Real-time data acquisition-based weftless tape machine intelligent monitoring system and method

Abstract

The invention relates to the technical field of industrial automatic monitoring, in particular to a weftless tape machine intelligent monitoring system and method based on real-time data acquisition. The system comprises a data acquisition module used for acquiring real-time data of a multi-heterogeneous sensor; the data processing module is used for preprocessing the real-time data; the intelligent analysis module covers a feature extraction unit, an anomaly detection unit, a fault diagnosis unit, a service life prediction unit and a data fusion unit and can extract data features, detect anomaly, diagnose faults, predict the service life of a part and fuse multi-unit output to make a decision; and the early warning response module triggers corresponding levels of alarms according to the analysis result and generates a diagnosis report. According to the invention, comprehensive real-time monitoring, early fault early warning, accurate fault diagnosis, residual life prediction, multi-source information fusion and the like of the operation state of the weftless tape machine are realized, the production efficiency and the product quality are effectively improved, and intelligent management of production of the weftless tape machine is assisted.

Description

Technical Field

[0001] The present invention relates to the technical field of industrial automation monitoring, and particularly to an intelligent monitoring system and method for a non-woven belt machine based on real-time data acquisition. Background Art

[0002] The non-woven belt machine is a key device for producing glass fiber non-woven binding belts, and is widely used in the insulation binding of electrical equipment such as motors and transformers. The traditional non-woven belt production process mainly relies on manual experience for monitoring, and there are the following problems: Lag in fault detection: Equipment faults are often discovered only after they have caused product quality problems or shutdowns, resulting in low production efficiency and increased costs.

[0003] Irrational maintenance strategy: Adopting the methods of regular maintenance or after-fact maintenance, it is impossible to carry out precise maintenance according to the actual state of the equipment, and it is easy to cause over-maintenance or under-maintenance.

[0004] Difficulty in quality control: The production quality of non-woven belts is affected by various factors, such as yarn tension, temperature, speed, etc. It is difficult for manual monitoring to achieve real-time collaborative control of multiple parameters.

[0005] Lack of data-driven decision-making: A large amount of data in the production process is not effectively utilized, and it is impossible to provide a basis for equipment optimization and process improvement.

[0006] In recent years, with the development of sensor technology, the Internet of Things and artificial intelligence, some data-driven equipment monitoring systems have emerged. However, the existing monitoring systems still have deficiencies when dealing with complex industrial equipment such as non-woven belt machines. For example, they have limited ability to fuse multi-source heterogeneous data, low accuracy in fault diagnosis, and lack of accurate prediction of the remaining service life of the equipment. Summary of the Invention

[0007] In view of the above-mentioned disadvantages and deficiencies, the present invention provides an intelligent monitoring system and method for a non-woven belt machine based on real-time data acquisition. Through multi-sensor data fusion and deep learning algorithms, it realizes real-time perception of equipment status, early diagnosis of faults and accurate prediction of remaining life, and improves equipment reliability and production efficiency.

[0008] An intelligent monitoring system for a non-woven belt machine based on real-time data acquisition includes: Data acquisition module: Obtain real-time data from multiple heterogeneous sensors through an industrial communication network. The sensors include vibration sensors deployed at the main shaft bearing part, yarn tension sensors installed at the yarn guiding roller position, optical sensors arranged in the winding area, and at least one selected from temperature sensors, yarn break sensors, position sensors, speed sensors, sound sensors, and power consumption sensors; Data processing module: It includes distributed edge computing nodes for performing data cleaning, filtering, normalization, segmentation, and synchronization operations; Intelligent analysis module: It includes a feature extraction unit, an anomaly detection unit, a fault diagnosis unit, a remaining useful life prediction unit, and a data fusion unit. The feature extraction unit is used to convert the preprocessed raw data into a multi-dimensional feature vector and output it to the subsequent units. The anomaly detection unit identifies outliers in the data based on the feature vector and outputs an anomaly score to the data fusion unit. The fault diagnosis unit classifies the feature vector to determine the specific fault type and confidence level and outputs it to the data fusion unit. The remaining useful life prediction unit predicts the remaining useful life based on the historical feature sequence and outputs probability distribution parameters to the data fusion unit. The data fusion unit integrates the outputs of the anomaly detection unit, the fault diagnosis unit, and the remaining useful life prediction unit to generate a final decision and trigger an early warning response; Early warning response module: It triggers multi-level alarms according to the anomaly level, generates a diagnostic report containing the fault probability and disposal suggestions, and executes the equipment protection strategy.

