposition layers is set to 5 layers. The high-frequency noise in the vibration data is removed by wavelet threshold filtering to obtain the denoised vibration signal.

[0044] Furthermore, perform a short-time Fourier transform on the denoised vibration signal to extract the time-frequency characteristics of the vibration signal, including parameters such as the amplitude, frequency, and energy distribution of each frequency component. Use the extracted time-frequency characteristics of the vibration signal as the input and the stress data in the winding stage as the output, and establish an association model between the vibration signal and the stress using the support vector regression algorithm. Train and optimize the model through the method of cross-validation. Utilize the established association model to conduct predictive analysis on the newly collected vibration data in the cooling stage to obtain the predicted stress values corresponding to the winding stage. If the predicted stress value exceeds the preset safety threshold range, trigger the warning mechanism, generate warning information, and send it to the equipment management personnel and relevant technical personnel in a timely manner via text message or email. According to the set model update period, regularly retrain and optimize the association model using the newly collected vibration data and the newly collected stress data. After each model update, evaluate the prediction performance of the model and adjust the hyperparameters and feature selection methods of the model to ensure that the association model can continuously and effectively predict the winding quality.

[0045] Specifically, when obtaining the original data in the cooling stage and the winding stage, vibration sensors and stress sensors can be installed on the equipment to collect vibration signals and stress signals. The vibration sensor can be an acceleration sensor with a sampling frequency set to 1000 Hz, and the stress sensor can be a strain gauge sensor with a sampling frequency set to 100 Hz. According to the production process flow, determine the time ranges of the cooling stage and the winding stage, and use timestamps to divide the collected data to obtain the vibration data in the cooling stage and the stress data in the winding stage. For the vibration data in the cooling stage, a wavelet transform algorithm is used for denoising. Select Daubechies wavelet as the wavelet basis, set the decomposition level to 5 layers, and through the wavelet threshold denoising method, set the high-frequency noise coefficients to zero and then perform wavelet reconstruction to obtain the denoised vibration signal. Perform time-frequency analysis on the denoised vibration signal, using the short-time Fourier transform, set the window length to 256 sampling points, and the overlap rate to 50%, and extract the time-frequency characteristics of the vibration signal, including characteristic parameters such as frequency, amplitude, and energy. According to the extracted vibration signal characteristics and the stress data in the winding stage, establish a correlation model between the vibration signal and the stress. Use the support vector machine algorithm, select the radial basis kernel function, optimize the model parameters through the grid search method, and train to obtain the optimal correlation model. Use the established correlation model to perform predictive analysis on the newly collected vibration data in the cooling stage. When new vibration data arrives, extract its time-frequency characteristics and input them into the correlation model to predict the corresponding stress level in the winding stage. Set the stress threshold range to ±10%, and when the predicted stress exceeds the threshold range, trigger the warning mechanism and notify relevant personnel to adjust process parameters or perform equipment maintenance via text message or email. To continuously optimize and update the correlation model, regularly collect new production data, such as performing data updates once a week or once a month. Use the newly collected data to retrain the model, evaluate the model performance through methods such as cross-validation, and continuously improve the prediction accuracy and adaptability of the model. At the same time, establish a data quality monitoring mechanism to perform anomaly detection and filtering on the collected data to ensure the reliability and consistency of the data and provide high-quality data support for model training. Through continuous data updates and model optimization, ensure that potential quality problems can be detected and solved in a timely manner during the production process, and improve production efficiency and product quality.

[0046] Step S104, according to the sampling frequency of the stress data, use the sliding window algorithm to segment the data, calculate the mean and variance of each segment of data, and obtain the stress data after segmentation processing.

[0047] Obtain stress data, where the stress data has a preset sampling frequency. According to the sampling frequency, determine the size of the sliding window. Use the sliding window to segment the stress data to obtain the stress data after segmentation processing.

