Turbine rotor idle running abnormity real-time identification method and device based on machine learning

Abstract

The invention discloses a machine learning-based real-time identification method and a machine learning-based real-time identification device for coasting running abnormity of a steam turbine rotor, and relates to the technical field of steam turbine state monitoring and fault diagnosis. Multi-dimensional parameters such as rotor rotating speed, bearing temperature, vibration amplitude, lubricating oil pressure and exhaust steam temperature are collected in real time through a multi-sensor system, after wavelet threshold denoising, K-nearest neighbor interpolation and Z-score standardization preprocessing, an identification model constructed based on a random forest algorithm is input, an abnormal probability value is output and compared with a preset threshold value, and the abnormal probability value is identified. And abnormal state determination is realized. Once an abnormality is detected, early warning information including an abnormality type, occurrence time and detailed parameters is automatically generated. According to the method, the multi-source data and the machine learning model are fused, the recognition accuracy and real-time performance are improved, the problems that a traditional method depends on an empirical threshold value, the anti-interference capacity is weak, and adaptability is poor are solved, and reliable guarantee is provided for safe operation of the steam turbine.

Description

Technical Field

[0001] This application relates to the field of turbine condition monitoring and fault diagnosis technology, specifically to a method and device for real-time identification of turbine rotor coasting anomalies based on machine learning. Background Technology

[0002] As a core piece of equipment in power generation and industrial power systems, the operational safety of steam turbines directly affects the stability and reliability of the entire system. During turbine shutdown, the rotor coasting phase refers to the process where the unit continues to rotate due to inertia after being disconnected from the grid until it stops. The dynamic characteristics of this process reflect the health level of the equipment's mechanical condition. Currently, the industry commonly uses multiple sensors to collect parameters such as speed, temperature, and vibration, and combines this with signal processing technology to analyze the coasting process. Existing technologies include methods based on empirical threshold judgment and traditional signal processing (such as Fourier transform and wavelet analysis) for anomaly identification. These methods achieve state discrimination by setting fixed thresholds or extracting time-frequency domain features, and have a certain engineering application foundation.

[0003] However, the aforementioned methods have significant drawbacks: relying on manually set thresholds makes it difficult to adapt to changes in different operating conditions and equipment aging, resulting in poor generalization ability; traditional signal processing methods are poorly adaptable to nonlinear and non-stationary data, feature extraction is complex and lacks real-time performance; and the lack of a multi-dimensional parameter fusion mechanism leads to low recognition accuracy and high false alarm rate. Furthermore, existing systems often lack automatic early warning and data traceability functions, failing to meet the demands of modern industry for intelligent, high-precision, and real-time fault diagnosis. Therefore, a novel recognition method that can integrate multi-source information, possesses strong learning capabilities, and exhibits real-time response characteristics is urgently needed to improve the accuracy and practicality of turbine rotor coasting anomaly detection. Summary of the Invention

[0004] The purpose of this invention is to provide a method and device for real-time identification of turbine rotor coasting anomalies based on machine learning, which can solve the technical problems of low accuracy, poor real-time performance and insufficient adaptability in the existing technology for identifying turbine rotor coasting anomalies.

[0005] To achieve the above objectives, this application provides the following technical solution: A machine learning-based real-time identification method for turbine rotor coasting anomalies, characterized by the following steps: S1, the multi-sensor system collects multi-dimensional operating parameters of the turbine rotor coasting process in real time, including rotor speed, bearing temperature, vibration amplitude, lubricating oil pressure and exhaust temperature; S2, preprocessing the collected multi-dimensional operating parameters to obtain the processed data. The preprocessing includes using wavelet threshold denoising algorithm for signal denoising, using K nearest neighbor algorithm for missing value imputation, and using Z-score method for data standardization. S3, a lazy walking anomaly identification model is built based on the random forest algorithm. The model is trained using historical normal samples and abnormal samples, and the model hyperparameters are optimized through cross-validation and grid search. S4. Input the processed data into the trained lazy walking anomaly recognition model and output the anomaly probability value. When the anomaly probability value exceeds the set threshold, it is determined to be a lazy walking anomaly state. S5 automatically generates warning information containing the abnormality type, occurrence time, and detailed parameters when an abnormal state is detected.

