Spiral duct predictive operation and maintenance method and system based on digital twinning

Abstract

The invention discloses a spiral air duct predictive operation and maintenance method and system based on digital twinning, and relates to the technical field of air duct system state monitoring and operation and maintenance, and the method comprises the following steps: deploying a high-frequency acoustic emission AE sensor and a low-frequency vibration acceleration sensor at key nodes of a spiral air duct, and taking a fan start-stop signal and operation power as working condition labels; a high-frequency acquisition mode is automatically triggered, noise reduction processing is carried out, and a multi-mode acoustic vibration original signal data set which is provided with a time sequence and a working condition label and is subjected to preliminary noise reduction is output; the high-frequency acoustic emission AE sensor and the low-frequency vibration acceleration sensor are arranged in the spiral duct, the health state of the thin-wall structure is monitored in real time through the high-frequency acoustic emission technology, weak early-stage fault signals in the duct can be accurately captured and early warning can be given out in the fault germination stage of fretting wear, sealant aging and the like of the spiral duct, and the fault early-warning effect is achieved. Real-time monitoring, early warning and accurate operation and maintenance of the spiral air pipe are achieved.

Description

Technical Field

[0001] This invention relates to the field of duct system condition monitoring and operation and maintenance technology, specifically to a predictive operation and maintenance method and system for spiral ducts based on digital twins. Background Technology

[0002] Spiral ducts are widely used in building ventilation and air conditioning systems. Their operating status directly affects the performance and energy consumption of the entire ventilation system. However, during long-term operation, duct systems may face problems such as fretting wear of the interlocking joints and aging of the sealant. If these problems are not detected and dealt with in time, they can lead to serious consequences such as duct leakage, increased energy consumption, and equipment damage.

[0003] For example, spiral ducts are made of steel strips wound and interlocked in a spiral. They are subject to cyclic stresses from fan start-up and shutdown, wind pressure fluctuations, slight deformation of the building structure, or vibration transmission. The interlocking joints of the spiral joints experience fretting wear due to the cyclic stress, which can lead to micro-cracks that can propagate and eventually cause fatigue fracture or air leakage. However, this wear is extremely small and invisible to the naked eye, making it impossible to detect by traditional inspections. But long-term accumulation can lead to a decrease in the tightness of the interlocking joints, forming micro-cracks that are not discovered until the air leakage increases significantly or the duct breaks in a localized area. In addition, the joints of spiral ducts are usually sealed with sealant. Over time and due to factors such as changes in environmental temperature and humidity, the sealant will gradually age, crack, and fail, leading to leaks in the duct. Leaks not only reduce ventilation efficiency but may also result in energy waste.

[0004] Application number CN202410565048.0 discloses a real-time monitoring system and method for air conditioning ducts. The system includes a measurement terminal installed inside the air conditioning duct, housing a high-precision microelectronic flow sensor, a particle counter, a noise sensor, and a fiber optic temperature and humidity sensor. The air conditioning unit itself contains a compressor detection block, a condenser detection module, a throttling device detection module, and an evaporator detection module. A server terminal is also included. The measurement terminal can monitor data such as flow rate, velocity, dust quantity, noise, temperature, and humidity within the air conditioning duct. It can also detect parameters of various components within the air conditioning unit to determine their normal operation. A high-precision digital twin model can provide early warnings of potential future problems, facilitating user prevention. When a fault occurs, it can analyze parameters of potentially abnormal areas based on previous data, enabling troubleshooting and facilitating user maintenance. However, it lacks consideration for the failure modes of the continuous spiral seams and joint sealant in the duct, such as fretting wear, fatigue cracks, and sealant aging and leakage, leading to untimely maintenance and high costs.

[0005] Existing technologies have the following shortcomings: Current duct monitoring and maintenance technologies mainly rely on periodic manual inspections or simple post-incident repairs, which cannot capture the weak fault signals of the aforementioned slow and hidden early faults. The post-incident repair mode leads to unplanned downtime and high repair costs, while periodic maintenance lacks specificity, has poor economic efficiency, and cannot achieve precise operation and maintenance of spiral ducts. Therefore, there is an urgent need to provide a predictive operation and maintenance method and system for spiral ducts based on digital twins, which can achieve early detection, accurate diagnosis, and prediction of remaining life of faults such as micro-motion wear and sealant aging in spiral ducts, and provide optimal maintenance decisions.

[0006] The information disclosed in the background section is only intended to enhance the understanding of the background of this disclosure, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0007] The purpose of this invention is to provide a predictive operation and maintenance method and system for spiral ducts based on digital twins. This invention accurately captures weak early fault signals inside the duct and issues early warnings through high-frequency acoustic emission (AE) sensors and low-frequency vibration acceleration sensors. It uses MVSPC and unsupervised learning algorithms to discover fault anomaly patterns, and integrates a hybrid diagnostic method that combines a physical mechanism knowledge base and a data-driven model. This further accurately identifies and distinguishes fault types, occurrence events, and fault locations based on fault anomaly patterns. Furthermore, it uses physical fatigue cumulative damage models and data-driven models with different physical mechanisms to differentiate and predict remaining lifespan, thereby solving the problems mentioned in the background art.

[0008] To achieve the above objectives, the present invention provides the following technical solution: a predictive operation and maintenance method for spiral ducts based on digital twins, comprising the following steps:

[0009] S1. Deploy high-frequency acoustic emission (AE) sensors and low-frequency vibration acceleration sensors at key nodes of the spiral duct. Use the fan start / stop signal and operating power as operating condition labels to automatically trigger the high-frequency acquisition mode and perform noise reduction processing. Output a multi-modal acoustic and vibration raw signal dataset with time sequence, operating condition labels and preliminary noise reduction, including acoustic emission signals and vibration signals.

[0010] S2. Extract features from acoustic emission signals and vibration signals to form multidimensional feature vectors, and input them into the multivariate statistical process control (MVSPC) model. Monitor the status of the spiral duct by calculating the T² statistic and squared prediction error. Use an unsupervised learning algorithm to learn the fused multidimensional feature space, identify abnormal points that deviate from the normal baseline state, and output structured event alarms.

[0011] S3. Based on structured event alarms, establish a physical mechanism knowledge base of fault characteristics, combine a multi-class machine learning model trained with historical data to determine the fault mode and output confidence level, and achieve fault segment-level localization by analyzing the time difference / energy attenuation of sensor signals and output an enhanced diagnostic report.