[0048] For each of the stress data after the segmented processing, calculate its mean value and variance to obtain the statistical characteristics of the stress data after the segmented processing. Preset the outlier judgment threshold according to the design requirements. Compare the statistical characteristics of each of the stress data after the segmented processing with the outlier judgment threshold. If the statistical characteristics exceed the range of the outlier judgment threshold, determine that the stress data segment after the segmented processing is a stress data segment containing outliers. For each stress data segment containing outliers, use data smoothing methods such as moving average or spline interpolation for processing to obtain the smoothed stress data segment. For each of the smoothed stress data segments, recalculate its mean value and variance to update the statistical characteristics of the stress data segment to eliminate the influence of outliers on the statistical characteristics. Summarize the statistical characteristics of all the smoothed stress data segments to obtain the overall statistical characteristic distribution of the stress data. According to the overall statistical characteristic distribution, use the k-means clustering algorithm to perform pattern recognition on the stress data to obtain the change pattern of the stress data. According to the overall statistical characteristic distribution, use the support vector machine classification algorithm to perform trend analysis on the stress data to obtain the change trend of the stress data.

[0049] Specifically, assuming that the sampling frequency of the stress data is 100 Hz, a sliding window with a size of 1 second can be selected, that is, each window contains 100 data points. Through the sliding window algorithm, the stress data is divided into multiple data segments with a length of 1 second. For each data segment, calculate its mean value and variance. For example, the mean value of the first segment is 10 MPa and the variance is 5. Set the outlier judgment threshold to ±3 times the standard deviation of the mean value, and judge the statistical characteristics of each segment. If the mean value of the second segment is 15 MPa and exceeds the threshold range, determine that this segment contains outliers. Use the Kalman filter algorithm to correct the outliers, smooth the abnormal data points, recalculate the segment mean value to be 12 MPa, the variance to be 3, and update the statistical characteristics of this segment. Summarize the statistical characteristics of all segments to obtain the mean value of the entire stress data to be 11 MPa and the variance to be 4, and draw a statistical characteristic distribution histogram. Use the K-means clustering algorithm to divide the data, identify 3 typical stress patterns: low stress state, normal working state, and high stress state, and analyze the change trends of the stress data under different patterns. Combining the work logs and maintenance records of the equipment, it is found that the high stress state often appears during equipment startup and failure, providing a basis for predicting the health status of the equipment.

[0050] Step S105, align the corrected temperature data, corrected pressure data, denoised vibration data, and stress data after segmented processing according to the time stamp to generate a multi-source data set in a unified format.

[0051] Obtain the temperature data and pressure data after calibration processing, obtain the vibration data after denoising processing, and obtain the stress data after segmentation processing. Judge whether the timestamps of each data are consistent. If the timestamps are inconsistent, use the linear interpolation method to perform time alignment processing on each data so that each data is consistent in the time dimension. According to the aligned temperature data, pressure data, vibration data, and stress data, use the Kalman filter algorithm for multi-source data fusion to obtain fusion data in a unified format. For the unified format data after fusion, use the hash algorithm to perform deduplication processing to eliminate redundant data. Use the median filter algorithm to perform denoising processing on the data, smooth the data and eliminate abnormal noise. Use the box plot method to identify outliers, and regard the data exceeding 5 times the interquartile range of the upper and lower quartiles as outliers and eliminate them to obtain a high-quality multi-source data set.

[0052] According to the data characteristics of the multi-source data set, use the feature selection algorithm based on mutual information to calculate the mutual information between each feature and the target variable, and select the features with high mutual information as the final feature subset. Use the min-max normalization algorithm to normalize the selected feature subset to eliminate the influence of the dimension of different data. Randomly divide the normalized multi-source data set into a training set, a validation set, and a test set according to the ratio of 8:1:1. Use the support vector regression algorithm to construct a prediction model, and optimize the model hyperparameters through the grid search method to obtain a multi-source data fusion prediction model with excellent performance.