[0006] Furthermore, in step S1, during the collection of multi-dimensional operating parameters, the sampling frequency is set to no less than 100Hz, and the data from each sensor are precisely aligned with the timestamps through a synchronous acquisition device to ensure the synchronicity and real-time nature of the data collection.

[0007] Furthermore, in step S2, the wavelet thresholding denoising is performed using a soft thresholding function, the mathematical expression of which is: η(w) = sign(w) × max(|w| - λ, 0); Where w is the wavelet coefficient, λ is the denoising threshold, sign(w) is the sign function, and max(·) is the maximum value function.

[0008] Furthermore, in S2, the K-nearest neighbor missing value imputation uses a weighted average method for samples containing missing values. The j-th missing feature value The interpolation formula is: ; Where k is the number of nearest neighbors. The weight is the j-th feature value of the i-th nearest neighbor sample. The calculation formula is:

[0009] in, For the sample With the i-th nearest neighbor sample The Euclidean distance between them To prevent division by zero, use a very small constant.

[0010] Furthermore, the specific process of S4 is as follows: The processed data is then input into the trained lazy walking anomaly detection model. The lazy walking anomaly identification model outputs the anomaly probability P through a multi-decision tree voting mechanism; The anomaly probability P is compared with a preset threshold T; When the abnormal probability P is greater than the threshold T, i.e. P>T, the current turbine rotor coasting state is determined to be abnormal. When the abnormal probability P is less than or equal to the threshold T, i.e. P ≤ T, the current turbine rotor coasting state is determined to be normal.

[0011] Furthermore, the formula for calculating the anomaly probability P is as follows: ; Where N is the total number of decision trees, This is the output of the i-th decision tree.

[0012] Furthermore, the specific process of S5 is as follows: When an abnormal state is detected, immediately capture and record the timestamp of the abnormal state as the time of the abnormality. Based on the auxiliary information output by the coasting anomaly identification model, the type of the current anomaly is identified and determined; the anomaly type includes, but is not limited to, one or more of the following: bearing wear, rotor imbalance, abnormal lubricating oil pressure, and steam seal friction. Extract the raw and preprocessed data of all multi-dimensional operating parameters within a preset time window before and after the time of the anomaly occurrence to form a detailed parameter set for traceability analysis; The time of the anomaly, the identified anomaly type, and the extracted detailed parameter set are encapsulated according to a predefined data structure and format and assembled into an early warning message.

[0013] The second aspect of this application provides a machine learning-based real-time identification system for turbine rotor coasting anomalies, comprising: Acquisition module: Used to acquire multi-dimensional operating parameters of the turbine rotor during coasting in real time through a multi-sensor system. The operating parameters include rotor speed, bearing temperature, vibration amplitude, lubricating oil pressure, and exhaust steam temperature. Preprocessing module: Used to preprocess the collected multi-dimensional operating parameters to obtain processed data. Preprocessing includes signal denoising using wavelet threshold denoising algorithm, missing value imputation using K-nearest neighbor algorithm, and data standardization using Z-score method; Training module: Used to build a lazy walking anomaly detection model based on the random forest algorithm, train the model using historical normal samples and abnormal samples, and optimize the model hyperparameters through cross-validation and grid search; Judgment module: It is used to input the processed data into the trained lazy walking anomaly detection model and output the anomaly probability value. When the anomaly probability value exceeds the set threshold, it is judged as a lazy walking anomaly state. Warning information generation module: When an abnormal state is detected, it automatically generates warning information containing the abnormality type, occurrence time and detailed parameters.

[0014] A third aspect of this application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the machine learning-based real-time identification method for turbine rotor coasting anomalies described in the first aspect.