[0009] Preferably, the data cleaning, filtering, normalization, segmentation, and synchronization operations include: Identifying outliers using the Z-score method and filling in missing values through cubic spline interpolation; Applying a low-pass filter for data filtering; Scaling the data to the [0, 1] interval using min-max normalization; Segmenting the data using a sliding time window with 1024 sampling points; Achieving multi-sensor data timestamp alignment based on the NTP (Network Time Protocol) protocol.

[0010] Preferably, the anomaly detection unit identifies outliers in the data based on the feature vector and outputs an anomaly score to the data fusion unit, specifically including: Feature standardization: Performing Z-score standardization on the raw feature vector output by the feature extraction unit; Isolation forest modeling: Constructing an isolation forest model containing 100 - 200 decision trees, with the maximum depth of each tree being 8 - 12 layers; randomly partitioning the standardized feature vector until each sample is isolated or the maximum tree depth is reached; Anomaly score calculation: Calculating the average path length of each sample in the isolation forest , and then converting the average path length into an anomaly score : ; where, is the normalization factor of the average path length of the tree, is the number of training samples; Dynamic threshold setting: Set the dynamic threshold based on the abnormal score distribution of historical normal data : ; where and are the mean and standard deviation of the abnormal scores of normal samples respectively, is a dynamically adjustable coefficient; Time series smoothing: Apply the exponentially weighted moving average (EWMA) filter to the abnormal scores within consecutive time windows to obtain smoothed abnormal scores; Result output: Output the smoothed abnormal scores to the data fusion unit.

[0011] Preferably, the fault diagnosis unit classifies the feature vectors, determines the specific fault type and confidence level, and outputs them to the data fusion unit, specifically including: Feature preprocessing: Perform L2 normalization on the original feature vectors output by the feature extraction unit; Apply principal component analysis for feature dimensionality reduction, and retain the principal components with a cumulative variance contribution rate ≥ 95%; Multi-scale feature extraction: Construct a multi-branch convolutional neural network, and each branch uses different-sized convolutional kernels to extract features of different scales; Perform global average pooling on the output of each branch to obtain multi-scale feature representations; Attention mechanism weighting: Calculate the importance weights of features in each dimension through the attention mechanism; Multiply the original feature vectors element-wise by the attention weights to obtain weighted feature vectors; Fault classification and confidence calculation: Input the weighted feature vectors into the fully connected layer for classification, and use the Softmax activation function to output the probability distribution of various faults; Take the maximum probability value as the confidence score of the classification result; Uncertainty quantification: Use the Monte Carlo Dropout method to quantify the classification uncertainty, specifically: Keep the Dropout activation state during the test phase and perform T forward propagations, where T ≥ 20; Calculate the variance of the probability distributions of T predictions; Take the variance as the uncertainty measure of the classification result; Result Output and Verification: Output the classification result, confidence score, and uncertainty measure to the data fusion unit; When the confidence is less than the confidence threshold or the uncertainty is greater than the uncertainty threshold, trigger the secondary verification process.

