Rubber tube production line monitoring method and system

Abstract

The invention relates to the technical field of industrial automation and intelligent manufacturing, in particular to a rubber tube production line monitoring method and system.The method comprises the steps that multi-mode production data of a production line is collected and preprocessed, a virtual production line model is generated through a dynamic twin model and compared with actual data, and virtual-real difference data is generated; performing anomaly detection on the virtual-real difference data based on a classification algorithm and a rule matching structure, and optimizing parameters of an anomaly detection model; generating abnormal trend prediction data by using the optimized model, and proposing a system correction suggestion according to the prediction data; and dynamically adjusting the operation parameters of the equipment through a closed-loop feedback mechanism. According to the invention, through innovative dynamic twin modeling, deep learning optimization and closed-loop control technologies, real-time monitoring and intelligent adjustment of the production line are realized, and the production efficiency and the product quality are effectively improved.

Description

Technical Field

[0001] The present invention relates to the technical field of industrial automation and intelligent manufacturing, and in particular to a hose production line monitoring method and system. Background Art

[0002] As an indispensable transmission medium in the fields of industry, agriculture and transportation, the quality and production efficiency of hoses directly affect the operating efficiency and product safety of related industries. Hose production line monitoring can ensure the stability of product quality and reduce resource waste and production stagnation caused by abnormalities through real-time detection and adjustment of key parameters in the production process, while improving the level of production automation and reducing the need for human intervention.

[0003] The existing hose production line data detection method (Chinese invention patent, publication number: CN118468008A, name: A real-time detection method and device for hose production line data) mainly relies on PCA dimensionality reduction analysis of multi-dimensional production data, and achieves data dimensionality reduction and anomaly detection by extracting the eigenvalues ​​and eigenvectors of the covariance matrix. However, the core defects of this solution are:

[0004] The PCA dimensionality reduction algorithm has a weak ability to handle noisy data or mutation data in the production environment, which may reduce the credibility of the eigenvalues ​​and thus affect the reliability of the dimensionality reduction results. Selecting eigenvectors as principal components based only on the size of the eigenvalues ​​ignores the global characteristics of the data distribution and may miss key information related to actual anomalies. The multi-step feature calculation and verification in the dimensionality reduction process leads to insufficient real-time data feedback and adjustment speed, which cannot quickly respond to emergencies in production. Summary of the invention

[0005] In view of the many problems existing in the above-mentioned prior art, the present invention provides a hose production line monitoring method and system. The present invention collects multimodal data of the hose production line and uses a dynamic twin model to perform virtual-real comparison analysis on the actual production status. Based on the virtual-real difference data, the classification algorithm and rule matching model are used to achieve accurate detection of production anomalies, and the model parameters are optimized through deep learning. Finally, combined with a closed-loop feedback mechanism, the equipment parameters are dynamically adjusted, which significantly improves the operating efficiency, product quality and adaptability of the production line.

[0006] A hose production line monitoring method comprises the following steps:

[0007] Collect multimodal production data of the hose production line, pre-process the collected data, and generate fused multimodal production data;

[0008] Input the fused multimodal production data into the digital twin modeling module to generate a dynamic twin model; compare the actual production line data with the dynamic twin model to generate virtual-real difference data;

[0009] Anomaly detection is performed based on virtual-real difference data. The virtual-real difference data is analyzed using an anomaly detection model. The anomaly detection model includes a structure based on a classification algorithm and rule matching, which is used to mark and classify abnormal points and generate classified abnormal data; the parameters of the anomaly detection model are optimized, where the optimization method includes model weight adjustment based on deep learning and threshold dynamic correction based on real-time data feedback, and the optimized parameters are fed back to the digital twin model to update its simulation rules;

[0010] Generate abnormal trend prediction data based on classified abnormal data; generate system correction suggestions based on abnormal trend prediction data, and adjust the production line operation through a closed-loop feedback mechanism, where the closed-loop feedback mechanism includes real-time adjustment of the temperature, pressure and flow rate parameters of the production equipment according to the prediction results, and generation of control signals to dynamically correct the equipment operation status.

[0011] Preferably, the preprocessing of the multimodal production data comprises the following steps:

[0012] Perform edge detection on the visual data of the hose production line and extract the surface texture features of the visual data;

[0013] Normalize the temperature data and pressure data of the hose production line respectively;

[0014] The vibration data of the hose production line is processed by frequency domain filtering.

[0015] Preferably, the dynamic twin model includes:

[0016] Flow field model, which generates flow field characteristic data through three-dimensional flow field simulation based on computational fluid dynamics algorithm;

[0017] The geometric model is generated by a point cloud-based reconstruction algorithm to describe the state of the hose surface.

[0018] Preferably, the virtual-real difference data includes the following contents:

[0019] Flow field abnormal data, including velocity gradient mutation data and pressure abnormal point data;

[0020] Geometric deviation data, including surface roughness deviation data and crack extension area data;

[0021] Temperature and pressure abnormal data, including abnormal range data and change trend data.

[0022] Preferably, the classification algorithm of the anomaly detection model includes the following contents:

[0023] Use K-means clustering algorithm to cluster the abnormal points in the virtual-real difference data and generate abnormal category data;

[0024] Use the random forest classification algorithm to predict the category of abnormal points in the virtual-real difference data and generate classified abnormal data;

[0025] The rule matching structure corrects the classified abnormal data through dynamic detection rules based on preset thresholds.

[0026] Preferably, the method for optimizing the anomaly detection model parameters comprises:

[0027] Dynamically adjust the location of the K-means clustering center based on real-time feedback data;

[0028] Update the depth and node splitting rules of decision trees in the random forest classification algorithm;

[0029] Adjust classification thresholds to improve the sensitivity and accuracy of anomaly detection;

[0030] The weight adjustment formula of the dynamic weighted feature fusion algorithm is as follows:

[0031]

[0032] Among them, W i represents the weight of the i-th feature; F i represents the real-time input value of the i-th feature; P i represents the importance coefficient of the i-th feature; n represents the total number of features.

[0033] Preferably, the abnormal trend prediction data is generated by the following steps:

[0034] Based on the Transformer time series model, the abnormal propagation path data in the virtual and real difference data is analyzed in time series;

[0035] Extract features of dynamic abnormal field data in virtual-real difference data, and predict the abnormal diffusion range and key abnormal points;

[0036] Output abnormal trend prediction data, including time, space and intensity distribution.

[0037] Preferably, the system correction suggestion includes the following adjustment parameters:

[0038] Adjust the temperature control range of production equipment to reduce the risk of local overheating;

[0039] Adjust the flow rate distribution optimization parameters of the production equipment to balance the fluid pressure in the production line;

[0040] Adjust the pressure regulation range of the production equipment to prevent the pressure from exceeding the limit during the operation of the production equipment.

[0041] Preferably, the closed-loop feedback mechanism comprises the following steps:

[0042] Collect the status data of the production line after adjustment in real time and verify the system correction suggestions;

[0043] Dynamically update the operating parameters of production equipment to optimize the operating status of production equipment;

[0044] New operating rules are formed based on the verification results to complete the self-learning optimization cycle.