[0053] Specifically, first, obtain the temperature data and pressure data that have been corrected. The correction error of the temperature data is controlled within ±1°C, and the correction error of the pressure data is controlled within ±5%. Then, obtain the vibration data that has been processed by wavelet denoising, and the signal-to-noise ratio after denoising is increased by more than 10 dB. Next, obtain the stress data that has been processed by piecewise linear fitting, and the fitting error is controlled within ±1%. According to the timestamp information of each data, use the dynamic time warping algorithm with a 1s time window to perform time alignment processing on each data, so that the error of each data in the time dimension is controlled within ±0.5s. Through the aligned temperature, pressure, vibration, and stress data, use the multiple linear regression algorithm with temperature, pressure, and vibration data as independent variables and stress data as the dependent variable to construct a data fusion model. The format of the fused data is (timestamp, temperature, pressure, vibration, stress). For the fused data, use the density clustering algorithm with the Euclidean distance as the similarity measure and a density threshold of 0.5 as the denoising parameter to remove noise data points. Then, use the box plot analysis method with the upper and lower quartiles as the outlier thresholds to detect and remove outliers. Next, use the maximum information coefficient method with a correlation threshold of 5 to select three features of temperature, pressure, and vibration as the key feature subset affecting stress. According to the selected feature subset, use the min-max normalization algorithm to normalize the temperature data to the interval [0,1], the pressure data to the interval [2,8], the vibration data to the interval [-1,1], and the stress data to the interval [0,100]. Randomly divide the normalized multi-source dataset into a training set, a validation set, and a test set according to the ratio of 8:1:1. Finally, select the support vector regression algorithm with the radial basis kernel function as the kernel function and grid search cross-validation as the parameter tuning method to construct a stress prediction model. Through model training and optimization, the prediction accuracy on the test set reaches more than 95%, and the mean absolute error is less than 5%, which can achieve high-precision multi-source data fusion stress prediction.

[0054] Step S106: Use the principal component analysis method to perform dimensionality reduction processing on the multi-source dataset, extract key features, and generate a low-dimensional feature matrix.

[0055] The principal component analysis method is used to preprocess the data and extract features, mapping the original high-dimensional data to a low-dimensional space, extracting the key features of the data, and eliminating the differences brought by different data sources and formats. The preprocessed data is dimensionally reduced, and the principal components are obtained through the principal component analysis method, removing redundant and noisy data to obtain a compressed low-dimensional feature matrix. The compressed low-dimensional feature matrix is used as the input of the support vector machine algorithm to train a classification model to achieve intelligent classification of multi-source heterogeneous data. The support vector machine can effectively process high-dimensional data by finding the optimal classification hyperplane and has good generalization ability. According to the classification results of the support vector machine model, the original data is labeled. By analyzing the feature differences of different category data, the internal connections and rules between different data sources are revealed, providing data support for business decisions. For new multi-source heterogeneous data, first use the same preprocessing and feature extraction methods as the original data to transform it into a low-dimensional feature matrix with the same input format as the model. Then, directly perform classification prediction through the trained support vector machine model to achieve fast processing and real-time application of the data, improving business efficiency and value.

[0056] Specifically, first preprocess the multi-source heterogeneous data, and use the principal component analysis method to map the data from different data sources to the same space. For example, convert data in different formats such as text, images, and audio into vector representations, and then reduce the vector dimension to 100 dimensions through principal component analysis, eliminating the differences between different data sources and formats. Then, perform feature extraction and dimensionality reduction on the preprocessed 100-dimensional data, and use the principal component analysis method to further reduce the dimension to 20 dimensions to extract the key features of the data. Next, use the 20-dimensional feature vector matrix obtained by principal component analysis to reconstruct the original data, removing redundant and noisy data to obtain a compressed 20-dimensional low-dimensional feature matrix. Then, use the singular value decomposition algorithm to reduce the 20-dimensional matrix to 10 dimensions, extracting the most important feature vectors, greatly reducing the data storage and calculation overhead. Finally, input the compressed 10-dimensional feature matrix into the support vector machine model for training to achieve classification and prediction of multi-source heterogeneous data. By analyzing the results output by the model, the original data can be labeled and interpreted, and the correlation rules between different data sources can be found. For example, there is a high correlation between the names mentioned in the text data and the faces appearing in the image data, providing strong data support for business decisions. When new multi-source heterogeneous data needs to be analyzed, the trained model can be directly used for prediction to achieve fast processing and real-time application of the data, greatly improving the processing efficiency.

[0057] Step S107, judge whether the processing time of the low-dimensional feature matrix meets the real-time requirement through a preset real-time threshold. If the processing time exceeds the real-time threshold, the incremental principal component analysis method is used to dynamically update the data.