[0012] S4. Utilizing the enhanced diagnostic report, the remaining life is predicted differentially using a physics-based fatigue cumulative damage model and a data-driven model. The remaining life prediction and decision optimization module has a built-in cost model that converts the predicted remaining life into risk costs. These costs are compared with planned maintenance costs, and a decision tree model is used to recommend the economically optimal maintenance timing and strategy, generating an executable strategy.

[0013] Optionally, the high-frequency acoustic emission (AE) sensor has a frequency range of 20kHz-1MHz and is used to capture transient elastic wave acoustic emission signals released by structural changes such as fretting wear, microcrack initiation and propagation at the spiral interlocking seam inside the spiral duct material.

[0014] The low-frequency vibration acceleration sensor has a frequency range of 0.1Hz-2kHz and is used to monitor structural vibration signals caused by decreased connection stiffness due to sealant aging, airflow impact, and fan vibration.

[0015] Optionally, the steps for outputting the multimodal acoustic vibration raw signal dataset are as follows:

[0016] High-frequency acoustic emission (AE) sensors and low-frequency vibration acceleration sensors are deployed at key nodes of the spiral duct.

[0017] Set operating condition labels, including fan start-up, shutdown signals, and operating power, and monitor fan start-up and shutdown signals and operating power in real time;

[0018] When a change in the start / stop switch signal of the wind turbine is detected or the rate of change of operating power exceeds a preset threshold, the high-frequency acquisition mode is automatically triggered.

[0019] When the high-frequency acquisition mode is triggered, all high-frequency acoustic emission (AE) sensors deployed on the duct are activated to acquire acoustic emission signals at a frequency of 1MHz. Simultaneously, all low-frequency vibration acceleration sensors deployed on the duct are activated to acquire vibration signals at a frequency of 2kHz. A fixed duration is sampled as the total length of the sampled signal.

[0020] The collected acoustic emission and vibration signals are processed by wavelet transform noise reduction using a low-pass filter to generate noise-reduced acoustic emission and vibration signals.

[0021] Add timing information and operating condition labels to the noise-reduced acoustic emission and vibration signals;

[0022] The acoustic emission signal and vibration signal, which have been labeled with time sequence and operating condition and have undergone preliminary noise reduction, are combined into a multimodal acoustic and vibration raw signal dataset for output.

[0023] Optionally, the steps for extracting the multidimensional feature vector are as follows:

[0024] The continuous acoustic emission signal and vibration signal are divided into analysis frames of fixed length, and each frame contains 1024 sampling points;

[0025] For each frame of data, multiple feature indicators are calculated in parallel: for acoustic emission signals, ring count, energy, amplitude, and wavelet packet energy spectrum are extracted; for vibration signals, root mean square value, kurtosis, and envelope spectrum are extracted.

[0026] The feature values ​​of all calculated feature indicators are combined in order into a multidimensional vector to form a multidimensional feature vector.

[0027] Optionally, the steps for monitoring the spiral duct status using the multivariate statistical process control (MVSPC) model are as follows:

[0028] Under normal operating conditions of the spiral duct, a large amount of multidimensional feature vector data is collected as samples, and the sample mean vector and covariance matrix are calculated.

[0029] For the multidimensional feature vector input at the new sampling point, calculate the T² statistic and the squared prediction error SPE;

[0030] Determine the control limits based on the distribution of the T² statistic;

[0031] The T² statistic calculated in real time is compared with the determined control limits to determine whether the spiral duct is in a normal state.

[0032] Optionally, the steps for identifying outliers using the unsupervised learning algorithm are as follows:

[0033] Subsamples are randomly selected from the multidimensional feature vector data as the training set, and the Isolation Forest algorithm in the unsupervised learning algorithm is used to train the training set.

[0034] Randomly select features and segmentation values ​​to construct multiple isolation trees to build an isolation tree forest;

[0035] For each data point, traverse each isolation tree and record the path length from the root node to the leaf node.

[0036] Calculate the anomaly score based on the average path length across all trees;

[0037] The anomaly score is used to assess the anomaly status of data points and determine whether each data point is an anomaly.

[0038] Optionally, the output steps of the structured event alarm are as follows:

[0039] The monitoring results of the MVSPC model and the abnormal scores of the unsupervised learning algorithm are fused to form abnormal evidence.

[0040] Determine the anomaly type based on the characteristics of the anomaly and its location;

[0041] By combining the location information from the sensors, the location of the anomaly can be determined;

[0042] Record the time when the exception occurred;

[0043] The severity of the deviation is scored by calculating the Euclidean distance based on the degree of deviation of outliers from the normal baseline.

[0044] The anomaly type, location, time, and severity score are combined into a structured event alert output.

[0045] Optionally, the steps for establishing the physical mechanism knowledge base are as follows:

[0046] List all types of faults that need to be diagnosed, including fretting wear of the interlocking joint, aging and leakage of sealant, external impact, and normal condition;

[0047] Collect various data related to spiral duct failures, including structured event alerts, spiral duct physics principles, experimental data, and expert experience;

[0048] Extract and map the physical mechanism knowledge related to fault characteristics from the collected data to determine the typical signal characteristics corresponding to each fault mode;

[0049] Mathematical modeling is performed on the typical performance of each multidimensional feature under a specific failure mode, and the result is expressed in the form of a probability distribution or numerical range.

[0050] By integrating well-represented knowledge into a knowledge base, a physical mechanism knowledge base is established, which facilitates subsequent querying and use.

[0051] The steps of the multi-class machine learning model in determining the fault mode and outputting the confidence score are as follows:

[0052] A logistic regression multi-class classification model is identified, combined with a physical mechanism knowledge base;

[0053] Historical data is divided into multi-dimensional feature vectors and fault mode types to train the multi-classification model, resulting in a well-trained multi-classification model.

[0054] Input a new multidimensional feature vector into the trained multi-classification model and output the predicted fault mode and the corresponding confidence level.

[0055] Optionally, the steps of the physics-based fatigue cumulative damage model to predict remaining life are as follows:

[0056] The low-frequency vibration acceleration sensor is converted into a stress time history at the joint through a mechanical model. The load spectrum of different stress amplitudes and stress cycle numbers is obtained by statistically analyzing the stress time history using the rainflow counting method, thereby determining the stress spectrum that the spiral pipe experiences during operation.