[0045] A system for implementing the hose production line monitoring method, comprising:

[0046] A multimodal data acquisition module is configured to collect visual data, temperature data, pressure data, vibration data and environmental data of the hose production line, and perform time synchronization, normalization and denoising on the collected multimodal production data to generate fused multimodal production data;

[0047] A digital twin modeling module is configured to receive the fused multimodal production data, generate a dynamic twin model, and compare the production data of the actual production line with the dynamic twin model to generate virtual-real difference data;

[0048] An anomaly detection module is configured to perform anomaly detection based on the virtual-real difference data, and the anomaly detection module includes:

[0049] A classification module based on classification algorithms is used to analyze virtual and real difference data and classify outliers;

[0050] A rule-matching-based detection module is used to mark abnormal points according to preset rules;

[0051] An optimization module is used to optimize the parameters of the anomaly detection model through model weight adjustment based on deep learning and dynamic correction of thresholds based on real-time data feedback, and feed the optimized parameters back to the digital twin modeling module to update its simulation rules;

[0052] An abnormal trend prediction module is configured to generate abnormal trend prediction data based on the classified abnormal data, and the abnormal trend prediction module predicts the abnormal diffusion range, key abnormal points and time, space and intensity distribution through a time series analysis model;

[0053] A closed-loop feedback control module is configured to generate system correction suggestions based on abnormal trend prediction data and adjust the production line operation through a closed-loop feedback mechanism, wherein the closed-loop feedback mechanism includes:

[0054] Adjust the temperature parameters, pressure parameters and flow rate parameters of production equipment in real time;

[0055] Generate control signals for dynamically correcting the operating status of production equipment;

[0056] Optimize system correction suggestions through real-time data verification and dynamic updates.

[0057] Compared with the prior art, the advantages and beneficial effects of the present invention are:

[0058] The present invention improves the modeling and evaluation accuracy of the real-time status of the production line by introducing a dynamic twin model and virtual-real comparison analysis technology, and solves the problem of inaccurate detection caused by noise data in the prior art.

[0059] The present invention utilizes an anomaly detection model based on a combination of classification algorithms (such as K-means clustering and random forest classification) and rule matching to effectively mark and classify production anomalies, thereby achieving efficient and accurate anomaly identification;

[0060] The present invention optimizes model weights through deep learning and dynamically corrects thresholds through real-time feedback, thereby improving the sensitivity and robustness of anomaly detection and ensuring the stability of the model in complex environments.

[0061] The present invention establishes a closed-loop feedback mechanism to adjust the equipment operating parameters in real time according to the abnormal trend prediction data, thereby optimizing the operating efficiency and product quality of the production line;

[0062] The present invention uses the Transformer time series model to predict abnormal trends, ensuring accurate control of the spatiotemporal dynamics of abnormal diffusion, and further improving the initiative and foresight of the production process. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] Figure 1 It is a schematic diagram of the method flow of the present invention;

[0064] Figure 2 It is a schematic diagram of the abnormality detection and optimization process of the present invention;

[0065] Figure 3 It is a schematic diagram of abnormal trend prediction and system correction of the present invention;

[0066] Figure 4 It is a structural block diagram of the system of the present invention. DETAILED DESCRIPTION

[0067] Hereinafter, embodiments of the present disclosure will be described with reference to the accompanying drawings. However, it should be understood that these descriptions are exemplary only and are not intended to limit the scope of the present disclosure. In the following detailed description, for ease of explanation, many specific details are set forth to provide a comprehensive understanding of the embodiments of the present disclosure.

[0068] like Figure 1 As shown, a hose production line monitoring method comprises the following steps:

[0069] Collect multimodal production data of the hose production line, pre-process the collected data, and generate fused multimodal production data;

[0070] Preferably, the preprocessing of the multimodal production data comprises the following steps:

[0071] Perform edge detection on the visual data of the hose production line and extract the surface texture features of the visual data;

[0072] Normalize the temperature data and pressure data of the hose production line respectively;

[0073] The vibration data of the hose production line is processed by frequency domain filtering.

[0074] During the hose production process, visual data is mainly used to monitor the quality of the hose surface. An industrial camera is used to obtain an image of the hose surface, and the edge information of the hose surface is extracted through an edge detection algorithm. Edge detection is a common image processing technology that extracts the boundary information of an object by calculating the change in pixel gradient. In the present invention, edge detection is used to identify the texture features of the hose surface, such as possible abnormal areas such as dents, scratches or cracks. In addition, the surface detail information, such as roughness distribution and local structural features, is further analyzed through a texture feature extraction algorithm. These feature data provide accurate geometric information and a basis for surface quality assessment for subsequent modeling.

[0075] Temperature and pressure data are collected in real time by sensors installed at key locations on the production line. These data directly reflect the operating status of the production equipment and the environmental conditions inside and outside the hose. However, due to the different dimensions and ranges of data from different sensors, these data need to be normalized in order to eliminate the dimensional differences between the data. Normalization maps the raw data to a uniform range (such as [0,1]) through linear transformation, which facilitates the fusion and unified analysis of multimodal data. The core of this step is to ensure the comparability and consistency of various types of data to improve the accuracy of subsequent feature fusion.

[0076] In hose production, vibration data is mainly used to reflect the mechanical state and running stability of the equipment. The collected original vibration signal usually contains a lot of high-frequency noise and environmental interference, so the signal needs to be denoised through frequency domain filtering. Frequency domain filtering is a signal processing method based on frequency decomposition. In the present invention, a bandpass filter is used to extract the effective frequency range in the mechanical vibration signal, such as the natural frequency of the equipment operation and the typical fault characteristic frequency. The filtered vibration signal can not only reflect the normal operating status of the equipment, but also identify possible abnormal vibration modes, providing data support for equipment health status monitoring and fault prediction.

[0077] The preprocessing of the above three types of data is achieved through a distributed data acquisition and processing system. Each type of data acquisition unit and its corresponding preprocessing module operate independently. After synchronization and calibration, the data of each module are uniformly transmitted to the fusion module. The preprocessed multimodal production data is output in a structured format, providing high-quality input data for the construction of the digital twin model.

[0078] The noise and interference of different types of data are eliminated to ensure the accuracy and consistency of the data; the normalization and standardization processing methods lay the foundation for subsequent data fusion and ensure the effective combination of multi-source data at the feature level; the feature extraction of various types of data provides comprehensive information support for the monitoring of hose production status, including surface quality, equipment operating status and environmental parameters.

[0079] Example, in actual application, a hose production line is equipped with a variety of sensors and industrial cameras. For example, in the process of visual data acquisition, the industrial camera collects high-definition images of the hose surface at a rate of 50 frames per second, extracts texture features and detects surface cracks through Sobel edge detection; the temperature and pressure sensors collect the operating parameters of each key part of the production line at intervals of 1 second, and the data change trend is clearly visible after normalization; the vibration sensor extracts the characteristic frequency of the vibration signal through fast Fourier transform, and successfully identifies the abnormal frequency peak caused by bearing damage during equipment operation. After the above processing, the fused multimodal production data was successfully used for real-time simulation of the digital twin model, discovering and locating potential problems in production, and improving production efficiency and product quality.

[0080] Input the fused multimodal production data into the digital twin modeling module to generate a dynamic twin model; compare the actual production line data with the dynamic twin model to generate virtual-real difference data;

[0081] Preferably, the dynamic twin model includes:

[0082] Flow field model, which generates flow field characteristic data through three-dimensional flow field simulation based on computational fluid dynamics algorithm;

[0083] The geometric model is generated by a point cloud-based reconstruction algorithm to describe the state of the hose surface.

[0084] Industrial cameras are deployed at key locations on the production line to capture high-resolution images of the hose surface in real time. The significant boundaries in the image are identified through edge detection algorithms (such as Sobel operator or Canny edge detection) to extract edge information of the hose surface structure. Subsequently, texture analysis methods (such as gray-level co-occurrence matrix or local binary pattern) are used to further extract surface texture features, which can reflect the quality status of the hose surface, such as whether there are cracks, dents, or changes in roughness.

[0085] Assuming that the input visual image is I(x,y), the edge strength is obtained by calculating the amplitude of the image gradient G(x,y):

[0086]

[0087] Where x and y represent the horizontal and vertical pixel positions of the image, and are the gradients of the image in the horizontal and vertical directions respectively. By performing threshold processing on the edge intensity, significant boundaries are extracted.