[0012] Preferably, the remaining useful life is predicted based on the historical feature sequence by the remaining useful life prediction unit, and the probability distribution parameters are output to the data fusion unit, specifically including: Temporal Feature Construction: Arrange the multi-dimensional feature vectors output by the feature extraction unit in chronological order to construct a sliding time window of length L , where is the feature vector at time; Perform differential processing on the feature sequence within each time window to calculate the change rate between adjacent time steps ; Combine the original features and the change rate features to form an enhanced feature sequence ; Sequence-to-Sequence Modeling: Construct a gated recurrent unit (GRU) model with at least 2 hidden layers, each layer having 64 - 128 neurons; Use the attention mechanism to weight the input sequence: ; where is the hidden state at the th time step, , are trainable weight matrices, is the bias term, is the activation function; Input the weighted hidden state sequence into the fully connected layer to output the predicted remaining useful life; Probability Distribution Fitting: Convert the point prediction result of the GRU model into Weibull distribution parameters: Collect historical prediction errors and calculate the shape parameter : ; where is the ratio of the sample standard deviation to the mean, is the gamma function; Calculate the scale parameter : ; where is the remaining useful life predicted by the GRU model; Uncertainty quantification: Estimate prediction uncertainty using the Monte Carlo Dropout method: Keep Dropout activated during the test phase and perform M forward propagations; Fit the Weibull distribution to each prediction result separately to obtain M sets of parameters ; Calculate the statistical distribution of the parameters and output the 95% confidence interval; Dynamically update the prediction result: When new sensor data arrives, update the prediction result using exponential weighting: ; where, is the forgetting factor, is the current predicted value, is the previous predicted value; Result output: Output the Weibull distribution parameters and their 95% confidence interval as the probability density function to the data fusion unit.

[0013] Preferably, the data fusion unit integrates the outputs of the anomaly detection unit, the fault diagnosis unit, and the remaining useful life prediction unit to generate a final decision and trigger an early warning response, specifically including: Data preprocessing and standardization: Map the anomaly scores output by the anomaly detection unit to the interval [0, 1] and convert them into anomaly probabilities; Normalize the confidence scores of various faults output by the fault diagnosis unit; Convert the Weibull distribution parameters output by the remaining useful life prediction unit into the probability density function of the remaining useful life; Temporal and spatial consistency verification: Check the consistency of the outputs of the three analysis units in terms of timestamps, and perform interpolation alignment on the data with a time deviation exceeding 50 ms; Verify spatial consistency: Ensure that the physical location located by the anomaly detection matches the fault diagnosis result; Evidence weight assignment: Dynamically adjust the weights of each analysis unit according to the equipment operating state: Startup phase: Increase the weight of the anomaly detection unit by 20%; Steady-state operation: Increase the weight of the fault diagnosis unit by 15%; Approaching the maintenance cycle: Increase the weight of the remaining useful life prediction unit by 25%; Calculate the credibility coefficient of each analysis unit based on historical accuracy; Decision fusion: Fusing the anomaly probability and the confidence level of the fault type using the Dempster-Shafer evidence theory: Convert the anomaly probability into a basic probability assignment (BPA); Convert the confidence level of the fault type into the corresponding BPA; Calculate the fused BPA through the evidence combination rule; Combine the results of life prediction to determine the comprehensive risk index ; Decision threshold setting: Set three-level risk thresholds according to the historical data of the equipment and expert knowledge: High risk: , immediately stop the machine for inspection and generate an emergency repair work order; Medium risk: , reduce the operating load and arrange preventive maintenance; Low risk: , operate normally and record the monitoring status; Early warning response trigger: Trigger the corresponding level of early warning response according to the comprehensive risk index; Generate a diagnostic report including the anomaly location, fault type, risk level, and disposal suggestions; Feed back the decision result to the model training unit for continuous optimization of the model.

[0014] Preferably, the intelligent analysis module further includes: Fuse multi-sensor features based on the attention mechanism and automatically adjust the weight coefficients of the features of each sensor according to different working conditions; Adopt the transfer learning method to transfer the model parameters trained on similar equipment to the target non-woven tape machine model, reducing the training sample requirements of the target model.

[0015] Preferably, the distributed edge computing node is configured to: Perform real-time data preprocessing at a position close to the sensor; Through the edge-cloud collaborative architecture, only transmit the feature data and anomaly events to the cloud server; Realize the edge incremental learning of the model based on the locally stored historical data.

[0016] An intelligent monitoring method for a non-woven tape machine based on real-time data acquisition, including: Obtain data in real time from multiple heterogeneous sensors; Perform data preprocessing operations on the obtained real-time data; Use a machine learning model to analyze the preprocessed data to detect running anomalies or predict impending faults; Trigger an early warning response based on the detection or prediction results.

[0017] Compared with the prior art, the advantages of the present invention are as follows: Real-time status monitoring: Through multi-sensor data acquisition and processing, comprehensive and real-time monitoring of the running status of the weftless tape machine is achieved.