[0057] Based on the stress spectrum, the fatigue damage model of Miner's linear cumulative damage theory is used to calculate the accumulated damage that has occurred.

[0058] The remaining service life is predicted based on the cumulative damage under the current operating conditions of the spiral pipe.

[0059] The steps for predicting remaining lifetime using the data-driven model are as follows:

[0060] Collect historical data related to the operating status of the spiral pipeline, including multimodal acoustic and vibration raw signal datasets, structured event alarms and enhanced diagnostic reports, and extract multiple feature data that can characterize the overall health status of the pipeline to construct a health index;

[0061] The historical health index HI values ​​are arranged in chronological order to form a degradation time series. A time series prediction LSTM model is trained using the historical data to learn the degradation pattern of the health index HI.

[0062] The new input feature vector is transformed into the current degradation time series, input into the trained LSTM model, and the predicted remaining lifetime is output.

[0063] The predictive operation and maintenance system for spiral duct based on digital twins includes a heterogeneous sensor network module for spiral duct deployment: acoustic emission sensors, vibration sensors, data acquisition devices and operating condition interfaces are deployed in the spiral duct to form a heterogeneous sensor network.

[0064] Data acquisition and preprocessing module: Real-time acquisition of sensor signals using industrial IoT protocols, using fan start / stop signals and operating power as operating condition labels, collaboratively triggering high-frequency acquisition mode and performing adaptive noise reduction processing, outputting a preliminary noise-reduced multimodal acoustic and vibration raw signal dataset;

[0065] Feature extraction and anomaly detection module: Extracts features from acoustic emission signals and vibration signals in the multimodal acoustic and vibration raw signal dataset, inputs multidimensional features into the multivariate statistical process control (MVSPC) model and unsupervised learning algorithm, and outputs structured event alarms;

[0066] Fault diagnosis and location module: Based on structured event alarms, combined with a physical mechanism knowledge base and a multi-class machine learning model, it determines the fault mode and confidence level, realizes fault segment-level location by analyzing sensor signals, and outputs an enhanced diagnostic report;

[0067] The Remaining Life Prediction and Decision Optimization Module utilizes enhanced diagnostic reports to predict remaining life using both a physics-based fatigue cumulative damage model and a data-driven model. It calculates risk costs, compares them with planned maintenance costs, recommends the most economically optimal maintenance timing and strategy, and generates executable strategies.

[0068] The technical effects and advantages provided by the present invention in the above technical solution are as follows:

[0069] This invention utilizes high-frequency acoustic emission (AE) sensors and low-frequency vibration acceleration sensors deployed within spiral ducts to monitor the health status of thin-walled structures in real time using high-frequency acoustic emission technology. It can accurately capture weak early fault signals within the duct and issue warnings during the nascent stages of faults such as fretting wear and sealant aging, achieving real-time monitoring, early warning, and precise operation and maintenance of spiral ducts. This results in early detection and warning of anomalies during the operation of spiral ducts. Furthermore, by employing MVSPC... This approach utilizes unsupervised learning algorithms to learn hidden fault and anomaly patterns in multidimensional feature vector space that are not readily apparent to the human eye. Furthermore, by constructing a hybrid diagnostic method that integrates a physical mechanism knowledge base and a data-driven model, it accurately identifies fault types, occurrence events, and fault locations, avoiding blind manual repairs and significantly improving operational efficiency. Additionally, it employs differentiated prediction of remaining lifespan using physics-based fatigue cumulative damage models and data-driven models with different physical mechanisms, predicting fault occurrence times in advance and rationally scheduling maintenance plans. This avoids emergency downtime losses and unnecessary repair costs, improving the operational efficiency and economy of the duct system. Moreover, it quantifies fault prediction into a cost dimension, recommending optimal maintenance times by comparing and optimizing maintenance actions for different faults. This facilitates the development of maintenance planning strategies based on the principle of economic optimization, maximizing return on investment. Attached Figure Description

[0070] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0071] Figure 1 This is a flowchart of the predictive operation and maintenance method for spiral ducts based on digital twins according to the present invention.

[0072] Figure 2This is a block diagram of the predictive operation and maintenance system for spiral ducts based on digital twins, as described in this invention. Detailed Implementation

[0073] Exemplary embodiments will now be described more fully with reference to the accompanying drawings. However, these exemplary embodiments can be implemented in many forms and should not be construed as limited to the examples set forth herein; rather, they are provided so that the description of this disclosure will be more complete and fully convey the concept of the exemplary embodiments to those skilled in the art.

[0074] Example 1

[0075] This invention provides, for example Figure 1 The predictive operation and maintenance method for spiral ducts based on digital twins, as shown, includes the following steps:

[0076] S1. Deploy high-frequency acoustic emission (AE) sensors and low-frequency vibration acceleration sensors at key nodes of the spiral duct. Use the fan start / stop signal and operating power as operating condition labels to automatically trigger the high-frequency acquisition mode and perform noise reduction processing. Output a multi-modal acoustic and vibration raw signal dataset with time sequence, operating condition labels and preliminary noise reduction, including acoustic emission signals and vibration signals.

[0077] Specifically, the high-frequency acoustic emission (AE) sensor has a frequency range of 20kHz-1MHz and is used to capture transient elastic wave acoustic emission signals released by microstructural changes such as fretting wear, microcrack initiation and propagation at the spiral interlocking seam inside the spiral duct material.

[0078] The low-frequency vibration acceleration sensor has a frequency range of 0.1Hz-2kHz and is used to monitor structural vibration signals caused by decreased connection stiffness due to sealant aging, airflow impact, and fan vibration.

[0079] Specifically, the steps for outputting the multimodal acoustic vibration raw signal dataset are as follows:

[0080] High-frequency acoustic emission (AE) sensors and low-frequency vibration acceleration sensors are deployed at key nodes of the spiral duct. The high-frequency AE sensors can capture high-frequency acoustic emission signals generated by internal material deformation and crack propagation in the duct, while the low-frequency vibration acceleration sensors measure the vibration of the duct.