[0058] Obtain the preset real-time threshold and the low-dimensional feature matrix. Record the processing time of the low-dimensional feature matrix through a timer, and determine whether the processing time meets the requirements of the preset real-time threshold. If the processing time of the low-dimensional feature matrix exceeds the preset real-time threshold, then according to the data volume and data distribution characteristics of the low-dimensional feature matrix, use the incremental principal component analysis method to perform dynamic update processing on the low-dimensional feature matrix. When performing incremental principal component analysis, calculate the covariance matrix of the low-dimensional feature matrix, apply the Jacobi method to obtain the eigenvalues and eigenvectors of the covariance matrix, and select the eigenvectors corresponding to the top k largest eigenvalues according to the eigenvalue magnitudes to determine the number k of principal components. According to the determined number k of principal components, use the online learning algorithm IPCA to dynamically update the k principal components of the low-dimensional feature matrix to obtain the updated low-dimensional feature matrix. Use the updated low-dimensional feature matrix as the new feature data for subsequent data processing and analysis to improve the real-time performance and efficiency of data processing. Continuously iterate and update the principal components of the low-dimensional feature matrix multiple times to enable the low-dimensional feature matrix to adapt to the dynamic changes of the data and maintain the effectiveness of the low-dimensional feature representation of the data. When the processing time of the low-dimensional feature matrix is continuously lower than the preset real-time threshold for multiple times, stop the iterative update.

[0059] Furthermore, input the dynamically updated low-dimensional feature matrix into the support vector machine SVM for pattern recognition and classification prediction, continuously optimize the parameters of the SVM, and improve the classification performance and generalization ability of the SVM model. Feed back the classification results of the SVM into the IPCA algorithm to further guide the dynamic update of the low-dimensional feature matrix, forming a closed loop for dynamically updating the low-dimensional features and optimizing the machine learning model.

[0060] Specifically, assume that the preset real-time threshold is 10 milliseconds, and the processing time of the low-dimensional feature matrix is 12 milliseconds, exceeding the preset real-time threshold. According to the data characteristics of the low-dimensional feature matrix, the incremental principal component analysis method is used to perform dynamic update processing on the low-dimensional feature matrix. First, calculate the covariance matrix of the low-dimensional feature matrix, and obtain the eigenvalues of the covariance matrix as [5, 8, 2, 8, 5] and the corresponding eigenvectors. According to the eigenvalue magnitudes, determine that the number of principal components is 3. Then, use an incremental learning algorithm, such as the online principal component analysis algorithm, to dynamically update the principal components of the low-dimensional feature matrix. Specifically, when new data samples arrive, calculate the covariance matrix, eigenvalues, and eigenvectors incrementally, and update the principal component space. The dimension of the updated low-dimensional feature matrix is reduced to 3, and the processing time is shortened to 8 milliseconds, meeting the real-time requirement. Use the updated low-dimensional feature matrix as new feature data for subsequent data processing and analysis, such as clustering analysis. Through the K-means clustering algorithm, divide the updated low-dimensional feature matrix into several clusters, and the data points within each cluster have similar characteristics. The clustering results can be used in application scenarios such as anomaly detection and user profiling. By continuously iteratively updating the principal components of the low-dimensional feature matrix, it can adapt to the dynamic changes of the data, maintain the effectiveness of the low-dimensional feature representation of the data, and improve the performance and generalization ability of the machine learning model.

[0061] Step S108: Input the low-dimensional feature matrix after real-time processing into the support vector machine model to train the classifier and determine whether the current production state is within the normal range.

[0062] Obtain the real-time operation data of the production equipment, where the real-time operation data includes key parameters such as the temperature, pressure, and vibration of the equipment; preprocess the real-time operation data to obtain preprocessed data, and the preprocessing includes data cleaning, standardization, and outlier removal; extract key feature parameters from the preprocessed data, and the key feature parameters include time-domain features and frequency-domain features; use a dimensionality reduction algorithm to perform dimensionality reduction processing on the key feature parameters to obtain a low-dimensional feature matrix; obtain the historical operation data of the equipment and the corresponding status labels, and construct a support vector machine classification model; divide the historical operation data into a training set and a test set, use the training set data to train the support vector machine classification model, and use the test set data to evaluate the performance of the trained support vector machine classification model; use the grid search method to optimize the hyperparameters of the support vector machine classification model, and select the parameter combination with the highest classification accuracy through cross-validation as the optimal parameters; input the low-dimensional feature matrix into the trained support vector machine classification model to perform fault diagnosis and classification of the equipment operation status; if the result of the fault diagnosis and classification is an abnormal state of the equipment, trigger an early warning mechanism and send a device inspection and maintenance notice to relevant personnel; continuously obtain the real-time operation data of the production equipment, regularly update the training data set of the support vector machine classification model, and retrain the support vector machine classification model.