[0015] The fourth aspect of this application provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the machine learning-based real-time identification method for turbine rotor coasting anomalies described in the first aspect.

[0016] Compared with the prior art, the present invention has the following beneficial technical effects: This invention provides a machine learning-based real-time identification method for turbine rotor coasting anomalies. It comprehensively characterizes the state changes during coasting by synchronously collecting multi-dimensional parameters such as rotor speed, bearing temperature, vibration amplitude, lubricating oil pressure, and exhaust temperature from multiple sensors. Wavelet threshold denoising effectively suppresses high-frequency noise interference, improving the signal-to-noise ratio. K-nearest neighbor weighted interpolation ensures data integrity, and Z-score standardization eliminates dimensional differences, improving model convergence efficiency. A random forest algorithm is used to construct the identification model, leveraging its ensemble learning mechanism and anti-overfitting ability to accurately capture complex nonlinear patterns. Clear decision-making logic is achieved by comparing the anomaly probability output with a threshold, supporting sensitivity adjustment. Structured early warning information containing anomaly type, time, and detailed parameters is generated for easy maintenance traceability. Therefore, this method solves the problems of low accuracy, poor real-time performance, and weak adaptability of traditional methods, significantly improving the intelligence level and engineering practicality of turbine rotor coasting anomaly identification. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of a real-time identification method for turbine rotor coasting anomalies based on machine learning in an embodiment of the present invention. Detailed Implementation

[0018] In the following description, only certain exemplary embodiments are briefly described. As those skilled in the art will recognize, the described embodiments can be modified in various ways without departing from the spirit or scope of the invention. Therefore, the drawings and description are considered to be exemplary in nature and not restrictive.

[0019] It should be understood that, when used in this specification and the appended claims, the terms "comprising" and "including" indicate the presence of the described features, integrals, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or collections thereof.

[0020] It should also be further understood that the term "and / or" as used in this specification and the appended claims refers to any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0021] See Figure 1 This invention provides a real-time identification method for turbine rotor coasting anomalies based on machine learning, comprising the following steps: S1 uses a multi-sensor system to collect multi-dimensional operating parameters of the turbine rotor during coasting in real time. These operating parameters include rotor speed, bearing temperature, vibration amplitude, lubricating oil pressure, and exhaust temperature. The multi-sensor system consists of various types of sensors installed at key parts of the turbine: rotor speed is acquired using magnetoelectric or photoelectric speed sensors, with sampling points typically located at the end of the main shaft or gear disk; bearing temperature is monitored continuously using platinum resistance (Pt100) or thermocouple sensors arranged in the temperature measurement holes of each supporting bearing; vibration amplitude is acquired using acceleration sensors or eddy current displacement sensors, used to measure bearing housing vibration and shaft radial runout, respectively; lubricating oil pressure is measured via a pressure transmitter installed in the middle of the oil supply pipeline, reflecting the oil supply stability of the lubrication system; and exhaust temperature is measured using an insertion-type thermistor to assess the exhaust energy state and thermal stress distribution in the low-pressure cylinder exhaust port area. All these sensors are connected to a synchronous acquisition device via an industrial-grade data acquisition card, ensuring that all channel timestamps are aligned and avoiding feature distortion due to time delay differences. The acquired data is a time-series signal, covering the entire process from shutdown to complete rotor standstill, with typical durations ranging from tens of minutes to several hours, forming a high-dimensional, dynamic, and non-stationary operating state characterization dataset.

[0022] The selection of the aforementioned multi-dimensional parameters has clear physical significance and fault correlation: a decreasing rotor speed reflects changes in overall inertia and frictional torque; abnormal deceleration may indicate mechanical jamming or abnormal load; increased bearing temperature often indicates poor lubrication or localized overheating; sudden increases in vibration amplitude and changes in spectral composition are closely related to rotor imbalance, misalignment, or rubbing faults; fluctuations in lubricating oil pressure can reveal oil pump performance degradation or pipeline leaks; abnormal exhaust temperature may reflect vacuum system faults or steam backflow. By integrating these five types of parameters, a state perception system covering multiple physical fields such as mechanics, thermodynamics, and fluid dynamics is constructed, providing a sufficient information foundation for subsequent intelligent identification.