[0081] Set operating condition labels, including fan start-up and shutdown signals and operating power, and monitor fan start-up and shutdown signals and operating power in real time. The expression for the operating condition label is: In the formula, This is represented as a working condition label. This indicates the start signal for the wind turbine. This indicates a fan shutdown signal. This is expressed as the operating power of the wind turbine;

[0082] When a change in the wind turbine start / stop switch signal is detected, or when the rate of change in operating power exceeds a preset threshold, the high-frequency acquisition mode is automatically triggered. The expression for automatically triggering the high-frequency acquisition mode is as follows: , In the formula, Indicated as in The system continuously monitors changes in the fan start / stop switch signal or the rate of change in operating power to automatically trigger the high-frequency acquisition mode. Expressed as operating power, Represented as time, This represents the trigger threshold for the rate of increase in wind turbine operating power. This represents the trigger threshold for the rate of decrease in wind turbine operating power. This represents the start / stop switch signal for the fan. This indicates that a change in the fan start / stop switch signal has been detected. This indicates that the wind turbine is in The rate of change of power detected at any given time;

[0083] When the high-frequency acquisition mode is triggered, all high-frequency acoustic emission (AE) sensors deployed on the duct are activated to acquire acoustic emission signals at a frequency of 1MHz. Simultaneously, all low-frequency vibration acceleration sensors deployed on the duct are activated to acquire vibration signals at a frequency of 2kHz. A fixed duration is sampled as the total length of the sampled signal.

[0084] The acquired acoustic emission and vibration signals are subjected to wavelet transform denoising using a low-pass filter. The input acoustic emission / vibration signal is decomposed into components of different scales and frequencies. After wavelet decomposition, thresholding, and wavelet reconstruction, the wavelet-transform-denoised acoustic emission / vibration signal is obtained, generating the denoised acoustic emission and vibration signals. The expression for the wavelet transform denoising of the acoustic emission signal using the low-pass filter is as follows: ,and ,as well as In the formula, Represented as the acoustic emission signal at the th The value of each sampling point, These represent the scale and frequency in wavelet decomposition, respectively. Represented as wavelet coefficients, Represented as wavelet basis functions, The sign function is represented as the wavelet coefficients. It is expressed as the absolute value of the wavelet coefficients. This is represented as a soft threshold. These are the wavelet coefficients after processing with the soft thresholding function. The acoustic emission signal after wavelet transform and noise reduction is represented as the signal at the th... The value of each sampling point;

[0085] The simplified expression for wavelet transform noise reduction of vibration signals using a low-pass filter is: In the formula, The vibration signal after wavelet transform and noise reduction is represented as the signal at the th wavelet transform. The value of each sampling point, This is represented as the inverse discrete wavelet transform operation, which reconstructs the signal from the wavelet coefficients after soft thresholding. This is represented as a soft thresholding operation, which will process values ​​less than the soft threshold. The wavelet coefficients are set to zero to remove noise. This is represented by the discrete wavelet transform operation, which decomposes the signal into components of different scales and frequencies. Represented as the vibration signal at the th The value of each sampling point;

[0086] Add timing information and operating condition tags to the denoised acoustic emission and vibration signals, where the timing information is the sampling timestamp;

[0087] The acoustic emission and vibration signals, labeled with time series and operating conditions and preliminarily denoised, are combined to form a multimodal acoustic-vibration raw signal dataset for output. The expression for this multimodal acoustic-vibration raw signal dataset is as follows: In the formula, This is represented as a dataset of raw acoustic and vibration signals from multiple modes. Represented as a sampling timestamp, This is represented as a working condition label. This indicates whether the high-frequency acquisition mode is automatically triggered. This represents the acoustic emission signal used for initial noise reduction. This represents the vibration signal as initially denoised.

[0088] To further clarify, the key nodes for deploying high-frequency acoustic emission (AE) sensors and low-frequency vibration acceleration sensors in spiral ducts include the middle section of the long straight section of the spiral duct, supports, elbows, flange connections, and the inlet and outlet of the fan.

[0089] High-frequency acoustic emission (AE) sensors and low-frequency vibration acceleration sensors collect multimodal acoustic and vibration raw signal datasets related to the operating status of the spiral duct in real time. These datasets provide raw data for subsequent analysis and processing, ensuring that the system can accurately perceive the actual operating status of the spiral duct.

[0090] S2. Extract features from acoustic emission signals and vibration signals to form multidimensional feature vectors, and input them into the multivariate statistical process control (MVSPC) model. Monitor the status of the spiral duct by calculating the T² statistic and squared prediction error. Use an unsupervised learning algorithm to learn the fused multidimensional feature space, identify abnormal points that deviate from the normal baseline state, and output structured event alarms, including anomaly type, location, time, and severity score.

[0091] Specifically, the steps for extracting multidimensional feature vectors are as follows:

[0092] The continuous acoustic emission signal and vibration signal are divided into analysis frames of fixed length, and each frame contains 1024 sampling points;

[0093] For each frame of data, multiple feature indicators are calculated in parallel: for acoustic emission signals, ring count, energy, amplitude, and wavelet packet energy spectrum are extracted; for vibration signals, root mean square value, kurtosis, and envelope spectrum are extracted.

[0094] The expressions for the feature indices extracted from acoustic emission signals include: , , , In the formula, This is represented as a ring count. Represented as a counting function, Represented as the discrete acoustic emission signal of the th The voltage amplitude corresponding to each sampling point This is expressed as the threshold voltage. Represented as energy, Represented as the first One sampling point, This represents the total number of sampling points. This can be represented as summing the squares of the amplitudes at all sampling points of the discrete acoustic emission signal. Represented as amplitude, This is represented as a function that takes the maximum value. Represented as the first The absolute value of the voltage amplitude corresponding to each sampling point is taken. Represented as the first Wavelet packet energy spectrum of each frequency band Represented as a frequency band index, and , Represented as the first The first frequency band Wavelet packet coefficients, Represented as the first At a certain point in time, This represents the number of wavelet packet decomposition levels;

[0095] The expressions for the feature indices extracted from vibration signals include: , , In the formula, Represented as root mean square value, This represents the total number of sampling points. Represented as the vibration signal at the th Vibration amplitude at each sampling point Indicated as kurtosis, It is expressed as the mean of the vibration amplitude of the vibration signal at all sampling points. It is expressed as the standard deviation of the vibration amplitude of the vibration signal at all sampling points. Represented as an envelope signal, the envelope spectrum is obtained by performing a Fourier transform (FFT). Represented as the Hilbert transform operator, It is represented as the absolute value of the analytic signal, and , Represented as the imaginary unit;

[0096] The eigenvalues ​​of all calculated feature indicators are combined sequentially into a multidimensional vector, forming a multidimensional feature vector. The expression for the multidimensional feature vector is: In the formula, Represented as a multidimensional feature vector, This is represented as the envelope spectrum extracted from the vibration signal.