[0063] Specifically, the real-time operation data of the production equipment can be obtained by installing sensors and a data acquisition system. For example, parameters such as temperature, pressure, and vibration are collected every 5 seconds. Perform preprocessing such as denoising and normalization on the collected raw data, and extract 20 key feature parameters such as mean, variance, and peak-to-peak value. Then use the principal component analysis algorithm to reduce the 20-dimensional feature data to 5 dimensions, with a cumulative contribution rate of over 95%. Use the production data for the past 1 year, a total of 10,000 pieces, including 8,000 pieces in the normal state and 2,000 pieces in the abnormal state, to construct a support vector machine binary classification model. Through 5-fold cross-validation and grid search, obtain the optimal Gaussian kernel function and penalty coefficient C = 10, and the accuracy of the model on the test set reaches 95%. Input the real-time collected and processed feature data into the trained support vector machine model. If the classification is an abnormal state, immediately send an alarm message to the equipment administrator, and at the same time display the location and diagnosis result of the abnormal equipment on the monitoring interface to guide the maintenance personnel to quickly locate and handle the fault. The system automatically updates the support vector machine model every day, performs incremental training using the newly collected data, and continuously improves the generalization ability and robustness of the model. After half a year of operation and testing, the equipment fault diagnosis system has achieved real-time fault detection and early warning of over 90%, and the average diagnosis time has been shortened from the previous 30 minutes to within 5 minutes, effectively improving the operation efficiency of the equipment and the product quality.

[0064] Step S109, generate a production status report according to the output result of the classifier, and guide the adjustment and optimization of the subsequent production process according to the production status report, including:

[0065] Obtain the real-time production data of production equipment sensors and the production management system, input the real-time production data into a pre-trained production status classification model to obtain the classification result of the current production status; according to the classification result, select the corresponding template from a preset production status report template library, fill in the real-time production data to generate an initial report; perform word segmentation, part-of-speech tagging and named entity recognition on the initial report, and extract the key information in the report; match the key information with a pre-configured production optimization rule library to generate optimization suggestions for the current production status; merge the optimization suggestions into the initial report to obtain a merged report; perform automatic summarization on the merged report, extract key sentences to generate a report summary; generate a chart presentation of the content of the merged report through a data visualization tool to obtain a visualized complete report; send the visualized complete report to relevant production management personnel and record the feedback information of the management personnel; adjust the production status report template library, key information extraction rules and production optimization rule library according to the feedback information; apply the production status classification model to real-time production data for continuous production status monitoring and early warning; re-train the production status classification model regularly according to the implementation effect of the optimization suggestions.

