Federal learning degradation prediction method for lubricating system of turboset
By constructing a communication protocol and deep neural network model for the turbine lubrication system using a federated learning method, the problems of data silos and privacy protection were solved. This enabled collaborative training and accurate prediction of multi-source data, improving the accuracy of degradation prediction and equipment status monitoring capabilities of the lubrication system.
Patent Information
- Application Number
- CN202511502414.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-10-21
- Publication Date
- 2026-01-16
AI Technical Summary
Existing technologies for predicting the degradation trend of turbine lubrication systems suffer from data silos, insufficient model generalization ability, and a balance between privacy protection and model accuracy. These issues lead to localized prediction results and high false alarm rates, making it difficult to achieve cross-institutional data sharing and effective prediction.
By employing a federated learning approach, a communication protocol between a central server and local nodes is constructed. Through data collection, preprocessing, feature extraction, and encryption, deep neural networks are used for model training and aggregation to achieve collaborative training and privacy protection of multi-source data, and to construct a multi-scale health index for prediction.
While ensuring data privacy, the accuracy of lubrication system degradation prediction and system generalization ability have been improved, prediction lag and false alarm rate have been reduced, and accurate prediction of equipment status and intelligent suggestions for maintenance strategies have been realized.
Smart Images
Figure CN121350474A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The application belongs to the technical field of steam turbine unit detection, and particularly relates to a steam turbine unit lubricating system federated learning degradation prediction method. BACKGROUND
[0002] The lubricating system of a steam turbine generator is a core subsystem of a key power equipment, and its operation state directly affects the safety and economy of the unit. The traditional degradation trend prediction method mainly relies on the monitoring data of a single unit to construct a prediction model, which has three technical defects: first, the modeling method based on single-point data cannot fully represent the overall degradation characteristics of a multi-component coupled system, resulting in localized prediction results; second, there is a data island phenomenon between different units, which is subject to data privacy and all authority restrictions, making it difficult to integrate cross-institutional data and limiting the model generalization ability; third, the existing federated learning framework has insufficient adaptability in the time series degradation prediction scene and cannot effectively handle the time-frequency characteristics and stage characteristics of the lubricating system degradation data.
[0003] The distributed learning method disclosed in the prior art mainly adopts a parameter averaging strategy, which fails to establish a model aggregation mechanism consistent with the equipment degradation mechanism, resulting in insufficient capture of the degradation inflection point by the federated model. At the same time, the traditional method does not consider the time domain correlation between the oil parameter and the vibration parameter of the lubricating system in the data preprocessing stage, resulting in insufficient feature extraction. In addition, the existing solution lacks an effective balance mechanism between privacy protection and model accuracy, and there is a risk of feature leakage during parameter transmission, affecting the feasibility of actual engineering application.
[0004] The core contradiction in the current field of intelligent operation and maintenance of lubricating systems is that improving prediction accuracy requires multi-source data support, while data privacy requirements limit cross-institutional data sharing. Although some research attempts to introduce transfer learning methods, there are problems such as poor model interpretability and long adaptation period for new units. The existing technology has not solved the following key problems: (1) time series alignment method for multi-unit degradation data; (2) federated representation learning mechanism for nonlinear degradation characteristics; (3) model generalization strategy for small sample units. These problems result in the existing prediction system having obvious prediction lag and high false alarm rate in actual application, which restricts the industrial application of intelligent operation and maintenance technology. SUMMARY
[0005] The problem to be solved by the present application is to accurately predict the failure of the lubricating system of a steam turbine unit, and a steam turbine unit lubricating system federated learning degradation prediction method is proposed.
[0006] To achieve the above-mentioned purpose, the following technical solutions are adopted in the present application:
[0007] A steam turbine unit lubricating system federated learning degradation prediction method, comprising the following steps:
[0008] S1. Establishing a federated learning network topology structure, configuring a communication protocol between a central server and local nodes of a turbine unit lubricating system;
[0009] S2. Constructing a data acquisition module, a local processing module and a degradation prediction module at the local nodes of the turbine unit lubricating system, the data acquisition module, the local processing module and the degradation prediction module being connected in sequence, deploying a federated aggregation module at the central server, the federated aggregation module being connected to the local processing module and the degradation prediction module respectively;
[0010] S3. The data acquisition module acquires lubricating system operating parameters in real time;
[0011] S4. The local processing module performs data preprocessing and feature extraction on the lubricating system operating parameters obtained in step S3 to obtain turbine unit lubricating system feature parameters;
[0012] S5. The degradation prediction module performs input processing and feature fusion on the turbine unit lubricating system feature parameters obtained in step S4, and then inputs them into a deep neural network model with time series processing capability for training to obtain updated model parameters;
[0013] S6. The federated aggregation module encrypts the updated model parameters obtained in step S5, and then transmits the encrypted updated model parameters to the central server;
[0014] S7. The central server aggregates the updated model parameters from each node and decrypts the data, updates the global model parameters, obtains an updated global model and distributes it to each local node;
[0015] S8. In the local node, based on the updated global model obtained in step S7, the degradation prediction module generates a device degradation trend report including health state classification early warning and maintenance strategy suggestion.
[0016] Further, the lubricating system operating parameters in step S3 include oil viscosity sensor data, bearing vibration signals, temperature gradient distribution parameters and pressure pulsation characteristic quantities.
[0017] Further, the turbine unit lubricating system feature parameters obtained in step S4 include time domain statistics of vibration signals, frequency domain wavelet packet energy entropy, dielectric constant change rate of oil parameters, particle concentration gradient and dynamic fluctuation variance of temperature and pressure.