[0097] Specifically, the steps for monitoring the status of a spiral duct using the Multivariate Statistical Process Control (MVSPC) model are as follows:

[0098] Under normal operating conditions of the spiral duct, a large amount of multidimensional feature vector data is collected as samples, and the sample mean vector and covariance matrix are calculated. The formula for calculating the sample mean vector is as follows: In the formula, Represented as a sample mean vector, This represents the total number of samples collected for multidimensional feature vector data. Represented as the first A multidimensional feature vector;

[0099] The formula for calculating the covariance matrix is: In the formula, Represented as the sample covariance matrix, Represented as vector transpose;

[0100] For the multidimensional feature vector input at new sampling points, calculate the T² statistic and the squared prediction error (SPE). The formula for calculating the T² statistic is: In the formula, It is expressed as the T² statistic, and the larger the T² statistic value, the further the new sampling point is from the multivariate center of the normal state. This is represented as a multidimensional feature vector input at the new sampling point. It is represented as the inverse of the covariance matrix;

[0101] The formula for calculating the squared prediction error (SPE) is as follows: In the formula, This is expressed as the squared prediction error. Represented as an identity matrix, This is represented as the principal load matrix extracted from the normal training data by the PCA model. It is expressed as the squared Euclidean norm of a vector. Represented as a projection matrix, Represented as the residual projection matrix, Represented as a residual vector;

[0102] Determine the control limits based on the distribution of the T² statistic;

[0103] The real-time calculated T² statistic is compared with the determined control limit to determine whether the spiral duct is in a normal state. When any calculated T² statistic exceeds the control limit, the spiral duct is determined to be in an abnormal state.

[0104] Specifically, the steps for unsupervised learning algorithms to identify outliers are as follows:

[0105] Subsamples are randomly selected from the multidimensional feature vector data as the training set, and the Isolation Forest algorithm in the unsupervised learning algorithm is used to train the training set.

[0106] Randomly select features and segmentation values ​​to construct multiple isolation trees to build an isolation tree forest;

[0107] For each data point, traverse each isolation tree and record the path length from the root node to the leaf node.

[0108] Anomaly scores are calculated based on the average path lengths across all trees. The formula for calculating the anomaly score is as follows: ,and In the formula, The anomaly score is represented as the feature vector data point of the subsample dimension. Represented as subsample size, Represented as, Represented as subsample dimensional feature vector data points The average path length across all isolated trees. It is represented as the harmonic number of the standardized path length over the subsamples. Represented as harmonic numbers;

[0109] The abnormality score is used to assess the abnormality status of data points. The closer the abnormality score is to 1, the more likely it is to be an abnormal data point. Points with an abnormality score much less than 0.5 are likely to be normal points. This is used to determine whether each data point is an abnormal point.

[0110] Specifically, the output steps for structured event alerts are as follows:

[0111] The monitoring results of the MVSPC model and the abnormal scores of the unsupervised learning algorithm are fused to form abnormal evidence.

[0112] Based on the characteristics of the anomaly and its location, determine the type of anomaly, including spiral duct impact failure, stiffness change, leakage or blockage, etc.

[0113] By combining the location information from the sensors, the location of the anomaly can be determined;

[0114] Record the time when the exception occurred;

[0115] The severity of the deviation is scored by calculating the Euclidean distance based on the degree of deviation of outliers from the normal baseline.

[0116] The anomaly type, location, time, and severity score are combined into a structured event alert output.

[0117] S3. Based on structured event alarms, establish a physical mechanism knowledge base of fault characteristics, combine a multi-class machine learning model trained with historical data to determine the fault mode and output confidence level, and achieve fault segment-level localization by analyzing the time difference / energy attenuation of sensor signals, and output an enhanced diagnostic report, including fault location, abnormal cause and confidence level.

[0118] Specifically, the steps for establishing the physical mechanism knowledge base are as follows:

[0119] List all fault types requiring diagnosis, including fretting wear of the interlocking seams, aging and leakage of sealant, external impact, and normal conditions. The fault type is represented as follows: ,and In the formula, Represented as the first Types of faults This indicates a fault type characterized by fretting wear in the interlocking seam. This indicates a fault type characterized by sealant aging and leakage. This indicates the type of failure caused by external impact. This indicates a normal state;

[0120] Collect various data related to spiral duct failures, including structured event alerts, spiral duct physics principles, experimental data, and expert experience;

[0121] Extract and map physical mechanism knowledge related to fault characteristics from the collected data to determine the typical signal characteristics corresponding to each fault mode, such as the variation law of sensor signals under different fault modes, the causes of fault occurrence and influencing factors;

[0122] Mathematical modeling is performed on the typical performance of each multidimensional feature under a specific failure mode, and the result is expressed in the form of a probability distribution or numerical range.

[0123] By integrating well-represented knowledge into a knowledge base, a physical mechanism knowledge base is established, which facilitates subsequent querying and use.

[0124] Specifically, the steps for a multi-class machine learning model to determine fault modes and output confidence scores are as follows:

[0125] A logistic regression multi-class classification model is identified, combined with a physical mechanism knowledge base;

[0126] Historical data is divided into multi-dimensional feature vectors and fault mode types to train the multi-classification model, resulting in a well-trained multi-classification model.

[0127] Input a new multi-dimensional feature vector into the trained multi-class classification model, and output the predicted fault mode and the corresponding confidence score. The confidence score is calculated using the following formula: ,and ,as well as In the formula, This represents the confidence level that the multidimensional feature vector predicted by the multi-classification model belongs to the fault mode type. This is represented as a multi-class classification model that calculates the category of a class based on a new multi-dimensional feature vector as input. The original scores for each failure mode category, This is represented as a multi-class classification model that calculates the category of a class based on a new multi-dimensional feature vector as input. The original scores for each failure mode category, This is represented as the sum of the index scores for all failure mode categories. This is represented as the maximum confidence level. This indicates that the operator with the highest probability is selected.