[0066] Specifically, after obtaining the output result of the classifier for the current production status, the system will generate a production status report that conforms to the management personnel's thinking logic according to the preset production status templates, such as indicators like production efficiency, equipment utilization rate, and product qualification rate. For example, if the current production efficiency is 85%, the equipment utilization rate is 90%, and the product qualification rate is 95%, the system will automatically generate a report: "The current production status is good, with relatively high production efficiency and equipment utilization rate, and stable product quality. It is recommended to maintain the existing production rhythm and continuously pay attention to equipment maintenance and quality control." Then, the system uses natural language processing techniques, such as the TF-IDF algorithm, to perform semantic analysis and key information extraction on the generated report content to ensure that the report content is concise and to the point, highlighting the key points. For example, keywords such as "production efficiency", "equipment utilization rate", and "product qualification rate" are extracted and given relatively high weights. Then, the extracted key information is matched with the preset production optimization rules. For example, when the production efficiency is lower than 80%, it is recommended to optimize the production process; when the equipment utilization rate is lower than 85%, it is recommended to increase the frequency of equipment maintenance, etc., to automatically generate targeted production process adjustment suggestions. The system uses visualization techniques to display the proportion of each indicator in a pie chart, the trend of indicators in a line chart, and the optimization suggestions in a flowchart, facilitating quick understanding and decision-making by management personnel. The generated report and optimization suggestions will be automatically sent to the email of relevant production management personnel, and their evaluations and opinions will be recorded through a feedback form for continuously optimizing the algorithm models for report generation and suggestion provision. The system will dynamically adjust the detail level of the report content and the pertinence of the optimization suggestions according to the feedback from management personnel, such as "The report content is too detailed and I hope it can be more concise", "The optimization suggestions are not specific enough and I hope to provide more operation details", etc., to improve the practicality and operability of the report and suggestions. At the same time, the system will continuously track the implementation situation and effects of the optimization suggestions. For example, if the production efficiency increases by 5% and the product qualification rate increases by 2%, etc., the implementation results will be used as new training data, and through machine learning algorithms, such as random forest, support vector machine, etc., the classifier and report generation algorithm will be iteratively optimized to form a closed-loop feedback mechanism, continuously improving the intelligent level and practical value of the system.

[0067] The present invention provides an on-line production monitoring system for modified mineral cables, mainly including:

[0068] A temperature data acquisition and correction module, which is used to acquire temperature data and pressure data in the extrusion stage, and correct the temperature data through a preset temperature compensation model to obtain the corrected temperature data;

[0069] A pressure data acquisition and correction module, which is used to establish a pressure drift model using the historical data of a pressure sensor and perform real-time correction on the current pressure data to obtain the corrected pressure data;

[0070] The vibration data denoising and feature extraction module is used to obtain the vibration data in the cooling stage and the stress data in the winding stage, denoise the vibration data through wavelet transform, extract the effective features in the vibration signal, and obtain the denoised vibration data;

[0071] The stress data segmented processing module is used to segment the data by using the sliding window algorithm according to the sampling frequency of the stress data, calculate the mean and variance of each segment of data, and obtain the stress data after segmented processing;

[0072] The multi-source data set alignment module is used to align the corrected temperature data, corrected pressure data, denoised vibration data, and stress data after segmented processing according to the time stamp, and generate a multi-source data set in a unified format;

[0073] The data dimensionality reduction processing module is used to perform dimensionality reduction processing on the multi-source data set by using the principal component analysis method, extract key features, and generate a low-dimensional feature matrix;

[0074] The real-time judgment and dynamic update module is used to judge whether the processing time of the low-dimensional feature matrix meets the real-time requirement through a preset real-time threshold. If the processing time exceeds the real-time threshold, the incremental principal component analysis method is used to dynamically update the data;

[0075] The classifier training module is used to input the low-dimensional feature matrix after real-time processing into the support vector machine model, train the classifier, and judge whether the current production state is within the normal range;

[0076] The production status report generation module is used to generate a production status report according to the output result of the classifier, and guide the adjustment and optimization of the subsequent production process according to the production status report.

[0077] The above are only the embodiments of the present invention, and do not limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made by using the content of the specification and drawings of the present invention, or directly or indirectly applied to other related technical fields, shall be equally included in the patent protection scope of the present invention.

Claims

1. An online production monitoring method for a modified mineral cable, characterized in that: The method comprises: Acquire the temperature data and pressure data in the extrusion stage, and correct the temperature data through a preset temperature compensation model to obtain corrected temperature data; The pressure drift model is established using the historical data of the pressure sensor, and the current pressure data is corrected in real time to obtain the corrected pressure data; Obtain vibration data in the cooling stage and stress data in the wrapping stage, perform denoising on the vibration data through wavelet transform, extract effective features in the vibration signal, and obtain denoised vibration data; According to the sampling frequency of stress data, the sliding window algorithm is used to process the data in sections, the mean and variance of each section of data are calculated, and the stress data after section processing is obtained; Align the corrected temperature data, corrected pressure data, denoised vibration data, and segmented stress data according to timestamps to generate a multi-source data set in a unified format; The principal component analysis method is used to reduce the dimensionality of multi-source data sets, extract key features, and generate a low-dimensional feature matrix; Through the preset real-time threshold, it is judged whether the processing time of the low-dimensional feature matrix meets the real-time requirement. If the processing time exceeds the real-time threshold, the incremental principal component analysis method is used to dynamically update the data. The low-dimensional feature matrix after real-time processing is input into the support vector machine model to train the classifier and determine whether the current production status is within the normal range; According to the output results of the classifier, a production status report is generated, and the adjustment and optimization of the subsequent production process are guided by the production status report.