[0088] The generated surface texture feature data can be used to identify surface defect areas and provide geometric information input for subsequent digital twin modeling. Through edge detection and texture extraction, the abnormal position of the hose surface can be accurately located.

[0089] Temperature and pressure data are collected in real time by sensors placed on the production line. These data directly reflect the environmental conditions of the equipment. However, due to the different dimensions and ranges of sensors, the data needs to be normalized to achieve cross-modal fusion. Normalization uses a linear scaling formula to normalize the data to the [0,1] interval:

[0090]

[0091] Among them, X norm represents the normalized data; X represents the original data; X min and X max Indicates the minimum and maximum values ​​of the original data.

[0092] The input data collected by the temperature sensor is T(t), and the input data collected by the pressure sensor is P(t). T(t) and P(t) are normalized according to the above formula. This eliminates the data inconsistency caused by different dimensions and ranges, facilitates the effective fusion of multimodal features in subsequent steps, and ensures the accuracy and comparability of the analysis results.

[0093] Vibration data is collected by the acceleration sensor installed on the equipment to reflect the operating status of the mechanical equipment. However, the original vibration signal usually contains a lot of noise (such as high-frequency electromagnetic interference or environmental noise). The present invention removes invalid frequency band signals through a bandpass filter and only retains the target frequency band (for example, the natural frequency range of the equipment or the typical fault frequency range).

[0094] Assume that the input vibration signal is V(t), and its spectrum is calculated as V(f) through fast Fourier transform (FFT). Apply the frequency response function H(f) of the bandpass filter to obtain the filtered signal:

[0095]

[0096] in, represents the inverse Fourier transform; H(f) represents the frequency response function of the bandpass filter, which is used to limit the frequency range of the signal.

[0097] Set the frequency range of the bandpass filter to [f low ,f high ], where f low and f high Represent the low-frequency and high-frequency boundaries respectively, and remove the frequencies below f through the filter. low and above f high The frequency domain filtering of vibration data ensures signal quality and can extract effective characteristic frequencies in the equipment operation status to help identify potential faults such as bearing wear or vibration imbalance.

[0098] The multimodal production data preprocessing of the present invention achieves the following goals: eliminating noise, dimensional differences and redundant signals to ensure data accuracy and consistency; providing high-dimensional, high-quality data input for subsequent modeling through characterization processing of visual, temperature, pressure and vibration data; and achieving comprehensive monitoring of production status through precise extraction of surface texture features, environmental parameter change trends and mechanical properties, providing a solid data foundation for anomaly detection.

[0099] Example, in actual application, a multimodal data acquisition device is deployed in a hose production line. The surface image of the hose is collected at 30 frames per second by an industrial camera, and the Sobel operator is used to extract edge information and detect surface cracks; the temperature sensor and pressure sensor are installed at the production equipment and pipeline interface respectively, and the data is collected at a period of 5 seconds. After normalization, the temperature range is standardized to [0,1]; the vibration sensor monitors the vibration signal of the equipment during operation, and the bandpass filter limits the frequency range to 50-200Hz, removing high-frequency electromagnetic interference. The preprocessed data is input into the digital twin model, realizing high-precision simulation of the production process, and identifying local overheating and surface crack problems in hose production through the anomaly detection module.

[0100] like Figure 2 As shown, anomaly detection is performed based on virtual-real difference data, and the virtual-real difference data is analyzed using an anomaly detection model. The anomaly detection model includes a structure based on a classification algorithm and rule matching, which is used to mark and classify abnormal points and generate classified abnormal data; the parameters of the anomaly detection model are optimized, wherein the optimization method includes model weight adjustment based on deep learning and threshold dynamic correction based on real-time data feedback, and the optimized parameters are fed back to the digital twin model to update its simulation rules;

[0101] The virtual-real difference data is generated by comparing the actual operation data of the hose production line with the simulation results of the dynamic twin model, reflecting the deviation between reality and theory. The anomaly detection model analyzes these data by combining classification algorithms with rule matching. Classification algorithms (such as K-means clustering and random forest classification) identify and mark anomalies by extracting features from the virtual-real difference data. The rule matching structure further verifies the marked anomalies and determines the type and severity of the anomalies based on preset dynamic detection rules (such as deviation threshold or change rate).

[0102] Assume that the difference between the real and the virtual data is D = {d 1 ,d 2 ,…,d n}, where d i Represents the i-th difference point. First, the data is clustered by the K-means clustering algorithm, the data is divided into k categories, and the abnormal categories that deviate from the center point are marked; then the random forest classification algorithm is used to further classify these abnormal points to generate classified abnormal data. The rule matching structure verifies the classification results, and according to the set dynamic threshold rule T dynamic , determine a difference point d i Whether it belongs to a specific exception type.

[0103] The performance of the anomaly detection model depends on the parameter settings of the classification algorithm (such as cluster center location, classification rules, etc.) and the dynamic threshold of the rule matching structure. In order to improve the accuracy and robustness of detection, the present invention introduces a deep learning method to dynamically adjust the weight of the model and optimize the rule threshold through real-time data feedback. The model weight adjustment is implemented through a deep neural network, which takes the virtual and real difference data as input and iteratively updates the model parameters. The real-time data feedback mechanism dynamically adjusts the rule threshold by analyzing the new difference data, so that the threshold can adapt to changes in the production process.

[0104] Let the loss function of the deep learning network be L(w), where w represents the model weight. Update the weight by gradient descent:

[0105]

[0106] Among them, w t represents the current weight; η represents the learning rate; Represents the gradient of the loss function with respect to the weight. For the dynamic threshold T dynamic , the real-time data feedback mechanism readjusts the threshold range according to the distribution characteristics of the new data:

[0107] T dynamic =T base +α·ΔD

[0108] Among them, T base represents the basic threshold; α represents the feedback coefficient; ΔD represents the variation range of the difference data.

[0109] The optimized anomaly detection model parameters are fed back to the digital twin model to update its simulation rules. By correcting the anomalies in the virtual-real comparison, the accuracy and adaptability of the digital twin model are further improved, thereby better reflecting the actual operating status of the production line. The update of the digital twin model is based on the optimized model weights and dynamic thresholds, and the feature weighting rules and simulation accuracy parameters of the twin model are readjusted, so that the model can capture dynamic changes in production in real time.

[0110] The detection method that combines classification algorithms with rule matching can effectively identify various types of abnormal conditions, including equipment failures, parameter out-of-limit, and operating deviations. The optimized classified abnormal data improves the accuracy and robustness of abnormality detection. Real-time optimization of model parameters enables the detection model to adapt to the changing operating conditions in the production line, reduce the probability of false detection and missed detection, and ensure the ability to identify new abnormal types. The optimized parameters update the digital twin model through a feedback mechanism to ensure the accuracy and timeliness of the simulation results, providing high-quality data support for subsequent production optimization.

[0111] Embodiment, in actual application, the virtual and real difference data of a hose production line include surface geometric deviations, flow field anomalies and environmental parameter anomalies. The K-means algorithm is used to cluster the difference data and divide the anomalies into geometric deviation, flow field anomaly and temperature anomaly; the random forest algorithm further predicts the specific categories of these anomalies (such as crack extension and local overheating). During the optimization stage, the deep learning network dynamically adjusted the parameters of the classification model, and the false detection rate dropped by 15%; the real-time data feedback mechanism adjusted the threshold of the temperature anomaly from a fixed value to a range dynamically updated based on environmental changes, thereby improving the sensitivity of anomaly detection. Finally, the optimized detection results are fed back to the digital twin model, enabling it to simulate the dynamic state of the production line in real time, significantly improving the reliability and stability of the production process.