[0023] S2, preprocessing the collected multi-dimensional operating parameters to obtain the processed data. The preprocessing includes using wavelet threshold denoising algorithm for signal denoising, using K nearest neighbor algorithm for missing value imputation, and using Z-score method for data standardization. The signal denoising stage addresses issues such as high-frequency electromagnetic interference, environmental noise, and sensor drift in the original data by employing wavelet transform combined with a soft thresholding function for filtering. Specifically, the original signal is first decomposed into wavelet domains of multiple scales, and multi-level decomposition is performed using a Daubechies wavelet basis (such as db4) to extract coefficients for different frequency sub-bands. Then, soft thresholding is applied to the detail coefficients at each level to suppress weak signal components below a set threshold while preserving the main characteristic components. Finally, the smoothed signal is reconstructed through inverse wavelet transform. Compared to traditional low-pass filters, this method more effectively removes random noise while preserving signal edge characteristics, thus improving the signal-to-noise ratio.

[0024] For data loss due to communication interruptions, sensor malfunctions, or other reasons, the K-Nearest Neighbors (KNN) algorithm is used for imputation. The basic idea is to find the k most similar neighbor samples in the historical dataset to the current sample in the feature space, and then use their weighted average to fill in the missing data. Euclidean distance is used as the distance metric, with the weight decreasing as the distance increases, thus enhancing the influence of samples with strong local correlation. This method fully utilizes the inherent correlation between data points and maintains consistency of nonlinear trends better than linear interpolation.

[0025] Z-score standardization is used to eliminate differences in dimensions and orders of magnitude between different parameters, ensuring that all features are on the same comparison benchmark. The calculation formula is as follows: ,in and These represent the mean and standard deviation of the corresponding feature across the entire training set. Standardized data have a mean of 0 and a variance of 1, which is beneficial for subsequent gradient convergence and classification performance optimization in machine learning models.

[0026] The above three preprocessing operations are performed sequentially, forming a complete data purification process: first, noise reduction to ensure signal quality; then, missing data to ensure data integrity; and finally, standardization to unify the scale, outputting high-quality feature vectors that can be used for modeling.

[0027] S3, a lazy walking anomaly identification model is built based on the random forest algorithm. The model is trained using historical normal samples and abnormal samples, and the model hyperparameters are optimized through cross-validation and grid search. Random Forest (RF) is an ensemble learning method consisting of a large number of decision trees. Each tree uses bootstrap sampling to extract subsamples from the training set and randomly selects some features to participate in the judgment when splitting nodes, thereby reducing the risk of overfitting and improving generalization ability. In this application, the input features are preprocessed five-dimensional parameter time window sliding statistics (such as mean, variance, slope, peak factor, etc.), and the labels are "normal" or "abnormal" status markers, derived from historical operation and maintenance records and expert annotations.

[0028] Before model training, historical data is divided into training and test sets in chronological order, with a ratio of 7:3 or 8:2, ensuring the test set is not involved in any parameter adjustment process. Five-fold cross-validation is used to evaluate model stability, with four training sets and one validation set used in each iteration. The final average performance metrics (such as accuracy, F1 score, and AUC) are used as the evaluation criterion. Simultaneously, a grid search is used to traverse key hyperparameter combinations, including the number of decision trees (n_estimators), maximum depth (max_depth), minimum number of sample splits (min_samples_split), and maximum number of feature selections (max_features), to find the optimal configuration.

[0029] The advantages of random forests lie in their strong robustness to noise and outliers, their ability to capture complex nonlinear relationships without precise feature engineering, and their ability to output feature importance rankings to aid in the diagnosis of key influencing factors. Furthermore, their parallel structure is suitable for deployment on edge computing devices, meeting real-time inference requirements.