[0018] Early fault warning: Using machine learning algorithms to detect and warn of equipment anomalies early, discover potential faults in advance, and reduce downtime and maintenance costs.

[0019] Precise fault diagnosis: Through multi-dimensional feature extraction and deep learning models, precise diagnosis of fault types is achieved, improving the accuracy and efficiency of fault location.

[0020] Remaining life prediction: Based on time series data analysis and probability models, prediction of the remaining service life of key components of the equipment is realized, supporting the formulation of preventive maintenance strategies.

[0021] Multi-source information fusion: Adopting advanced data fusion technologies, integrating information from multiple aspects such as anomaly detection, fault diagnosis, and life prediction, generating comprehensive decisions, and improving the reliability and robustness of the monitoring system. Detailed implementation manners

[0022] To better explain the present invention for easy understanding, the present invention will be described in detail below through specific implementation manners.

[0023] This embodiment provides a weftless tape machine intelligent monitoring system based on real-time data acquisition, including a data acquisition module, a data processing module, an intelligent analysis module, and an early warning response module; The data acquisition module obtains real-time data from multiple heterogeneous sensors through an industrial communication network. The sensors include vibration sensors deployed at the main shaft bearing part, yarn tension sensors installed at the yarn guiding roller position, optical sensors arranged in the winding area, and at least one selected from temperature sensors, yarn break sensors, position sensors, speed sensors, sound sensors, and power consumption sensors.

[0024] The data processing module includes distributed edge computing nodes for performing data cleaning, filtering, normalization, segmentation, and synchronization operations.

[0025] Specifically, the data cleaning, filtering, normalization, segmentation, and synchronization operations include: Using the Z-score method to identify outliers and filling missing values through cubic spline interpolation; Applying a low-pass filter for data filtering; Using min-max normalization to scale the data to the [0, 1] interval; Segment the data using a sliding time window of 1024 sampling points; Implement multi-sensor data timestamp alignment based on the NTP protocol.

[0026] The intelligent analysis module includes a feature extraction unit, an anomaly detection unit, a fault diagnosis unit, a remaining useful life prediction unit, and a data fusion unit. The feature extraction unit is used to convert the preprocessed raw data into a multi-dimensional feature vector and output it to the subsequent units. The anomaly detection unit identifies outliers in the data based on the feature vector and outputs the anomaly score to the data fusion unit. The fault diagnosis unit classifies the feature vector to determine the specific fault type and confidence level and outputs it to the data fusion unit. The remaining useful life prediction unit predicts the remaining useful life based on the historical feature sequence and outputs the probability distribution parameters to the data fusion unit. The data fusion unit integrates the outputs of the anomaly detection unit, the fault diagnosis unit, and the remaining useful life prediction unit, generates a final decision, and triggers an early warning response.

[0027] Specifically, the anomaly detection unit identifying outliers in the data based on the feature vector and outputting the anomaly score to the data fusion unit includes: Feature standardization: Perform Z-score standardization on the raw feature vector output by the feature extraction unit; Isolation forest modeling: Construct an isolation forest model containing 100 - 200 decision trees, with the maximum depth of each tree being 8 - 12 layers; randomly partition the standardized feature vector until each sample is isolated or the maximum tree depth is reached; Anomaly score calculation: Calculate the average path length of each sample in the isolation forest , and then convert the average path length into an anomaly score : ; where is the normalization factor of the average path length of the tree, is the number of training samples; Dynamic threshold setting: Set the dynamic threshold based on the anomaly score distribution of historical normal data : ; where and are the mean and standard deviation of the anomaly scores of normal samples respectively, is the dynamically adjustable coefficient; Time series smoothing: Apply the exponentially weighted moving average (EWMA) filter to the anomaly scores within consecutive time windows to obtain the smoothed anomaly scores; Result output: Output the smoothed anomaly scores to the data fusion unit.