[0112] Preferably, the virtual-real difference data includes the following contents:

[0113] Flow field abnormal data, including velocity gradient mutation data and pressure abnormal point data;

[0114] Geometric deviation data, including surface roughness deviation data and crack extension area data;

[0115] Temperature and pressure abnormal data, including abnormal range data and change trend data.

[0116] Flow field anomaly data mainly include velocity gradient mutation data and pressure abnormal point data, which are used to monitor the dynamic state of the fluid inside and outside the hose. Flow field anomaly detection calculates the spatial gradient changes of velocity and pressure to identify possible abnormal areas, such as vortex areas with sudden changes in velocity or blockage points with sudden pressure increases.

[0117] The velocity gradient mutation data is obtained by calculating the gradient of the velocity field:

[0118]

[0119] in, is the velocity gradient vector, v x 、v y 、v z They represent the components of flow velocity in the x, y, and z directions respectively.

[0120] The pressure anomaly data is used to calculate the local extreme value through the distribution function of the pressure field:

[0121] P max / min =max / min(P(x,y,z))

[0122] Among them, P(x, y, z) represents the pressure value in three-dimensional space.

[0123] After real-time acquisition of flow field data, gradient mutations and pressure extremes are identified through numerical calculations. When the set threshold is exceeded, it is marked as a mutation area; when the pressure value P of a certain point in the pressure distribution is significantly higher than the average value of the neighborhood, it is marked as an abnormal point. Flow field abnormality data can help detect pipeline blockage, fluid turbulence or overpressure risks in the production process, and provide a basis for timely adjustment of production parameters.

[0124] Geometric deviation data includes surface roughness deviation data and crack propagation area data, which are used to monitor the geometric characteristics of the hose surface. Roughness deviation is extracted through surface texture analysis, such as calculating the standard deviation or mean square deviation of surface height; crack propagation area is calculated by the growth rate of crack boundaries in the visual image.

[0125] The surface roughness deviation can be calculated by the following formula:

[0126]

[0127] Among them, R q Represents the root mean square value of roughness; h i Represents the height of the i-th sampling point on the surface; represents the average surface height; N represents the total number of sampling points.

[0128] The crack extension area uses image segmentation technology to identify the crack boundary and calculate its time series growth rate:

[0129]

[0130] Among them, G c represents the crack growth rate; ΔA represents the change in the crack area; Δt represents the time interval.

[0131] An industrial camera is used to collect the surface image of the hose, and the crack feature points are extracted by combining the image processing algorithm to analyze the crack extension area and surface roughness deviation in real time. q Exceeding the standard range or crack growth rate G c When the set threshold is exceeded, it is marked as a geometric anomaly. Geometric deviation data can effectively identify wear, damage or potential failure problems on the hose surface, ensuring that product quality meets design requirements.

[0132] Temperature and pressure anomaly data mainly include anomaly range data and change trend data, which reflect the operating status of production equipment and environment. Temperature anomalies are usually caused by uneven heat distribution or insufficient heat dissipation, while pressure anomalies may be caused by equipment overload or fluid blockage. Anomaly range data calculates extreme value areas through spatially distributed temperature and pressure fields; change trend data predicts future change trends through time series analysis.

[0133] The abnormal range data is calculated through spatial interpolation to find the out-of-limit area. For example, when the temperature T(x,y,z) of a certain area exceeds the set threshold T threshold When the temperature and pressure exceed the limit, it is marked as an abnormal area. The trend data is used to predict the future temperature and pressure changes through time series analysis (such as ARIMA model or sliding average method) to generate trend data. The temperature and pressure abnormal data can identify local overheating or pressure exceeding the limit in the production process, providing a timely basis for adjusting the equipment operating parameters.

[0134] Abnormal flow field data helps identify fluid dynamics problems, geometric deviation data locates surface defects, and abnormal temperature and pressure data reveal the equipment operating status. Providing abnormal range and trend data provides a basis for optimizing production equipment parameters and improving operating strategies. Detailed differential data classification and quantitative analysis ensure the sensitivity and reliability of the monitoring system, significantly reducing missed detection and false detection rates.

[0135] Embodiment, in actual application, the virtual and real difference data of a hose production line includes the following contents:

[0136] In the flow field data, the velocity gradient in the middle section of the pipeline was detected Far exceeds the set threshold of 8.0m / s 2 , which was marked as a sudden change area of ​​flow velocity; at the same time, an abnormal pressure point P = 10.2MPa was detected at the pipeline joint, exceeding the design limit of 9.5MPa.

[0137] In surface inspection, the root mean square value of the roughness of a hose section is R q =0.75μm, which obviously exceeds the standard value of 0.50μm. At the same time, the crack extension area increases by G per minute. c =0.12cm 2 / min.

[0138] In the temperature field, the temperature distribution range of a certain equipment detection point is [80,92]℃, which exceeds the upper limit of 85℃. The prediction model shows that the temperature will further rise to 95℃ in the next 15 minutes.

[0139] Through the analysis of the above data, the production line timely adjusted the fluid flow rate, reduced the pipeline pressure, and stopped the production to replace the abnormal hose section, thus avoiding potential production accidents.

[0140] Preferably, the classification algorithm of the anomaly detection model includes the following contents:

[0141] Use K-means clustering algorithm to cluster the abnormal points in the virtual-real difference data and generate abnormal category data;

[0142] Use the random forest classification algorithm to predict the category of abnormal points in the virtual-real difference data and generate classified abnormal data;

[0143] The rule matching structure corrects the classified abnormal data through dynamic detection rules based on preset thresholds.

[0144] K-means clustering is an unsupervised learning algorithm that finds the intrinsic distribution structure between data points by dividing the data points into k clusters. In the present invention, the K-means clustering algorithm is used to perform cluster analysis on the abnormal points in the virtual-real difference data, and generate abnormal category data according to the characteristic value distribution of the data points. For example, flow rate mutation points, pressure abnormal points and geometric deviation points can be classified into different categories for subsequent processing.

[0145] The update of the cluster center is achieved by minimizing the sum of the squares of the distances from the data points to the cluster center. The objective function is:

[0146]

[0147] Where J represents the objective function; represents the jth data point belonging to the i-th class; μ i represents the cluster center of the i-th category; n i represents the number of data points in the i-th class; k represents the number of clustering categories.

[0148] In the implementation process, it is assumed that the input virtual-real difference data D = {d 1 ,d 2 ,…,d n}, each data point d i Including multi-dimensional features (such as flow rate, pressure, roughness, etc.). By setting k = 3, corresponding to the flow rate, pressure and geometric anomaly categories respectively, the K-means algorithm is used to calculate the cluster center and assign category labels.

[0149] The K-means clustering algorithm can quickly classify outliers into different categories according to feature similarity, provide basic input data for the random forest classification algorithm, and improve the robustness of outlier classification.

[0150] Random forest is an ensemble learning algorithm based on decision trees, which performs classification by constructing multiple decision trees and taking majority votes. In the present invention, the random forest classification algorithm uses the results of K-means clustering and the eigenvalues ​​of the virtual-real difference data to further accurately predict the category of each outlier.

[0151] The core of random forest is to improve the generalization ability of the model by introducing randomness of data subsets and feature subsets. The generation of each decision tree follows the optimal partitioning principle of information gain or Gini coefficient, and the classification result is determined by the voting results of all trees.

[0152] Assume that for each data point d i The input feature is F = {f 1 ,f 2 ,…,f m}, where m represents the feature dimension. The data after K-means clustering is used as the training set, and the classification model is trained by the random forest algorithm. i Predict its abnormal category C i The random forest classification algorithm improves classification accuracy through the integration of multiple decision trees and has strong anti-interference ability for noisy data. The classification results can accurately mark the categories of abnormal points and provide more reliable data for the correction of rule matching structure.