[0030] S4. Input the processed data into the trained lazy walking anomaly recognition model and output the anomaly probability value. When the anomaly probability value exceeds the set threshold, it is determined to be a lazy walking anomaly state. In the real-time operation phase, the system slides through the latest collected and preprocessed multidimensional parameter sequence at fixed time windows (e.g., 10 seconds), extracts a set of feature vectors, and inputs them into the trained random forest model. Internally, the model uses independent voting by all decision trees to statistically determine the proportion of predictions as "abnormal," which is then used as the anomalous probability P. For example, if the forest contains 100 decision trees, and 78 of them are judged as anomalous, then the output P = 0.78.

[0031] S5 automatically generates warning information containing the abnormality type, occurrence time, and detailed parameters when an abnormal state is detected.

[0032] Once the system determines that it has entered an abnormal state, it immediately activates the early warning generation mechanism. First, it records the current UTC timestamp as the time of the abnormality, with an accuracy of milliseconds. Second, it infers the type of abnormality by combining the classification results output by the model or by adding a classifier. Common types include, but are not limited to: bearing wear (manifested as a continuous increase in temperature and enhanced vibration subharmonics), rotor imbalance (significantly increased first-harmonic vibration), abnormal lubricating oil pressure (sudden drop in oil supply pressure or increased fluctuations), and steam seal friction (intermittent vibration spikes superimposed with localized temperature increases). Third, it automatically extracts the original sensor data and preprocessed feature data within a preset time window (e.g., ±30 seconds) before and after the abnormality, forming a complete parameter snapshot for subsequent traceability analysis. Finally, it encapsulates the above content according to a unified data structure (e.g., JSON format), generates a standardized early warning message, and pushes it to the monitoring platform via protocols such as OPC UA and MQTT, while simultaneously triggering audible and visual alarms, SMS notifications, or email alerts.

[0033] Through the above-described steps, this application achieves fully automated anomaly identification during the coasting process of a steam turbine rotor: comprehensive status acquisition is achieved through multi-sensor collaborative sensing; data quality is improved by wavelet denoising, KNN interpolation, and Z-score standardization; high-precision pattern recognition is achieved using a random forest model; flexible anomaly determination is achieved by combining a probability threshold mechanism; and finally, structured early warning output supports rapid response and fault tracing. This method overcomes the limitations of traditional methods that rely on single parameters and human experience, significantly improving identification accuracy, real-time performance, and intelligence. It can detect potential hazards in the early stages of a fault, prevent the escalation of accidents, and ensure the safe and stable operation of the unit. It is applicable to various application scenarios such as thermal power, nuclear power, and marine power.

[0034] The present invention also provides a more specific embodiment in which the sampling frequency is set to no less than 100Hz, and the data from each sensor are precisely aligned with the timestamps through a synchronous acquisition device to ensure the synchronicity and real-time performance of the data acquisition. The sampling frequency of no less than 100Hz means that at least 100 data samples are taken per second when collecting multi-dimensional operating parameters during the turbine rotor coasting process. This high sampling frequency can effectively capture the rapid dynamic changes that occur in the rotor during coasting, especially transient characteristics such as vibration amplitude and speed fluctuations, avoiding the omission or masking of key abnormal signals due to excessively large sampling intervals.

[0035] The present invention also provides a more specific embodiment, in which wavelet thresholding denoising is performed using a soft thresholding function, the mathematical expression of which is: η(w) = sign(w) × max(|w| - λ, 0); Where w is the wavelet coefficient, λ is the denoising threshold, sign(w) is the sign function, and max(·) is the maximum value function.