[0028] Specifically, the fault diagnosis unit classifies the feature vectors, determines the specific fault types and confidence levels, and outputs them to the data fusion unit, including: Feature preprocessing: Perform L2 normalization on the original feature vectors output by the feature extraction unit; Apply principal component analysis for feature dimensionality reduction, and retain the principal components with a cumulative variance contribution rate ≥ 95%; Multi-scale feature extraction: Construct a multi-branch convolutional neural network, and each branch uses convolutional kernels of different sizes to extract features of different scales; Perform global average pooling on the output of each branch to obtain multi-scale feature representations; Attention mechanism weighting: Calculate the importance weights of the features in each dimension through the attention mechanism; Multiply the original feature vectors element-wise with the attention weights to obtain weighted feature vectors; Fault classification and confidence calculation: Input the weighted feature vectors into the fully connected layer for classification, and use the Softmax activation function to output the probability distribution of various faults; Take the maximum probability value as the confidence score of the classification result; Uncertainty quantification: Adopt the Monte Carlo Dropout method to quantify the classification uncertainty, specifically: Keep the Dropout activation state during the test phase, and perform T forward propagations, where T ≥ 20; Calculate the variance of the probability distributions of the T predictions; Take the variance as the uncertainty measure of the classification result; Result output and verification: Output the classification result, confidence score, and uncertainty measure to the data fusion unit; When the confidence is less than the confidence threshold or the uncertainty is greater than the uncertainty threshold, trigger the secondary verification process.

[0029] Specifically, the remaining useful life prediction unit predicts the remaining useful life based on the historical feature sequences, and outputs the probability distribution parameters to the data fusion unit, including: Temporal feature construction: Arrange the multi-dimensional feature vectors output by the feature extraction unit in chronological order to construct a sliding time window of length L , where, is the feature vector at time Differentiate the feature sequence within each time window and calculate the change rate between adjacent time steps. ; Combine the original features and the change rate features to form an enhanced feature sequence. ; Sequence-to-sequence modeling: Construct a gated recurrent unit (GRU) model with at least 2 hidden layers, each layer having 64 - 128 neurons. Adopt an attention mechanism to weight the input sequence: ; where is the hidden state at the -th time step, , are trainable weight matrices, is the bias term, is the activation function; Input the weighted hidden state sequence into a fully connected layer and output the predicted remaining useful life. Probability distribution fitting: Convert the point prediction result of the GRU model into Weibull distribution parameters: Collect historical prediction errors and calculate the shape parameter : ; where is the ratio of the sample standard deviation to the mean, is the gamma function; Calculate the scale parameter : ; where is the remaining useful life predicted by the GRU model; Uncertainty quantification: Adopt the Monte Carlo Dropout method to estimate the prediction uncertainty: Keep the Dropout activation state during the test phase and perform M forward propagations. Fit the Weibull distribution to each prediction result respectively to obtain M sets of parameters ; Calculate the statistical distribution of the parameters and output the 95% confidence interval. Dynamic update of the prediction result: When new sensor data arrives, update the prediction result using exponential weighting: ; where is the forgetting factor. is the current predicted value, is the previous predicted value; Result output: Output the Weibull distribution parameters and their 95% confidence intervals to the data fusion unit as the probability density function.

[0030] Specifically, the data fusion unit integrates the outputs of the anomaly detection unit, the fault diagnosis unit, and the life prediction unit to generate a final decision and trigger an early warning response, including: Data preprocessing and standardization: Map the anomaly scores output by the anomaly detection unit to the interval [0, 1] and convert them into anomaly probabilities; Normalize the confidence scores of various faults output by the fault diagnosis unit; Convert the Weibull distribution parameters output by the life prediction unit into the probability density function of the remaining useful life; Temporal and spatial consistency verification: Check the consistency of the outputs of the three analysis units in terms of timestamps, and perform interpolation alignment on the data with a time deviation exceeding 50 ms; Verify spatial consistency: Ensure that the physical location located by the anomaly detection matches the fault diagnosis result; Evidence weight assignment: Dynamically adjust the weights of each analysis unit according to the device operating state: Startup phase: Increase the weight of the anomaly detection unit by 20%; Steady-state operation: Increase the weight of the fault diagnosis unit by 15%; Approaching the maintenance cycle: Increase the weight of the life prediction unit by 25%; Calculate the credibility coefficient of each analysis unit based on historical accuracy; Decision fusion: Adopt Dempster-Shafer evidence theory to fuse the anomaly probability and the fault type confidence: Convert the anomaly probability into the basic probability assignment BPA; Convert the fault type confidence into the corresponding BPA; Calculate the fused BPA through the evidence synthesis rule; Combine the life prediction results to determine the comprehensive risk index ; Decision threshold setting: Set three-level risk thresholds according to the device historical data and expert knowledge: High risk: , immediately stop the machine for inspection and generate an emergency repair work order; Medium risk: , reduce the operating load and arrange preventive maintenance; Low risk: , normal operation, record the monitoring status; Early warning response trigger: Trigger the corresponding level of early warning response according to the comprehensive risk index; Generate a diagnostic report including the abnormal location, fault type, risk level and disposal suggestions; Feed back the decision result to the model training unit for continuous optimization of the model.