[0153] The rule matching structure corrects the classified abnormal data through the preset dynamic detection rules to ensure the accuracy of the classification results. In the present invention, the dynamic detection rules are adjusted according to the real-time changes of the virtual and real difference data. Common rules include deviation threshold, change rate and neighborhood consistency. For example, if the velocity deviation value Δv of a classified abnormal point exceeds the threshold T v , then further verify whether it is a mutation point; if the characteristic value distribution of its neighborhood points is inconsistent, re-evaluate its category.

[0154] Assume that the classified abnormal data is A={a 1 ,a 2 ,…,a p}, each data point a i Contains its category C i and the eigenvalue F i The rule matching structure is corrected according to the following rules:

[0155] If Δv i >T v , then verify the flow rate mutation condition; if the classification of the classification neighborhood points is inconsistent, readjust the classification. The adjustment of dynamic rules is based on real-time feedback data to optimize the threshold T and conditional rules.

[0156] The rule matching structure enhances the credibility of classification results, reduces misclassification, and improves the applicability of classified abnormal data in subsequent processing.

[0157] The K-means algorithm effectively clusters abnormal points, and the random forest further refines the classification results, ensuring accurate identification of abnormal categories; the rule matching structure corrects the classification results through dynamic rules, enabling the model to adapt to real-time changes in the production line; the combination of ensemble learning and rule matching significantly improves the robustness and anti-interference ability of classification.

[0158] Embodiment: In a hose production line, the comparison between the actual operation data and the digital twin model generates virtual-real difference data. Through K-means clustering, the abnormal points are divided into three categories: velocity anomaly, pressure anomaly and geometric anomaly; random forest classification further subdivides the velocity anomaly into eddy current area and mutation area, and subdivides the pressure anomaly into overpressure point and low pressure point. The rule matching structure corrects some misclassified eddy current area points according to the threshold rule, and the accuracy of the optimized classified abnormal data reaches 95%. Ultimately, these classification results are used to guide the adjustment of production equipment, successfully avoiding downtime accidents caused by pipeline blockage.

[0159] Preferably, the method for optimizing the anomaly detection model parameters comprises:

[0160] Dynamically adjust the location of the K-means clustering center based on real-time feedback data;

[0161] Update the depth and node splitting rules of decision trees in the random forest classification algorithm;

[0162] Adjust classification thresholds to improve the sensitivity and accuracy of anomaly detection;

[0163] The weight adjustment formula of the dynamic weighted feature fusion algorithm is as follows:

[0164]

[0165] Among them, W i represents the weight of the i-th feature; F i represents the real-time input value of the i-th feature; P i represents the importance coefficient of the i-th feature; n represents the total number of features.

[0166] The performance of the K-means clustering algorithm depends on the location of the initial cluster center. In the actual production process, the distribution of virtual and real difference data will be dynamically adjusted as production conditions change. Therefore, it is necessary to dynamically update the location of the cluster center based on real-time feedback data to ensure the accuracy of the clustering results.

[0167] The position update formula of the cluster center point is based on minimizing the distance between the data point and the cluster center:

[0168]

[0169] Among them, μ i represents the i-th cluster center; n i represents the number of data points in the i-th cluster; represents the jth data point in the ith cluster.

[0170] Each time new data is input, the new data point distribution is calculated in real time to adjust the location of the current cluster center. For example, when the center value of a certain type of velocity anomaly changes from 5.2m / s to 5.8m / s, the cluster center will recalculate its location based on the new data to adapt to the change.

[0171] Dynamically adjusting the location of cluster centers can significantly improve the adaptability to the distribution of virtual and real difference data and reduce the probability of misjudgment of abnormal categories.

[0172] The random forest classification algorithm classifies data through a combination of multiple decision trees. The depth of the decision tree and the node splitting rule directly affect the accuracy and generalization ability of classification. In the present invention, the maximum depth and splitting rule of the decision tree are dynamically adjusted according to the real-time feedback data to improve the classification accuracy of the virtual-real difference data.

[0173] The node splitting rule is based on the optimal partitioning of information gain or Gini coefficient:

[0174]

[0175] Where G represents the Gini coefficient; p i represents the proportion of data points belonging to the i-th category; k represents the number of categories.

[0176] Assuming that the maximum depth of the current decision tree is 5, when the classification accuracy does not meet the set standard in real-time feedback, the model automatically increases the maximum depth to 6 to capture more fine-grained features; at the same time, by analyzing new virtual and real difference data, the node splitting rules are adjusted to reduce the Gini coefficient and make the classification results more accurate. The updated random forest algorithm can more accurately capture the complex characteristics of anomalies, improve the adaptability to new anomaly categories, and reduce the false detection rate.

[0177] The classification threshold is a key parameter for determining whether an outlier belongs to a certain category. In the present invention, the classification threshold is dynamically adjusted by real-time analysis of the distribution characteristics of the virtual-real difference data, thereby finding the best balance between sensitivity and accuracy.

[0178] The dynamic threshold adjustment formula is:

[0179] T new =T old +α·ΔD

[0180] Among them, T new represents the adjusted classification threshold; T old represents the original classification threshold; α represents the adjustment coefficient; ΔD represents the deviation value of the real-time feedback data.

[0181] Assuming that the current classification threshold of the abnormal velocity category is 5.0m / s, when new data feedback shows that the distribution range of abnormal points is concentrated at 5.5m / s, the system will automatically adjust the threshold to 5.4m / s, improving the detection sensitivity while avoiding misjudgment. The dynamic adjustment of the classification threshold significantly enhances the model's ability to respond to changes in production line status and improves the accuracy and sensitivity of anomaly detection.

[0182] The dynamic weighted feature fusion algorithm achieves the optimal fusion of multidimensional data by assigning different weights to multimodal features. In the present invention, the feature weights are dynamically adjusted according to the real-time input data and the importance coefficient of the feature to ensure that the feature fusion result can accurately reflect the real state of the production line.

[0183] The weight adjustment formula is:

[0184]

[0185] Assuming that the multimodal features include flow rate, pressure, and surface roughness, the real-time input value is F 流速 =5.8m / s, F 压力 =9.2MPa, F 粗糙度 =0.75μm, the importance coefficients are P 流速 =0.4, P 压力 =0.3, P 粗糙度 =0.3. The weight is calculated according to the formula:

[0186]

[0187] Dynamic weighted feature fusion ensures the rationality and real-time performance of multimodal feature fusion, highlights the contribution of important features to the detection results, and improves the overall reliability of anomaly detection.

[0188] The dynamic adjustment of the classification threshold and the optimization of the random forest algorithm improve the sensitivity of detection while reducing the occurrence of false detections and missed detections; the real-time adjustment of the clustering center and dynamic weights enhances the model's adaptability to dynamic changes in the production line; the dynamic weighted feature fusion algorithm improves the comprehensive analysis capability of multimodal data and provides reliable support for the real-time update of the digital twin model.

[0189] In practical applications, the real-time feedback data of a hose production line showed that the distribution range of the pressure anomaly category expanded from 8.5MPa to 9.5MPa. The system dynamically adjusted the classification threshold, raising the threshold from 8.6MPa to 9.0MPa, and accurately captured the new anomaly points; at the same time, the K-means clustering center was recalculated based on the new data, and the classification accuracy was improved by 12%. Through the dynamic weighted feature fusion algorithm, the feature weights of flow rate and pressure were adjusted to 0.4 and 0.35 respectively, effectively improving the sensitivity of anomaly detection.

[0190] like Figure 3 As shown, abnormal trend prediction data is generated based on the classified abnormal data; system correction suggestions are generated based on the abnormal trend prediction data, and the production line operation is adjusted through a closed-loop feedback mechanism, wherein the closed-loop feedback mechanism includes real-time adjustment of the temperature, pressure and flow rate parameters of the production equipment according to the prediction results, and generation of control signals to dynamically correct the equipment operation status.