[0128] Specifically, the steps for generating an enhanced diagnostic report are as follows:

[0129] By analyzing the arrival time difference or energy attenuation of sensor signals, and combining this with a digital twin model of the spiral duct, the location of the faulty duct segment can be determined. The expression for the location of the faulty duct segment is as follows: In the formula, This represents the distance from the fault point to a sensor. This represents the speed at which a signal propagates in the pipe. It is expressed as the time difference between the arrival of the signal detected by the two sensors. This is expressed as the distance between the two sensors;

[0130] Based on the physical mechanism knowledge base and the output of the multi-class machine learning model, analyze the abnormal causes of the fault types;

[0131] The confidence scores from the multi-class machine learning model output and the fault location prediction are combined to obtain the overall confidence score. The formula for calculating the overall confidence score is as follows: ,and In the formula, This is expressed as the overall confidence level. This represents the weighting coefficient of the confidence level of the corresponding multi-class classification model in determining the fault mode type. This represents the weighting coefficient of the confidence level for locating the corresponding fault location. This represents the confidence level for fault location. This represents the confidence level of the multi-classification model in determining the fault mode type.

[0132] The fault location, cause of the anomaly, and overall confidence level are combined to create an enhanced diagnostic report output.

[0133] S4. Utilizing the enhanced diagnostic report, the remaining life is predicted differentially using a physics-based fatigue cumulative damage model and a data-driven model. The remaining life prediction and decision optimization module has a built-in cost model that converts the predicted remaining life into risk costs, including expected energy consumption loss, downtime loss, and emergency repair costs. These costs are compared with planned maintenance costs, and a decision tree model is used to recommend the economically optimal maintenance timing and strategy, generating an executable strategy.

[0134] Specifically, the steps for predicting remaining life using a physics-based fatigue cumulative damage model are as follows:

[0135] The low-frequency vibration acceleration sensor is converted into a stress time history at the joint through a mechanical model. The load spectrum of different stress amplitudes and stress cycle numbers is obtained by statistically analyzing the stress time history using the rainflow counting method, thereby determining the stress spectrum that the spiral pipe experiences during operation.

[0136] Based on the stress spectrum, the accumulated damage is calculated using the fatigue damage model derived from Miner's linear cumulative damage theory. The expression for the fatigue damage model is as follows: ,and In the formula, This represents the current accumulated damage level, and when... This indicates that the material has experienced fatigue failure. Expressed as the number of stress levels, This represents the number of cycles that have been completed at the specified stress amplitude. This is represented as the value obtained by querying the stress-life SN curve at the 1st... The total number of cycles required for a material to undergo fatigue failure at a given stress amplitude. Expressed as material constants, Expressed as the material stress amplitude, Expressed as a material coefficient;

[0137] Based on the cumulative damage under the current operating conditions of the spiral pipe, the predicted remaining service life is given by the following expression: In the formula, This represents the fatigue damage model used to predict the remaining service life of a spiral pipe. This represents the cumulative degree of damage over time. Indicates future runtime. This represents the future rate of damage accumulation. When future operating conditions are the same as in the recent period, then... .

[0138] Specifically, the steps for predicting remaining lifetime using a data-driven model are as follows:

[0139] Historical data related to the operating status of the spiral pipeline were collected, including multimodal acoustic and vibration raw signal datasets, structured event alarms, and enhanced diagnostic reports. Multiple feature data characterizing the overall health status of the pipeline were extracted, and a health index was constructed. The expression for the health index is as follows: ,and ,as well as In the formula, Indicated as in Health index value over time Represented as the first extracted from historical data A characteristic value related to aging, Represented as the first extracted from historical data A characteristic value related to aging, This is represented as extracting a set of aging-related features from historical data. Indicated as in The first time A characteristic value related to aging, Represented as the first Weighting coefficients for aging-related eigenvalues;

[0140] The historical health index (HI) values ​​are arranged chronologically to form a degradation time series, where the expression for the degradation time series is: In the formula, Indicated as in Time's health index This is expressed as a health index at the present time;

[0141] A time-series prediction LSTM model is trained using historical data to learn the degradation pattern of the health index HI. The expression for the time-series prediction LSTM model is: ,and In the formula, Indicated as current The hidden state and cellular state at any given moment. This represents the calculation of the LSTM model for time series prediction. Indicated as will The health index values ​​over time are input into the LSTM model. Indicated as the previous one The hidden state and cellular state at any given moment. Represented as the next prediction of the LSTM model Constant health index This represents the weight coefficients of the output layer of the LSTM model. This is represented as the bias term of the output layer of the LSTM model;

[0142] The new input feature vector is transformed into the current degraded time series, input into the trained LSTM model, and the output is the predicted remaining lifetime, where the expression for the predicted remaining lifetime is: ,and In the formula, This represents the remaining service life of the spiral pipe as predicted by the LSTM model. This represents the prediction time step. Represented as a function that takes the minimum value, This represents the number of steps to predict backwards from the current moment. This is represented as a predicted health index at the current moment. This represents the preset health index failure threshold.

[0143] To further clarify, the remaining life prediction of fretting wear / fatigue cracks in the interlocking seams of spiral ducts adopts a physics-based fatigue cumulative damage model, using historical operating data as input; the remaining life prediction of sealant aging in spiral ducts adopts a data-driven model, using stiffness characteristic frequency offset and other parameters as health indices.

[0144] Specifically, the steps for converting risk costs are as follows:

[0145] Determine the cost parameters for expected energy loss, downtime loss, emergency repair costs, and planned maintenance costs;

[0146] Based on the cost model, the predicted remaining life is transformed into risk cost, where the expression for risk cost is: ,and , , In the formula, Expressed as risk cost, This is expressed as expected energy loss. Expressed as energy loss per unit time, This is expressed as downtime loss. Expressed as the loss per unit of downtime. This represents the downtime. This is expressed as emergency repair cost. This is represented as the fixed cost of emergency repairs. Represented as a function that takes the minimum value;

[0147] Compare and analyze the risk costs with the planned maintenance costs to implement planned periodic maintenance.