[0031] In a preferred embodiment of this embodiment, the intelligent analysis module further includes: Fuse multi-sensor features based on the attention mechanism and automatically adjust the weight coefficients of each sensor feature according to different working conditions; Adopt the transfer learning method to transfer the model parameters trained on similar devices to the target weftless tape machine model, reducing the training sample requirements of the target model.

[0032] The distributed edge computing node is configured to: Perform real-time data preprocessing at a position close to the sensor; Through the edge-cloud collaborative architecture, only transmit the feature data and abnormal events to the cloud server; Realize the edge incremental learning of the model based on the historical data stored locally.

[0033] In another embodiment, there is also a weftless tape machine intelligent monitoring method based on real-time data acquisition, including the following steps: Obtain data in real time from multiple heterogeneous sensors; Perform data preprocessing operations on the obtained real-time data; Use a machine learning model to analyze the preprocessed data to detect abnormal operations or predict impending faults; Trigger an early warning response according to the detection or prediction result.

[0034] Through the above specific implementation manners, the intelligent monitoring system of the present invention can realize the real-time status monitoring, fault diagnosis and life prediction of the weftless tape machine, improve production efficiency and product quality, and reduce maintenance costs and downtime.

[0035] In the description of this specification, the description referring to terms such as "one embodiment", "example", "specific example", etc. means that the specific features, structures, materials or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in a suitable manner in any one or more embodiments or examples.

[0036] The preferred embodiments of the present invention disclosed above are only used to help illustrate the present invention. The preferred embodiments do not describe all the details in detail, nor do they limit the invention to the specific embodiments described. Obviously, many modifications and variations can be made according to the content of this specification. These embodiments are selected and specifically described in this specification in order to better explain the principles and practical applications of the present invention, so that those skilled in the art can well understand and utilize the present invention. The present invention is only limited by the claims and their full scope and equivalents.

Claims

1. An intelligent monitoring system for a weftless belt machine based on real-time data acquisition, characterized in that: include: Data acquisition module: acquires real-time data from multiple heterogeneous sensors through an industrial communication network, wherein the sensors include a vibration sensor deployed at the spindle bearing, a yarn tension sensor installed at the yarn guide roller, an optical sensor arranged in the winding area, and at least one selected from a temperature sensor, a yarn break sensor, a position sensor, a speed sensor, a sound sensor, and a power consumption sensor; Data processing module: includes distributed edge computing nodes for performing data cleaning, filtering, normalization, segmentation, and synchronization operations; Intelligent analysis module: It includes a feature extraction unit, an anomaly detection unit, a fault diagnosis unit, a life prediction unit and a data fusion unit, wherein the feature extraction unit is used to convert the preprocessed raw data into a multi-dimensional feature vector and output it to the subsequent unit; the anomaly detection unit identifies outliers in the data based on the feature vector and outputs the anomaly score to the data fusion unit; the fault diagnosis unit classifies the feature vector, determines the specific fault type and confidence level, and outputs them to the data fusion unit; the life prediction unit predicts the remaining service life based on the historical feature sequence and outputs the probability distribution parameters to the data fusion unit; the data fusion unit integrates the outputs of the anomaly detection unit, the fault diagnosis unit and the life prediction unit to generate a final decision and trigger an early warning response; Early warning response module: triggers multi-level alarms according to the abnormality level, generates a diagnostic report containing failure probability and disposal suggestions, and executes equipment protection strategies; The data cleaning, filtering, normalization, segmentation and synchronization operations include: The Z-score method was used to identify outliers, and missing values ​​were filled by cubic spline interpolation; Apply a low-pass filter to filter the data; Use minimum-maximum normalization to scale the data to the [0,1] interval; Use a sliding time window with a length of 1024 sampling points to segment the data; Multi-sensor data timestamp alignment is achieved based on the NTP protocol.