[0036] Wavelet thresholding denoising is a signal processing technique based on wavelet transform, widely used for noise suppression of non-stationary signals. Its core idea is to decompose the original time-domain signal into different scale spaces through wavelet transform, concentrating the energy of the useful signal on a few large coefficients, while the noise energy is dispersed among multiple small coefficients. By applying a threshold operation to the wavelet coefficients, low-amplitude noise components can be effectively filtered out. The signal is then reconstructed through inverse wavelet transform, achieving the denoising goal. In this embodiment, the soft thresholding function η(w) continuously shrinks the wavelet coefficients. Compared to the hard thresholding method (directly setting coefficients less than the threshold to zero), this avoids signal reconstruction oscillations and the Gibbs effect caused by abrupt coefficient changes, thereby improving the smoothness and fidelity of the denoised signal.

[0037] The present invention also provides a more specific embodiment, wherein in S2, the K-nearest neighbor missing value imputation adopts a weighted average method for samples containing missing values. The j-th missing feature value The interpolation formula is: ; Where k is the number of nearest neighbors. The weight is the j-th feature value of the i-th nearest neighbor sample. The calculation formula is:

[0038] in, For the sample With the i-th nearest neighbor sample The Euclidean distance between them To prevent division by zero, use a very small constant.

[0039] K-nearest neighbor (KNN) imputation is a data completion technique based on the similarity assumption. It is used to reasonably estimate missing feature values ​​when some data points are lost during multi-dimensional operational parameter acquisition due to sensor malfunctions, signal interference, or communication delays. The weighted average method, instead of using the arithmetic mean of the feature values ​​of each nearest neighbor sample when calculating the imputed value, assigns different weights based on their similarity to the target sample, achieving a more refined estimate.

[0040] The present invention also provides a more specific embodiment, wherein the specific process of S4 is as follows: The processed data is then input into the trained lazy walking anomaly detection model. The lazy walking anomaly identification model outputs the anomaly probability P through a multi-decision tree voting mechanism; The anomaly probability P is compared with a preset threshold T; When the abnormal probability P is greater than the threshold T, i.e. P>T, the current turbine rotor coasting state is determined to be abnormal. When the anomaly probability P is less than or equal to the threshold T, i.e., P ≤ T, the current turbine rotor coasting state is determined to be normal. The coasting anomaly identification model is built based on the Random Forest algorithm. Its core structure consists of multiple independently trained decision trees. Each tree splits and grows based on different sample subsets and feature subsets, thus forming diverse classification paths. During the inference phase, each decision tree independently judges the input data and outputs a local prediction result indicating that the sample belongs to the "anomaly" category.

[0041] The present invention also provides a more specific embodiment, wherein the formula for calculating the anomaly probability P is: ; Where N is the total number of decision trees, Let be the output of the i-th decision tree. The calculation formula is introduced to quantify and integrate the classification results of multiple decision trees in a random forest model to obtain a statistically significant anomaly probability value. This formula is based on the voting mechanism principle in ensemble learning. It converts the binary classification output (normal or anomalous) of each decision tree into a numerical representation of 0 or 1, and then averages them to obtain the final probability estimate.

[0042] The present invention also provides a more specific embodiment, wherein the specific process of S5 is as follows: when an abnormal state is detected, the timestamp of the abnormal state is immediately captured and recorded as the time of the abnormality occurrence; based on the auxiliary information output by the coasting anomaly identification model, the type of the current anomaly is identified and determined; the anomaly type includes, but is not limited to, one or more of bearing wear, rotor imbalance, abnormal lubricating oil pressure, and steam seal friction; the raw data and preprocessed data of all multi-dimensional operating parameters within a preset time window before and after the time of the anomaly occurrence are extracted to form a detailed parameter set for traceability analysis; the time of the anomaly occurrence, the identified anomaly type, and the extracted detailed parameter set are encapsulated according to a predefined data structure and format and assembled into early warning information.