[0148] Specifically, the steps of the decision tree model in recommending the economically optimal maintenance timing and strategy are as follows:

[0149] A training dataset is constructed using risk cost, planned maintenance cost, and predicted remaining life as features to determine the timing decision options for maintaining spiral pipes.

[0150] The decision tree algorithm is used to train the training dataset to build a decision tree model, where each decision option is a branch of the decision tree. For each possible decision option, the total expected cost is calculated using the risk-cost model.

[0151] Compare the total cost of all decision branches and select the branch with the lowest cost as the optimal maintenance timing decision.

[0152] Example 2

[0153] This invention provides, for example Figure 2 The predictive operation and maintenance system for spiral duct based on digital twins shown includes a heterogeneous sensor network module for spiral duct deployment: acoustic emission sensors, vibration sensors, data acquisition devices and operating condition interfaces are deployed in the spiral duct to form a heterogeneous sensor network;

[0154] Data acquisition and preprocessing module: Real-time acquisition of sensor signals using industrial IoT protocols, using fan start / stop signals and operating power as operating condition labels, collaboratively triggering high-frequency acquisition mode and performing adaptive noise reduction processing, outputting a preliminary noise-reduced multimodal acoustic and vibration raw signal dataset;

[0155] Feature extraction and anomaly detection module: Extracts features from acoustic emission signals and vibration signals in the multimodal acoustic and vibration raw signal dataset, inputs multidimensional features into the multivariate statistical process control (MVSPC) model and unsupervised learning algorithm, and outputs structured event alarms;

[0156] Fault diagnosis and location module: Based on structured event alarms, combined with a physical mechanism knowledge base and a multi-class machine learning model, it determines the fault mode and confidence level, realizes fault segment-level location by analyzing sensor signals, and outputs an enhanced diagnostic report;

[0157] The Remaining Life Prediction and Decision Optimization Module utilizes enhanced diagnostic reports to predict remaining life using both a physics-based fatigue cumulative damage model and a data-driven model. It calculates risk costs, compares them with planned maintenance costs, recommends the most economically optimal maintenance timing and strategy, and generates executable strategies.

[0158] The predictive operation and maintenance system for spiral ducts based on digital twins provided in this embodiment of the invention is implemented through the above-described predictive operation and maintenance method for spiral ducts based on digital twins. For details on the specific methods and processes of the predictive operation and maintenance system for spiral ducts based on digital twins, please refer to the embodiments of the above-described predictive operation and maintenance method for spiral ducts based on digital twins, which will not be repeated here.

[0159] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0160] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more sets of available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium. A semiconductor medium can be a solid-state drive.

[0161] It should be understood that in the various embodiments of this application, the order of the above-mentioned processes does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0162] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0163] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A predictive operation and maintenance method for spiral ducts based on digital twins, characterized in that, Includes the following steps: S1. Deploy high-frequency acoustic emission (AE) sensors and low-frequency vibration acceleration sensors at key nodes of the spiral duct. Use the fan start / stop signal and operating power as operating condition labels to automatically trigger the high-frequency acquisition mode and perform noise reduction processing. Output a multi-modal acoustic and vibration raw signal dataset with time sequence, operating condition labels and preliminary noise reduction, including acoustic emission signals and vibration signals. S2. Extract features from acoustic emission signals and vibration signals to form multidimensional feature vectors, and input them into the multivariate statistical process control (MVSPC) model. Monitor the status of the spiral duct by calculating the T² statistic and squared prediction error. Use an unsupervised learning algorithm to learn the fused multidimensional feature space, identify abnormal points that deviate from the normal baseline state, and output structured event alarms. S3. Based on structured event alarms, establish a physical mechanism knowledge base of fault characteristics, combine a multi-class machine learning model trained with historical data to determine the fault mode and output confidence level, and achieve fault segment-level localization by analyzing the time difference / energy attenuation of sensor signals and output an enhanced diagnostic report. S4. Utilizing the enhanced diagnostic report, the remaining life is predicted differentially using a physics-based fatigue cumulative damage model and a data-driven model. The remaining life prediction and decision optimization module has a built-in cost model that converts the predicted remaining life into risk costs. These costs are compared with planned maintenance costs, and a decision tree model is used to recommend the economically optimal maintenance timing and strategy, generating an executable strategy.

2. The predictive operation and maintenance method for spiral ducts based on digital twins according to claim 1, characterized in that, The high-frequency acoustic emission (AE) sensor has a frequency range of 20kHz-1MHz and is used to capture transient elastic wave acoustic emission signals released by structural changes such as fretting wear, microcrack initiation and propagation at the spiral interlocking seam inside the spiral duct material. The low-frequency vibration acceleration sensor has a frequency range of 0.1Hz-2kHz and is used to monitor structural vibration signals caused by decreased connection stiffness due to sealant aging, airflow impact, and fan vibration.

3. The predictive operation and maintenance method for spiral ducts based on digital twins according to claim 1, characterized in that, The steps for outputting the multimodal acoustic vibration raw signal dataset are as follows: High-frequency acoustic emission (AE) sensors and low-frequency vibration acceleration sensors are deployed at key nodes of the spiral duct. Set operating condition labels, including fan start-up, shutdown signals, and operating power, and monitor fan start-up and shutdown signals and operating power in real time; When a change in the start / stop switch signal of the wind turbine is detected or the rate of change of operating power exceeds a preset threshold, the high-frequency acquisition mode is automatically triggered. When the high-frequency acquisition mode is triggered, all high-frequency acoustic emission (AE) sensors deployed on the duct are activated to acquire acoustic emission signals at a frequency of 1MHz. Simultaneously, all low-frequency vibration acceleration sensors deployed on the duct are activated to acquire vibration signals at a frequency of 2kHz. A fixed duration is sampled as the total length of the sampled signal. The collected acoustic emission and vibration signals are processed by wavelet transform noise reduction using a low-pass filter to generate noise-reduced acoustic emission and vibration signals. Add timing information and operating condition labels to the noise-reduced acoustic emission and vibration signals; The acoustic emission signal and vibration signal, which have been labeled with time sequence and operating condition and have undergone preliminary noise reduction, are combined into a multimodal acoustic and vibration raw signal dataset for output.