2. According to claim 1, the intelligent monitoring system for a weftless belt machine based on real-time data acquisition is characterized in that: The anomaly detection unit identifies outliers in the data based on the feature vector and outputs anomaly scores to the data fusion unit, specifically including: Feature standardization: performing Z-score standardization on the original feature vector output by the feature extraction unit; Isolation forest modeling: Build an isolation forest model containing 100-200 decision trees, with a maximum depth of 8-12 layers for each tree; randomly partition the standardized feature vector until each sample is isolated or the maximum tree depth is reached; Anomaly score calculation: Calculate the average path length of each sample in the isolation forest , and then convert the average path length into anomaly score : ; in, is the normalization factor for the average path length of the tree, is the number of training samples; Dynamic threshold setting: Set dynamic thresholds based on the anomaly score distribution of historical normal data : ; in, and are the mean and standard deviation of the abnormal scores of normal samples, is a dynamically adjustable coefficient; Time series smoothing: Apply exponentially weighted moving average (EWMA) filtering to the anomaly scores in continuous time windows to obtain smoothed anomaly scores; Result output: The smoothed anomaly score is output to the data fusion unit.

3. According to claim 1, the intelligent monitoring system for a weftless belt machine based on real-time data acquisition is characterized in that: The fault diagnosis unit classifies the feature vector, determines the specific fault type and confidence level, and outputs the result to the data fusion unit, which specifically includes: Feature preprocessing: Performing L2 normalization on the original feature vector output by the feature extraction unit; Principal component analysis was used to reduce feature dimensionality, and principal components with cumulative variance contribution rate ≥ 95% were retained; Multi-scale feature extraction: Construct a multi-branch convolutional neural network, where each branch uses convolution kernels of different sizes to extract features of different scales; Perform global average pooling on the output of each branch to obtain multi-scale feature representation; Attention Mechanism Weighting: Calculate the importance weight of each dimension feature through the attention mechanism; Multiply the original feature vector by the attention weight element by element to get the weighted feature vector; Fault classification and confidence calculation: The weighted feature vector is input into the fully connected layer for classification, and the Softmax activation function is used to output the probability distribution of various types of faults; The maximum probability value is taken as the confidence score of the classification result; Uncertainty Quantification: The Monte Carlo Dropout method is used to quantify classification uncertainty, specifically: Keep Dropout activated during the test phase and perform T forward propagations, where T ≥ 20; Calculate the variance of the probability distribution of T predictions; Use variance as a measure of uncertainty in classification results; Result output and verification: Outputting the classification results, confidence scores and uncertainty measures to the data fusion unit; When the confidence is less than the confidence threshold or the uncertainty is greater than the uncertainty threshold, the secondary verification process is triggered.