[0043] The auxiliary information output by the coasting anomaly identification model refers to information generated internally by the model, besides the anomaly probability, that can be used for classification and interpretation. Examples include feature importance ranking, leaf node path distribution, category confidence vectors, or attention weights (if using ensemble interpretation methods such as SHAP values). By analyzing this auxiliary information, the system can infer the main factors causing the anomaly. For example, when vibration amplitude-related features frequently appear at split nodes in the decision tree and have high weights, the system tends to classify it as "rotor imbalance"; when bearing temperature and lubricating oil pressure jointly exhibit a significant abnormal pattern, it may be attributed to "bearing wear" or "abnormal lubricating oil pressure"; if exhaust temperature and vibration signals show periodic coupled fluctuations, it suggests the possible existence of "steam seal friction." Anomaly type identification can be achieved through rule engine matching, cluster analysis, or a two-level classifier.

[0044] In another embodiment of the present invention, a real-time identification system for turbine rotor coasting anomalies based on machine learning is also provided, comprising: The data acquisition module is used to collect multi-dimensional operating parameters of the turbine rotor during coasting in real time through a multi-sensor system. These operating parameters include rotor speed, bearing temperature, vibration amplitude, lubricating oil pressure, and exhaust steam temperature. Preprocessing module: Used to preprocess the collected multi-dimensional operating parameters to obtain processed data. Preprocessing includes signal denoising using wavelet threshold denoising algorithm, missing value imputation using K-nearest neighbor algorithm, and data standardization using Z-score method; Training module: Used to build a lazy walking anomaly detection model based on the random forest algorithm, train the model using historical normal samples and abnormal samples, and optimize the model hyperparameters through cross-validation and grid search; Judgment module: It is used to input the processed data into the trained lazy walking anomaly detection model and output the anomaly probability value. When the anomaly probability value exceeds the set threshold, it is judged as a lazy walking anomaly state. Warning information generation module: When an abnormal state is detected, it automatically generates warning information containing the abnormality type, occurrence time and detailed parameters.

[0045] The present invention also discloses a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the above-described machine learning-based real-time identification method for turbine rotor coasting anomalies.

[0046] The present invention also discloses a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the above-mentioned machine learning-based real-time identification method for turbine rotor coasting anomalies.

[0047] This invention is described based on flowcharts and / or block diagrams of methods, apparatus (systems), and computer program products according to specific embodiments. It should be understood that each block of the flowcharts and / or block diagrams, and combinations of blocks in the flowcharts and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing device to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing device, generate instructions for implementing the flowcharts and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0048] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0049] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0050] It should be understood that the above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit them. Those skilled in the art can modify the technical solutions described in the above embodiments, or make equivalent substitutions for some of the technical features; and all such modifications and substitutions should fall within the protection scope of the present invention.

[0051] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A real-time identification method for turbine rotor coasting anomalies based on machine learning, characterized in that, Includes the following steps: S1, the multi-sensor system collects multi-dimensional operating parameters of the turbine rotor coasting process in real time, including rotor speed, bearing temperature, vibration amplitude, lubricating oil pressure and exhaust temperature; S2, preprocessing the collected multi-dimensional operating parameters to obtain the processed data. The preprocessing includes using wavelet threshold denoising algorithm for signal denoising, using K nearest neighbor algorithm for missing value imputation, and using Z-score method for data standardization. S3, a lazy walking anomaly identification model is built based on the random forest algorithm. The model is trained using historical normal samples and abnormal samples, and the model hyperparameters are optimized through cross-validation and grid search. S4. Input the processed data into the trained lazy walking anomaly recognition model and output the anomaly probability value. When the anomaly probability value exceeds the set threshold, it is determined to be a lazy walking anomaly state. S5 automatically generates warning information containing the abnormality type, occurrence time, and detailed parameters when an abnormal state is detected.

2. The method for real-time identification of turbine rotor coasting anomalies based on machine learning according to claim 1, characterized in that, In step S1, during the collection of multi-dimensional operating parameters, the sampling frequency is set to no less than 100Hz, and the data from each sensor are precisely aligned with the timestamps through a synchronous acquisition device to ensure the synchronicity and real-time performance of the data collection.