4. The predictive operation and maintenance method for spiral ducts based on digital twins according to claim 1, characterized in that, The steps for extracting the multidimensional feature vector are as follows: The continuous acoustic emission signal and vibration signal are divided into analysis frames of fixed length, and each frame contains 1024 sampling points; For each frame of data, multiple feature indicators are calculated in parallel: for acoustic emission signals, ring count, energy, amplitude, and wavelet packet energy spectrum are extracted; for vibration signals, root mean square value, kurtosis, and envelope spectrum are extracted. The feature values ​​of all calculated feature indicators are combined in order into a multidimensional vector to form a multidimensional feature vector.

5. The predictive operation and maintenance method for spiral ducts based on digital twins according to claim 1, characterized in that, The steps for monitoring the status of the spiral duct using the Multivariate Statistical Process Control (MVSPC) model are as follows: Under normal operating conditions of the spiral duct, a large amount of multidimensional feature vector data is collected as samples, and the sample mean vector and covariance matrix are calculated. For the multidimensional feature vector input at the new sampling point, calculate the T² statistic and the squared prediction error SPE; Determine the control limits based on the distribution of the T² statistic; The T² statistic calculated in real time is compared with the determined control limits to determine whether the spiral duct is in a normal state.

6. The predictive operation and maintenance method for spiral ducts based on digital twins according to claim 1, characterized in that, The steps of the unsupervised learning algorithm for identifying outliers are as follows: Subsamples are randomly selected from the multidimensional feature vector data as the training set, and the Isolation Forest algorithm in the unsupervised learning algorithm is used to train the training set. Randomly select features and segmentation values ​​to construct multiple isolation trees to build an isolation tree forest; For each data point, traverse each isolation tree and record the path length from the root node to the leaf node. Calculate the anomaly score based on the average path length across all trees; The anomaly score is used to assess the anomaly status of data points and determine whether each data point is an anomaly.

7. The predictive operation and maintenance method for spiral ducts based on digital twins according to claim 1, characterized in that, The output steps for the structured event alarm are as follows: The monitoring results of the MVSPC model and the abnormal scores of the unsupervised learning algorithm are fused to form abnormal evidence. Determine the anomaly type based on the characteristics of the anomaly and its location; By combining the location information from the sensors, the location of the anomaly can be determined; Record the time when the exception occurred; The severity of the deviation is scored by calculating the Euclidean distance based on the degree of deviation of outliers from the normal baseline. The anomaly type, location, time, and severity score are combined into a structured event alert output.

8. The predictive operation and maintenance method for spiral ducts based on digital twins according to claim 1, characterized in that, The steps for establishing the physical mechanism knowledge base are as follows: List all types of faults that need to be diagnosed, including fretting wear of the interlocking joint, aging and leakage of sealant, external impact, and normal condition; Collect various data related to spiral duct failures, including structured event alerts, spiral duct physics principles, experimental data, and expert experience; Extract and map the physical mechanism knowledge related to fault characteristics from the collected data to determine the typical signal characteristics corresponding to each fault mode; Mathematical modeling is performed on the typical performance of each multidimensional feature under a specific failure mode, and the result is expressed in the form of a probability distribution or numerical range. By integrating well-represented knowledge into a knowledge base, a physical mechanism knowledge base is established, which facilitates subsequent querying and use. The steps of the multi-class machine learning model in determining the fault mode and outputting the confidence score are as follows: A logistic regression multi-class classification model is identified, combined with a physical mechanism knowledge base; Historical data is divided into multi-dimensional feature vectors and fault mode types to train the multi-classification model, resulting in a well-trained multi-classification model. Input a new multidimensional feature vector into the trained multi-classification model and output the predicted fault mode and the corresponding confidence level.

9. The predictive operation and maintenance method for spiral ducts based on digital twins according to claim 1, characterized in that, The steps for predicting remaining life using the physics-based fatigue cumulative damage model are as follows: The low-frequency vibration acceleration sensor is converted into a stress time history at the joint through a mechanical model. The load spectrum of different stress amplitudes and stress cycle numbers is obtained by statistically analyzing the stress time history using the rainflow counting method, thereby determining the stress spectrum that the spiral pipe experiences during operation. Based on the stress spectrum, the fatigue damage model of Miner's linear cumulative damage theory is used to calculate the accumulated damage that has occurred. The remaining service life is predicted based on the cumulative damage under the current operating conditions of the spiral pipe. The steps for predicting remaining lifetime using the data-driven model are as follows: Collect historical data related to the operating status of the spiral pipeline, including multimodal acoustic and vibration raw signal datasets, structured event alarms and enhanced diagnostic reports, and extract multiple feature data that can characterize the overall health status of the pipeline to construct a health index; The historical health index HI values ​​are arranged in chronological order to form a degradation time series. A time series prediction LSTM model is trained using the historical data to learn the degradation pattern of the health index HI. The new input feature vector is transformed into the current degradation time series, input into the trained LSTM model, and the predicted remaining lifetime is output.

10. A predictive operation and maintenance system for spiral ducts based on digital twins, implemented by the predictive operation and maintenance method for spiral ducts based on digital twins as described in any one of claims 1-9, characterized in that, This includes deploying heterogeneous sensor network modules in spiral ducts: deploying acoustic emission sensors, vibration sensors, data acquisition devices, and operating condition interfaces in spiral ducts to form a heterogeneous sensor network; Data acquisition and preprocessing module: Real-time acquisition of sensor signals using industrial IoT protocols, using fan start / stop signals and operating power as operating condition labels, collaboratively triggering high-frequency acquisition mode and performing adaptive noise reduction processing, outputting a preliminary noise-reduced multimodal acoustic and vibration raw signal dataset; Feature extraction and anomaly detection module: Extracts features from acoustic emission signals and vibration signals in the multimodal acoustic and vibration raw signal dataset, inputs multidimensional features into the multivariate statistical process control (MVSPC) model and unsupervised learning algorithm, and outputs structured event alarms; Fault diagnosis and location module: Based on structured event alarms, combined with a physical mechanism knowledge base and a multi-class machine learning model, it determines the fault mode and confidence level, realizes fault segment-level location by analyzing sensor signals, and outputs an enhanced diagnostic report; The Remaining Life Prediction and Decision Optimization Module utilizes enhanced diagnostic reports to predict remaining life using both a physics-based fatigue cumulative damage model and a data-driven model. It calculates risk costs, compares them with planned maintenance costs, recommends the most economically optimal maintenance timing and strategy, and generates executable strategies.

Images