2. The method according to claim 1, characterized in that: The step of acquiring temperature data and pressure data in the extrusion stage and correcting the temperature data by a preset temperature compensation model to obtain corrected temperature data comprises: Obtaining original temperature data and pressure data during the extrusion stage, and using the original temperature data and pressure data as inputs of a temperature compensation model; According to the preset temperature compensation model, the original temperature data is corrected and calculated to obtain the corrected temperature data; Determine whether the corrected temperature data is within the preset normal range. If the corrected temperature data exceeds the normal range, a temperature abnormality alarm is triggered.

3. The method according to claim 1, characterized in that: The method of using the historical data of the pressure sensor to establish a pressure drift model and correcting the current pressure data in real time to obtain the corrected pressure data includes: Read historical data of the pressure sensor from a database, preprocess the historical data, and obtain preprocessed historical data; Based on the pre-processed historical data, the LSTM neural network algorithm is used to build and train the pressure drift prediction model. The trained pressure drift prediction model is encapsulated as an API interface and deployed to the real-time data processing system. The real-time system calls the API to input the real-time data collected by the pressure sensor into the model to obtain the pressure drift prediction results; According to the pressure drift prediction result, the drift value is obtained and it is determined whether the current pressure sensor has drifted; If the drift value exceeds a preset threshold, the drift value is deducted from the real-time pressure value to obtain the corrected pressure data.

4. The method according to claim 1, characterized in that The step of obtaining vibration data in the cooling stage and stress data in the wrapping stage, denoising the vibration data by wavelet transform, extracting effective features in the vibration signal, and obtaining denoised vibration data includes: According to the vibration sensor in the cooling stage and the stress sensor in the wrapping stage, the vibration data in the cooling stage and the stress data in the wrapping stage are collected in real time; According to the timestamp information contained in the vibration data and the stress data, the data are time synchronized and aligned, and according to a preset stage division rule, the data is divided into two parts: a cooling stage and a wrapping stage; The vibration data in the cooling stage are processed by wavelet denoising to obtain the denoised vibration signal.

5. The method according to claim 1, characterized in that The method uses a sliding window algorithm to process the data in sections according to the sampling frequency of the stress data, calculates the mean and variance of each section of the data, and obtains the stress data after the section processing, including: Determine the size of the sliding window according to the sampling frequency; Using the sliding window to process the stress data in sections to obtain stress data after the section processing; For each of the stress data after the segment processing, the mean and variance are calculated to obtain the statistical characteristics of the stress data after the segment processing.

6. The method according to claim 1, characterized in that The corrected temperature data, the corrected pressure data, the denoised vibration data, and the segmented stress data are aligned according to timestamps to generate a multi-source data set in a unified format, including: Determine whether the timestamps of each data are consistent. If the timestamps are inconsistent, use linear interpolation to align the data so that the data are consistent in the time dimension. Based on the aligned temperature data, pressure data, vibration data and stress data, the Kalman filter algorithm is used to fuse multi-source data to obtain fused data in a unified format; For the fused data in a unified format, deduplication, denoising and outlier identification operations are performed to obtain a high-quality multi-source data set.

7. The method according to claim 1, characterized in that The principal component analysis method is used to reduce the dimension of the multi-source data set, extract key features, and generate a low-dimensional feature matrix, including: The principal component analysis method is used to preprocess the data and extract features, mapping the original high-dimensional data to a low-dimensional space, extracting the key features of the data and eliminating the differences caused by different data sources and formats; The preprocessed data is reduced in dimension, the principal components are obtained through principal component analysis, redundant and noisy data are removed, and a compressed low-dimensional feature matrix is ​​obtained; The compressed low-dimensional feature matrix is ​​used as the input of the support vector machine algorithm to train the classification model, and the original data is labeled according to the classification results of the support vector machine model; The annotated raw data is analyzed, and the analyzed multi-source heterogeneous data is converted into a low-dimensional feature matrix consistent with the model input format using the same preprocessing and feature extraction methods as the raw data.