[0191] Preferably, the abnormal trend prediction data is generated by the following steps:

[0192] Based on the Transformer time series model, the abnormal propagation path data in the virtual and real difference data is analyzed in time series;

[0193] Extract features of dynamic abnormal field data in virtual-real difference data, and predict the abnormal diffusion range and key abnormal points;

[0194] Output abnormal trend prediction data, including time, space and intensity distribution.

[0195] The Transformer time series model is a deep learning model based on the self-attention mechanism, which is particularly suitable for modeling long time series data. In the present invention, the time series characteristics of the virtual-real difference data (such as flow rate changes, pressure fluctuations and temperature anomalies) are analyzed by the Transformer time series model to extract the time dependency of the abnormal propagation path, thereby identifying the time trend of abnormal diffusion.

[0196] The Transformer model captures global temporal dependencies by calculating the correlation between each time step in the input data. The core of the Transformer model is the self-attention mechanism:

[0197]

[0198] Where Q, K, and V represent the query matrix, key matrix, and value matrix respectively; k Represents the dimension of the key matrix; the softmax function is used to calculate the attention weights.

[0199] The time series features in the virtual and real difference data (such as the flow rate, pressure and temperature changes at the abnormal point) are input into the Transformer time series model, and the model extracts the time features through a multi-layer encoder. For example, when a flow rate anomaly is detected from time t 1 to 5 The changing trend of is accelerated diffusion, and the model can capture this trend and output the future velocity anomaly prediction time series. Through time series analysis, the time diffusion law of the anomaly points is identified, providing high-precision basic data for dynamic prediction of anomaly propagation.

[0200] Dynamic abnormal field data is the spatial distribution characteristics of virtual and real difference data, reflecting the propagation range and diffusion intensity of abnormal points in the production line. Through feature extraction methods, the dynamic evolution law of abnormal points in multidimensional space (such as velocity field, temperature field and pressure field) is analyzed to predict the abnormal diffusion range and key abnormal points.

[0201] Feature extraction is performed by calculating the spatial gradient and propagation direction of the outlier points:

[0202]

[0203] Among them, G represents the gradient value of the outlier point; V represents the propagation vector field of the outlier point.

[0204] The dynamic anomaly field is constructed using the velocity, pressure and temperature data collected by multimodal sensors. The boundary features of the anomaly diffusion range are extracted by calculating the gradient and propagation direction of each anomaly point. For example, when the gradient value of the velocity anomaly point P exceeds the set threshold, the point is marked as a key anomaly point, and its propagation range coverage area is predicted. The feature extraction of dynamic anomaly field data can accurately locate the key anomaly points and anomaly range, and provide comprehensive information on the spatial distribution of anomalies.

[0205] The abnormal trend prediction data is generated by integrating the results of time series analysis and dynamic abnormal field feature extraction, and contains prediction information of time, space and intensity distribution. These data can not only reveal the propagation law of abnormalities, but also quantify the scope and degree of the impact of abnormalities, thus providing a scientific basis for the active adjustment of the production line.

[0206] The time, space and intensity distributions are respectively expressed as: time, the time path of the anomaly diffusion; space, the specific location of the anomaly in the hose production line; intensity, the change amplitude of the anomaly (such as flow velocity increment, temperature gradient and pressure difference).

[0207] In the implementation process, the results of time series analysis and spatial feature extraction are combined to fit and quantify the time, space and intensity distribution using interpolation algorithms. For example, when the diffusion speed of the flow anomaly point in a certain area is 2.5m / s, the predicted impact range in the next 10 seconds is [x 1 ,x 2 The output abnormal trend prediction data provides clear guidance for the early warning and adjustment strategy of the production line, significantly improving the production line's initiative in handling abnormalities.

[0208] The Transformer time series model can accurately capture the time diffusion trend of anomalies and provide high-precision predictions for anomaly propagation paths. The analysis of dynamic anomaly fields can comprehensively describe the spatial propagation range and key positions of anomalies. The intensity of anomalies can be quantified to provide a clear degree of anomaly impact and provide a decision-making basis for production line adjustments. The comprehensive analysis of time, space and intensity distribution significantly improves the comprehensiveness and accuracy of anomaly trend predictions.

[0209] In a hose production line, the virtual-real difference data collected by the sensor in real time shows that at time t 1 to 3 During this period, the propagation speed of the abnormal flow velocity point in a certain area was 2.5m / s, and the gradient value of the abnormal pressure point was 15Pa / m. Based on the Transformer time series model, the time series characteristics of the abnormal points were analyzed and predicted in the future t 4 to 6 During this period, the diffusion range of the outliers is [x 1 ,x 2 ]. At the same time, through dynamic abnormal field analysis, the intersection area of ​​abnormal flow rate and abnormal pressure is identified as the key abnormal point, and its intensity is predicted to reach a change amplitude of 20 Pa. The output abnormal trend prediction data clearly describes the time, space and intensity distribution, providing decision support for parameter adjustment of production equipment.

[0210] Preferably, the system correction suggestion includes the following adjustment parameters:

[0211] Adjust the temperature control range of production equipment to reduce the risk of local overheating;

[0212] Adjust the flow rate distribution optimization parameters of the production equipment to balance the fluid pressure in the production line;

[0213] Adjust the pressure regulation range of the production equipment to prevent the pressure from exceeding the limit during the operation of the production equipment.

[0214] During the production of hoses, excessively high temperatures may cause degradation of hose material properties or damage to equipment due to overheating. By real-time monitoring of the temperature distribution of the equipment, local overheating areas can be identified, and the temperature control range of the equipment can be dynamically adjusted to ensure that the temperature remains within a safe range. The present invention uses temperature field analysis combined with a feedback control mechanism to achieve precise temperature adjustment.

[0215] Assume that the real-time temperature distribution of a certain area of ​​the equipment is T(x,y,z). When the temperature at a certain point is T i >T threshold (Set threshold), the system will automatically reduce the heating power P in this area heat Or increase the cooling flow Q cool .

[0216] Use multi-point temperature sensors to collect temperature data on the surface of the equipment, build a real-time temperature field, and control the temperature by adjusting the heating power or coolant flow. For example, when the temperature of a local area of ​​the equipment rises from 80°C to 95°C, the system will increase the coolant flow ΔQ cool = 10% to reduce the temperature to 85°C. The dynamic adjustment of the temperature control range effectively prevents the risk of local overheating and improves the stability of equipment operation and the reliability of hose quality.

[0217] In a production line, unbalanced fluid pressure may lead to unbalanced load of production equipment or abnormal local flow rate in the pipeline, thus affecting production stability. The present invention optimizes and adjusts the flow rate distribution of the fluid in the equipment, balances the pressure distribution, and reduces the occurrence of abnormal pressure points.

[0218] The adjustment of flow rate optimization parameters is based on the relationship between the pressure field and the flow rate field:

[0219]

[0220] in, represents the pressure gradient; μ represents the fluid viscosity coefficient; The Laplace operator represents the flow rate. The system adjusts the flow rate Q at the fluid inlet of the device in and the outlet flow rate Q out To achieve a balanced flow velocity distribution.

[0221] By real-time monitoring of the flow rate and pressure distribution in the production line, the abnormal flow rate area is identified and the fluid inlet and outlet flow rates are adjusted. For example, when the flow rate v(x,y,z)=3.5m / s in a certain section of the pipeline is significantly higher than the neighborhood average of 2.8m / s, the system will reduce the inlet flow rate Q in Or increase the outlet flow Q out , to balance the flow velocity distribution. The adjustment of the flow velocity distribution optimization parameters significantly reduced the amplitude of pressure fluctuations, improved the fluid balance of the production line, and avoided abnormal equipment load caused by uneven pressure.