4. According to claim 1, the intelligent monitoring system for a weftless belt machine based on real-time data acquisition is characterized in that: The life prediction unit predicts the remaining service life based on the historical feature sequence and outputs the probability distribution parameters to the data fusion unit, specifically including: Time series feature construction: Arrange the multidimensional feature vectors output by the feature extraction unit in chronological order to construct a sliding time window of length L ,in, for The feature vector at the moment; Perform differential processing on the feature sequence in each time window and calculate the rate of change between adjacent time steps ; Combine the original features and the rate of change features to form an enhanced feature sequence ; Sequence-to-Sequence Modeling: Build a gated recurrent unit (GRU) model with at least 2 hidden layers, each with 64-128 neurons. Use the attention mechanism to weight the input sequence: ; in, For the The hidden state of time steps, , is the trainable weight matrix, is the bias term, is the activation function; Input the weighted hidden state sequence into the fully connected layer and output the predicted remaining useful life; Probability distribution fitting: Convert the point prediction results of the GRU model to Weibull distribution parameters: Collect historical forecast errors and calculate shape parameters : ; in, is the ratio of the sample standard deviation to the mean, is the gamma function; Calculate scale parameters : ; in, The remaining useful life predicted by the GRU model; Uncertainty Quantification: Monte Carlo Dropout method is used to estimate forecast uncertainty: Keep Dropout activated during the test phase and perform M forward propagations; Fit the Weibull distribution for each prediction result and obtain M sets of parameters ; Calculate the statistical distribution of the parameters and output 95% confidence intervals; Dynamically update prediction results: When new sensor data arrives, the prediction results are updated using an exponential weighting method: ; in, For the forgetting factor, is the current predicted value, is the last predicted value; Result output: The Weibull distribution parameters and their 95% confidence intervals are output to the data fusion unit as a probability density function.

5. According to claim 1, the intelligent monitoring system for a weftless belt machine based on real-time data acquisition is characterized in that: The data fusion unit integrates the outputs of the anomaly detection unit, the fault diagnosis unit, and the life prediction unit to generate a final decision and trigger an early warning response, specifically including: Data preprocessing and standardization: Map the anomaly score output by the anomaly detection unit to the interval [0,1] and convert it into anomaly probability; Normalize the confidence scores of various faults output by the fault diagnosis unit; Convert the Weibull distribution parameters output by the life prediction unit into the probability density function of the remaining service life; Time and space consistency check: Check the consistency of the timestamps of the outputs of the three analysis units, and perform interpolation alignment on the data with a time deviation of more than 50ms; Verify spatial consistency: Ensure that the physical location of the anomaly detection localization matches the fault diagnosis result; Weight of evidence allocation: Dynamically adjust the weight of each analysis unit according to the equipment operation status: Startup phase: the weight of the anomaly detection unit increases by 20%; Steady-state operation: The weight of the fault diagnosis unit is increased by 15%; Approaching the maintenance cycle: the weight of the life prediction unit increases by 25%; Calculate the credibility coefficient of each analysis unit based on historical accuracy; Decision Fusion: The Dempster-Shafer evidence theory is used to fuse the abnormal probability and fault type confidence: Convert abnormal probability into basic probability allocation BPA; Convert the fault type confidence into the corresponding BPA; Calculate the fused BPA through evidence synthesis rules; Combine lifespan prediction results to determine comprehensive risk index ; Decision threshold setting: Based on historical equipment data and expert knowledge, three levels of risk thresholds are set: High risk: , immediately shut down for inspection and generate an emergency maintenance work order; Medium risk: , reduce operating load and arrange preventive maintenance; Low risk: , operate normally, record and monitor the status; Early warning response trigger: Trigger early warning responses of corresponding levels based on the comprehensive risk index; Generates a diagnostic report containing the abnormality location, fault type, risk level and treatment recommendations; The decision results are fed back to the model training unit for continuous optimization of the model.

6. The intelligent monitoring system for a weftless belt machine based on real-time data acquisition according to claim 1 is characterized in that: The intelligent analysis module also includes: Based on the attention mechanism, the multi-sensor features are integrated and the weight coefficient of each sensor feature is automatically adjusted according to different working conditions; The transfer learning method is used to migrate the model parameters trained on similar equipment to the target weftless belt machine model, reducing the training sample requirements of the target model.

7. The intelligent monitoring system for a weftless belt machine based on real-time data acquisition according to claim 1 is characterized in that: The distributed edge computing node is configured as follows: Perform real-time data preprocessing close to the sensor; Through the edge-cloud collaborative architecture, only feature data and abnormal events are transmitted to the cloud server; Implement edge incremental learning of models based on locally stored historical data.

8. An intelligent monitoring method for a weftless belt machine based on real-time data acquisition, characterized in that: include: Acquire data from multiple heterogeneous sensors in real time; Performing data preprocessing operations on the acquired real-time data; analyzing the preprocessed data using a machine learning model to detect operational anomalies or predict impending failures; Trigger early warning responses based on detection or prediction results.

Images