8. The method according to claim 1, characterized in that The preset real-time threshold is used to determine whether the processing time of the low-dimensional feature matrix meets the real-time requirement. If the processing time exceeds the real-time threshold, the incremental principal component analysis method is used to dynamically update the data, including: Recording the processing time of the low-dimensional feature matrix by a timer, and determining whether the processing time meets a preset real-time threshold requirement; If the processing time of the low-dimensional feature matrix exceeds the preset real-time threshold, the low-dimensional feature matrix is ​​dynamically updated using the incremental principal component analysis method according to the data volume and data distribution characteristics of the low-dimensional feature matrix; The updated low-dimensional feature matrix is ​​used as new feature data, and the principal components of the low-dimensional feature matrix are updated through multiple iterations. When the processing time of the low-dimensional feature matrix is ​​lower than the preset real-time threshold for multiple consecutive times, the iterative update is stopped.

9. The method according to claim 1, characterized in that: The low-dimensional feature matrix after real-time processing is input into the support vector machine model, and the classifier is trained to determine whether the current production status is within the normal range, including: Acquire real-time operating data of production equipment, including key parameters of temperature, pressure and vibration of the equipment; Preprocessing the real-time operation data to obtain preprocessed data; Extracting key feature parameters from the preprocessed data, and performing dimensionality reduction processing on the key feature parameters using a dimensionality reduction algorithm to obtain a low-dimensional feature matrix; Obtain the historical operation data and corresponding status labels of the equipment and build a support vector machine classification model; Dividing the historical operation data into a training set and a test set, using the training set data to train the support vector machine classification model, and using the test set data to perform performance evaluation on the trained support vector machine classification model; Inputting the low-dimensional feature matrix into the trained support vector machine classification model to perform fault diagnosis and classification of the equipment operation status; If the result of the fault diagnosis and classification is that the equipment is in an abnormal state, an early warning mechanism is triggered to send an equipment inspection and maintenance notice to relevant personnel; Continuously acquire real-time operation data of production equipment, regularly update the training data set of the support vector machine classification model, and retrain the support vector machine classification model.

10. An online production monitoring system for modified mineral cables, characterized in that: The system comprises: The temperature data acquisition and correction module is used to acquire the temperature data and pressure data of the extrusion stage, and correct the temperature data through a preset temperature compensation model to obtain the corrected temperature data; The pressure data acquisition and correction module is used to establish a pressure drift model using the historical data of the pressure sensor, perform real-time correction on the current pressure data, and obtain the corrected pressure data; The vibration data denoising and feature extraction module is used to obtain the vibration data of the cooling stage and the stress data of the wrapping stage, denoise the vibration data through wavelet transform, extract the effective features in the vibration signal, and obtain the denoised vibration data; The stress data segmentation processing module is used to segment the data using a sliding window algorithm according to the sampling frequency of the stress data, calculate the mean and variance of each segment of data, and obtain the stress data after segmentation processing; The multi-source data set alignment module is used to align the corrected temperature data, the corrected pressure data, the denoised vibration data, and the segmented stress data according to the timestamps to generate a multi-source data set in a unified format; The data dimension reduction processing module is used to reduce the dimension of multi-source data sets using principal component analysis, extract key features, and generate a low-dimensional feature matrix; The real-time judgment and dynamic update module is used to judge whether the processing time of the low-dimensional feature matrix meets the real-time requirements through the preset real-time threshold. If the processing time exceeds the real-time threshold, the incremental principal component analysis method is used to dynamically update the data; The classifier training module is used to input the low-dimensional feature matrix after real-time processing into the support vector machine model to train the classifier and determine whether the current production status is within the normal range; The production status report generation module is used to generate a production status report according to the output results of the classifier, and guide the adjustment and optimization of the subsequent production process according to the production status report.

Images