[0222] During the operation of production equipment, overpressure may cause equipment failure or pipeline rupture. The present invention controls the operating pressure of the production equipment within a safe range by dynamically adjusting the pressure regulation range, thereby preventing overpressure or excessive pressure fluctuations.

[0223] The setting of the pressure adjustment range is based on the real-time pressure field distribution:

[0224] P safe =P mean ±k·σ

[0225] Among them, P safe Indicates safe pressure range; Pmean represents the average pressure; σ represents the standard deviation of pressure; and k represents the safety factor. The system dynamically adjusts the pressure range by controlling the pressure relief valve or flow control valve of the equipment.

[0226] The pressure sensor monitors the pressure value inside the device in real time. When the pressure P at a certain point is detected, i >P safe , the system will automatically start the pressure relief valve or adjust the flow control valve. For example, when the pressure in a certain area reaches 10.5MPa, exceeding the safety upper limit of 10.0MPa, the system releases the excessive pressure and restores the pressure to 9.8MPa. The dynamic adjustment of the pressure regulation range prevents the occurrence of pressure over-limit problems and ensures the safe operation of equipment and pipelines.

[0227] Dynamically adjust key parameters of temperature, flow rate and pressure to solve problems such as local overheating, fluid imbalance and pressure overrun; parameter adjustment effectively reduces fluctuations in equipment operation and improves production line operation efficiency and product quality; parameter adjustment is based on real-time monitoring data and dynamic feedback mechanism to ensure the stability and safety of the production process.

[0228] Example: In a hose production line, the temperature in a certain area reached 95°C when the equipment was running, the flow rate fluctuated to 4.0m / s near the outlet, and the pressure rose to 10.5MPa. The system was corrected by the following measures:

[0229] Increase the coolant flow rate by 15% and control the temperature within 85°C; reduce the inlet flow rate by 10% and adjust the outlet flow rate by 8% to restore the flow rate to the equilibrium value of 3.2m / s; start the pressure relief valve and reduce the pressure to 9.8MPa. Through the above adjustments, the equipment resumed stable operation and avoided potential production interruptions and equipment damage.

[0230] Preferably, the closed-loop feedback mechanism comprises the following steps:

[0231] Collect the status data of the production line after adjustment in real time and verify the system correction suggestions;

[0232] Dynamically update the operating parameters of production equipment to optimize the operating status of production equipment;

[0233] New operating rules are formed based on the verification results to complete the self-learning optimization cycle.

[0234] After the production line implements the system correction suggestion, the sensor collects the adjusted operating status data in real time to verify whether the correction suggestion has achieved the expected effect. This step evaluates the effectiveness of the adjustment measures by comparing the changes in key parameters before and after the adjustment. For example, real-time data of temperature, flow rate and pressure are collected and compared with the target value.

[0235] The validation metric may include the error calculation formula:

[0236]

[0237] Where E represents the average error; M i Indicates the actual parameters collected in real time; T i represents the target parameter value; n represents the number of sampling points.

[0238] Assume that the adjusted target temperature is 85°C, the flow rate is 3.0m / s, and the pressure is 10.0MPa. The real-time data collected by the sensor is 87°C, 3.1m / s, and 10.2MPa. The system calculates the error of each parameter. If the error is less than the preset tolerance range, the correction suggestion is considered effective; otherwise, further adjustment is required. Real-time collection and verification ensure that the effect of the adjustment measures can be quickly fed back, which helps to detect and correct deviations in a timely manner.

[0239] Based on the verification results, the system dynamically adjusts the operating parameters of the production equipment to gradually approach the optimal state. This adjustment relies on feedback control theory and iteratively optimizes the parameters through closed-loop control. The core of dynamic update lies in the sensitive adjustment of control parameters to avoid excessive adjustment that leads to system instability.

[0240] The dynamic update formula of operating parameters can be described as:

[0241] P new =P old +k·ΔE

[0242] Among them, P new Indicates the updated operating parameters; P old represents the current operating parameters; k represents the adjustment coefficient; ΔE represents the change in parameter error.

[0243] Assume that the current operating parameters are temperature 85°C, flow rate 3.0m / s, and pressure 10.0MPa. If the verification result shows that the temperature deviation is +2°C, the temperature is adjusted by reducing the heating power, and the corrected temperature parameter is 84°C. Similarly, adjust the flow valve and pressure control valve to correct the flow rate and pressure parameters. Dynamically updating the operating parameters ensures that the production equipment always operates in the best state, improves production efficiency and stability, and avoids the risk of equipment failure caused by parameter mismatch.

[0244] The self-learning optimization cycle is the core of the closed-loop feedback mechanism. By using the verification results and optimized operating parameters as new learning samples, the system's operating rules are dynamically updated, so that production equipment can gradually adapt to different working conditions and form the optimal operating mode. This optimization is based on the principle of reinforcement learning and can improve system performance through continuous adjustments.

[0245] The goal in reinforcement learning is to maximize the cumulative reward function:

[0246]

[0247] Among them, R represents the total reward value; r t represents the immediate reward at time t; T represents the total time steps of the adjustment process.

[0248] The system uses the optimized parameters (such as temperature, flow rate and pressure) as input and updates the operating rules through the reinforcement learning algorithm. For example, when the temperature control error is reduced to less than 1°C for three consecutive times, the system will automatically write the current control strategy (such as heating power and cooling flow ratio) as a new rule into the parameter library for use in similar scenarios in the future. The self-learning optimization loop makes the system dynamically adaptable and can continuously optimize the operating rules according to real-time changes, realizing the transition from passive adjustment to active optimization.

[0249] Real-time collection and verification of adjusted status data ensures the timeliness and accuracy of system correction suggestions; dynamic update of operating parameters of production equipment enables the system to maintain optimal operating status under different working conditions; new operating rules are formed based on optimization cycles to enhance the intelligence level and adaptability of the system; fully automated monitoring and adjustment are achieved through a closed-loop feedback mechanism, reducing the uncertainty of manual operation and improving the reliability and efficiency of the production line.

[0250] Example: In a hose production line, the system correction suggested adjustments to the temperature, flow rate and pressure. The corrected real-time collected data was 87°C, 3.1m / s, and 10.2MPa, which deviated from the target values. Through the closed-loop feedback mechanism, the system dynamically adjusted the temperature heating power, inlet flow rate, and pressure relief valve parameters in turn, and finally adjusted the operating parameters to 85°C, 3.0m / s, and 10.0MPa. At the same time, the system writes these adjustment strategies into the rule base to guide the handling of similar anomalies in the future.

[0251] like Figure 4 As shown, a system for implementing the hose production line monitoring method comprises:

[0252] The multimodal data acquisition module is configured to collect visual data, temperature data, pressure data, vibration data and environmental data of the hose production line, and perform time synchronization, normalization and denoising on the collected multimodal production data to generate fused multimodal production data; through the comprehensive collection of visual, temperature, pressure, vibration and environmental data, combined with time synchronization, normalization and denoising, fused multimodal production data is generated to provide accurate input for subsequent modeling and analysis. These data can fully reflect the real-time status of the production line.

[0253] The digital twin modeling module is configured to receive the fused multimodal production data, generate a dynamic twin model, compare the production data of the actual production line with the dynamic twin model, and generate virtual-real difference data; construct a dynamic twin model, compare the actual operating status of the production line with the virtual model in real time, and generate virtual-real difference data. This virtual-real comparison method can accurately capture potential anomalies and performance deviations in the production line.