3. The method for real-time identification of turbine rotor coasting anomalies based on machine learning according to claim 1, characterized in that, In step S2, the wavelet thresholding denoising is performed using a soft thresholding function, the mathematical expression of which is: η(w) = sign(w) × max(|w| - λ, 0); Where w is the wavelet coefficient, λ is the denoising threshold, sign(w) is the sign function, and max(·) is the maximum value function.

4. The method for real-time identification of turbine rotor coasting anomalies based on machine learning according to claim 1, characterized in that, In S2, the K-nearest neighbor missing value imputation uses a weighted average method for samples containing missing values. The j-th missing feature value The interpolation formula is: ； Where k is the number of nearest neighbors. The weight is the j-th feature value of the i-th nearest neighbor sample. The calculation formula is: in, For the sample With the i-th nearest neighbor sample The Euclidean distance between them To prevent division by zero for extremely small constants.

5. The method for real-time identification of turbine rotor coasting anomalies based on machine learning according to claim 1, characterized in that, The specific process of S4 is as follows: The processed data is then input into the trained lazy walking anomaly detection model. The lazy walking anomaly identification model outputs the anomaly probability P through a multi-decision tree voting mechanism; The anomaly probability P is compared with a preset threshold T; When the abnormal probability P is greater than the threshold T, i.e. P > T, the current turbine rotor coasting state is determined to be abnormal. When the abnormal probability P is less than or equal to the threshold T, i.e. P ≤ T, the current turbine rotor coasting state is determined to be normal.

6. The method for real-time identification of turbine rotor coasting anomalies based on machine learning according to claim 5, characterized in that, The formula for calculating the anomaly probability P is: ； Where N is the total number of decision trees, This is the output of the i-th decision tree.

7. The method for real-time identification of turbine rotor coasting anomalies based on machine learning according to claim 1, characterized in that, The specific process of S5 is as follows: When an abnormal state is detected, immediately capture and record the timestamp of the abnormal state as the time of the abnormality. Based on the auxiliary information output by the coasting anomaly identification model, the type of the current anomaly is identified and determined; the anomaly type includes, but is not limited to, one or more of the following: bearing wear, rotor imbalance, abnormal lubricating oil pressure, and steam seal friction. Extract the raw and preprocessed data of all multi-dimensional operating parameters within a preset time window before and after the time of the anomaly occurrence to form a detailed parameter set for traceability analysis; The time of the anomaly, the identified anomaly type, and the extracted detailed parameter set are encapsulated according to a predefined data structure and format and assembled into an early warning message.

8. A real-time identification system for turbine rotor coasting anomalies based on machine learning, characterized in that, include: Acquisition module: Used to acquire multi-dimensional operating parameters of the turbine rotor during coasting in real time through a multi-sensor system. The operating parameters include rotor speed, bearing temperature, vibration amplitude, lubricating oil pressure, and exhaust steam temperature. Preprocessing module: Used to preprocess the collected multi-dimensional operating parameters to obtain processed data. Preprocessing includes signal denoising using wavelet threshold denoising algorithm, missing value imputation using K-nearest neighbor algorithm, and data standardization using Z-score method; Training module: Used to build a lazy walking anomaly detection model based on the random forest algorithm, train the model using historical normal samples and abnormal samples, and optimize the model hyperparameters through cross-validation and grid search; Judgment module: It is used to input the processed data into the trained lazy walking anomaly detection model and output the anomaly probability value. When the anomaly probability value exceeds the set threshold, it is judged as a lazy walking anomaly state. Warning information generation module: When an abnormal state is detected, it automatically generates warning information containing the abnormality type, occurrence time and detailed parameters.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the real-time identification method for turbine rotor coasting anomalies based on machine learning as described in any one of claims 1-7.

10. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the real-time identification method for turbine rotor coasting anomalies based on machine learning as described in any one of claims 1-7.

Images