[0254] An anomaly detection module is configured to perform anomaly detection based on the virtual-real difference data, and the anomaly detection module includes:

[0255] A classification module based on classification algorithms is used to analyze virtual and real difference data and classify outliers;

[0256] A rule-matching-based detection module is used to mark abnormal points according to preset rules;

[0257] The anomaly detection model that combines classification algorithms and rule matching is used to mark and classify anomalies in virtual and real difference data. The detection model is made more sensitive and robust by optimizing parameters through deep learning, and the optimized model parameters are fed back to the digital twin modeling module to dynamically update the simulation rules.

[0258] An optimization module is used to optimize the parameters of the anomaly detection model through model weight adjustment based on deep learning and dynamic correction of thresholds based on real-time data feedback, and feed the optimized parameters back to the digital twin modeling module to update its simulation rules;

[0259] An abnormal trend prediction module is configured to generate abnormal trend prediction data based on the classified abnormal data, and the abnormal trend prediction module predicts the abnormal diffusion range, key abnormal points and time, space and intensity distribution through a time series analysis model;

[0260] Based on the classified abnormal data, the time series analysis model is used to predict the spread of the abnormality, key abnormal points, and the time, space and intensity distribution, helping the production line to identify and prevent potential problems in advance.

[0261] A closed-loop feedback control module is configured to generate system correction suggestions based on abnormal trend prediction data and adjust the production line operation through a closed-loop feedback mechanism, wherein the closed-loop feedback mechanism includes:

[0262] Adjust the temperature parameters, pressure parameters and flow rate parameters of production equipment in real time;

[0263] Generate control signals for dynamically correcting the operating status of production equipment;

[0264] Optimize system correction suggestions through real-time data verification and dynamic updates.

[0265] Based on the abnormal trend prediction data, system correction suggestions are proposed, and dynamic control signals are generated by real-time adjustment of the temperature, pressure and flow rate parameters of the production equipment to achieve closed-loop feedback optimization. By verifying the effect of the correction suggestions, the operating rules are dynamically updated to build a self-learning optimization cycle.

[0266] The above are only embodiments of the present application and are not intended to limit the present application. For those skilled in the art, the present application may have various changes and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application should be included in the scope of the claims of the present application.

Claims

1. A hose production line monitoring method, characterized in that: The following steps are involved: Collect multimodal production data of the hose production line, pre-process the collected data, and generate fused multimodal production data; Input the fused multimodal production data into the digital twin modeling module to generate a dynamic twin model; Compare the actual production line data with the dynamic twin model to generate virtual and real difference data; Anomaly detection is performed based on virtual-real difference data. The virtual-real difference data is analyzed using an anomaly detection model. The anomaly detection model includes a structure based on a classification algorithm and rule matching, which is used to mark and classify abnormal points and generate classified abnormal data; Optimize the parameters of the anomaly detection model, where the optimization method includes model weight adjustment based on deep learning and threshold dynamic correction based on real-time data feedback, and feed the optimized parameters back to the digital twin model to update its simulation rules; Generate abnormal trend prediction data based on classified abnormal data; generate system correction suggestions based on abnormal trend prediction data, and adjust the production line operation through a closed-loop feedback mechanism, where the closed-loop feedback mechanism includes real-time adjustment of the temperature, pressure and flow rate parameters of the production equipment according to the prediction results, and generation of control signals to dynamically correct the equipment operation status.

2. The hose production line monitoring method according to claim 1, characterized in that: The preprocessing of the multimodal production data comprises the following steps: Perform edge detection on the visual data of the hose production line and extract the surface texture features of the visual data; Normalize the temperature data and pressure data of the hose production line respectively; The vibration data of the hose production line is processed by frequency domain filtering.

3. The hose production line monitoring method according to claim 1, characterized in that: The dynamic twin model includes: Flow field model, which generates flow field characteristic data through three-dimensional flow field simulation based on computational fluid dynamics algorithm; The geometric model is generated by a point cloud-based reconstruction algorithm to describe the state of the hose surface.

4. The hose production line monitoring method according to claim 3, characterized in that: The virtual-real difference data includes the following contents: Flow field abnormal data, including velocity gradient mutation data and pressure abnormal point data; Geometric deviation data, including surface roughness deviation data and crack extension area data; Temperature and pressure abnormal data, including abnormal range data and change trend data.

5. The hose production line monitoring method according to claim 1, characterized in that: The classification algorithm of the anomaly detection model includes the following: Use K-means clustering algorithm to cluster the abnormal points in the virtual-real difference data and generate abnormal category data; Use the random forest classification algorithm to predict the category of abnormal points in the virtual-real difference data and generate classified abnormal data; The rule matching structure corrects the classified abnormal data through dynamic detection rules based on preset thresholds.

6. The hose production line monitoring method according to claim 5, characterized in that: The method for optimizing the anomaly detection model parameters comprises: Dynamically adjust the location of the K-means clustering center based on real-time feedback data; Update the depth and node splitting rules of decision trees in the random forest classification algorithm; Adjust classification thresholds to improve the sensitivity and accuracy of anomaly detection; The weight adjustment formula of the dynamic weighted feature fusion algorithm is as follows: Among them, W i represents the weight of the i-th feature; F i represents the real-time input value of the i-th feature; P i represents the importance coefficient of the i-th feature; n represents the total number of features.

7. The hose production line monitoring method according to claim 1, characterized in that: The abnormal trend prediction data is generated by the following steps: Based on the Transformer time series model, the abnormal propagation path data in the virtual and real difference data is analyzed in time series; Extract features of dynamic abnormal field data in virtual-real difference data, and predict the abnormal diffusion range and key abnormal points; Output abnormal trend prediction data, including time, space and intensity distribution.

8. The hose production line monitoring method according to claim 1, characterized in that: The system correction suggestions include the following adjustment parameters: Adjust the temperature control range of production equipment to reduce the risk of local overheating; Adjust the flow rate distribution optimization parameters of the production equipment to balance the fluid pressure in the production line; Adjust the pressure regulation range of the production equipment to prevent the pressure from exceeding the limit during the operation of the production equipment.

9. The hose production line monitoring method according to claim 1, characterized in that: The closed-loop feedback mechanism comprises the following steps: Collect the status data of the production line after adjustment in real time and verify the system correction suggestions; Dynamically update the operating parameters of production equipment to optimize the operating status of production equipment; New operating rules are formed based on the verification results to complete the self-learning optimization cycle.

10. A system for implementing the hose production line monitoring method according to any one of claims 1 to 9, characterized in that: include: A multimodal data acquisition module is configured to collect visual data, temperature data, pressure data, vibration data and environmental data of the hose production line, and perform time synchronization, normalization and denoising on the collected multimodal production data to generate fused multimodal production data; A digital twin modeling module is configured to receive the fused multimodal production data, generate a dynamic twin model, and compare the production data of the actual production line with the dynamic twin model to generate virtual-real difference data; An anomaly detection module is configured to perform anomaly detection based on the virtual-real difference data, and the anomaly detection module includes: A classification module based on classification algorithms is used to analyze virtual and real difference data and classify outliers; A rule-matching-based detection module is used to mark abnormal points according to preset rules; An optimization module is used to optimize the parameters of the anomaly detection model through model weight adjustment based on deep learning and dynamic correction of thresholds based on real-time data feedback, and feed the optimized parameters back to the digital twin modeling module to update its simulation rules; An abnormal trend prediction module is configured to generate abnormal trend prediction data based on the classified abnormal data, and the abnormal trend prediction module predicts the abnormal diffusion range, key abnormal points and time, space and intensity distribution through a time series analysis model; A closed-loop feedback control module is configured to generate system correction suggestions based on abnormal trend prediction data and adjust the production line operation through a closed-loop feedback mechanism, wherein the closed-loop feedback mechanism includes: Adjust the temperature parameters, pressure parameters and flow rate parameters of production equipment in real time; Generate control signals for dynamically correcting the operating status of production equipment; Optimize system correction suggestions through real-time data verification and dynamic updates.

Images