[0018] Further, the specific implementation method of step S5 includes the following steps:
[0019] S5.1. Input processing is performed on the turbine unit lubricating system feature parameters obtained in step S4, including feature alignment and time stamp synchronization processing;
[0020] S5.2. Feature fusion employs a gating mechanism to achieve cross-modal feature interaction across multi-source data, and the output satisfies the following relation:
[0021]
[0022] in, Represents the vibration eigenvector. Represents the temperature eigenvector. and These are the first trainable parameters and the second trainable parameters, respectively. It is the sigmoid activation function;
[0023] S5.3. Constructing a deep neural network model with time series processing capabilities includes the following steps: First, establish a feature encoder based on a temporal convolutional network. The mathematical expression is:
[0024]
[0025] in, express Sensor data sequences within a time window For encoder parameters; Indicates the temporal convolution symbol. The implicit feature representation of time t;
[0026] Then, a degradation decoder was designed. Calculate the difference between the current state and the baseline state, expressed as:
[0027]
[0028] in, This is the weight matrix corresponding to the degradation degree decoder. For the bias term corresponding to the degradation decoder, The degree of degradation.
[0029] Furthermore, in step S6, the federated aggregation module encrypts the updated model parameters obtained in step S5. The expression for the encryption process is as follows:
[0030]
[0031] in, For the first Model parameters for each local node. Indicates the use of a public key encryption function, For the first The encrypted model parameters of each local node.
[0032] Furthermore, in step S7, the central server uses a federated averaging algorithm for aggregation, with the aggregation rules as follows:
[0033]
[0034] in, The total number of participating nodes. Indicates the use of private key The decryption function, These are global model parameters;
[0035] A dynamic weighting strategy is adopted, based on node data quality evaluation indicators. Adjusting aggregate weights :
[0036]
[0037] in, For temperature coefficient, It is calculated by combining data completeness, sampling frequency, and signal-to-noise ratio. Score the data quality of the j-th node.
[0038] Furthermore, the multi-scale health index constructed in step S8 The expression is:
[0039]
[0040] in, For the first One characteristic in time The normalized value, for The feature weight coefficients, Indicates the rate of change of the characteristic. For temperature coefficient, The total number of features;
[0041] Then, the multi-scale health index is predicted using the following prediction function:
[0042]
[0043] in, These are the global model parameters obtained through federated learning. The time window length, This represents the federated learning prediction function. For health index series, This is a predicted value for a multi-scale health index;
[0044] The federal learning cycle is set to perform model aggregation updates every 24 hours, and in abnormal conditions, the adaptive learning mode is started to shorten the update interval to 1 hour when the multi-scale health index prediction value exceeds the baseline threshold When the primary warning is triggered, the failure threshold is exceeded When the shutdown protection is started.
[0045] The beneficial effects of the present application are:
[0046] The steam turbine unit lubrication system federal learning degradation prediction method provided by the application solves the privacy protection and data sharing contradiction problem of equipment state prediction in a multi-data source environment. By constructing a distributed machine learning framework, the collaborative training and optimization of the degradation prediction model are realized under the premise of protecting the data privacy of each participant, and the prediction accuracy and system generalization ability are effectively improved. BRIEF DESCRIPTION OF DRAWINGS
[0047] Figure 1 A flowchart of the steam turbine unit lubrication system federal learning degradation prediction method provided by the application. DETAILED DESCRIPTION
[0048] In order to make the purpose, technical scheme and advantages of the present application clearer, the present application will be further described in detail below in combination with the drawings and specific embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application, that is, the described specific embodiments are only a part of the embodiments of the present application, but not all the specific embodiments. The components of the specific embodiments of the present application described and shown in the drawings herein can be arranged and designed in various different configurations, and the present application can also have other embodiments.
[0049] Therefore, the detailed description of the specific embodiments of the present application provided in the drawings below is not intended to limit the scope of the claimed application, but only represents selected specific embodiments of the present application. Based on the specific embodiments of the present application, all other specific embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.
[0050] In order to further understand the invention content, characteristics and effects of the present application, the following specific embodiments are exemplified, and the drawings are Figure 1 The detailed description is as follows:
[0051] Example 1
[0052] A steam turbine unit lubrication system federal learning degradation prediction method comprises the following steps:
[0053] S1. Establishing a federated learning network topology, configuring a communication protocol between a central server and local nodes of a turbine lubrication system;
[0054] S2. Constructing a data acquisition module, a local processing module, a degradation prediction module, and a federated aggregation module at the local nodes of the turbine lubrication system, wherein the data acquisition module, the local processing module, the degradation prediction module, and the federated aggregation module are connected in sequence;
[0055] S3. The data acquisition module acquires lubrication system operating parameters in real time;
[0056] Further, the lubrication system operating parameters in step S3 include oil viscosity sensor data, bearing vibration signals, temperature gradient distribution parameters, and pressure pulsation characteristic quantities.
[0057] Further, the vibration sensor group is installed at key measurement points of the bearing seat, and a three-axis acceleration sensor is used to collect vibration frequency spectrum data; the oil quality monitoring unit integrates a dielectric constant detection probe and a particle counter to monitor lubricating oil physical and chemical indicators in real time; a temperature and pressure sensor array is arranged at key nodes of the lubricating oil pipeline, including a PT100 temperature sensor and a piezoresistive pressure sensor.
[0058] S4. The local processing module performs data preprocessing and feature extraction on the lubrication system operating parameters obtained in step S3 to obtain turbine lubrication system characteristic parameters;
[0059] Further, the turbine lubrication system characteristic parameters obtained in step S4 include time-domain statistics of vibration signals, frequency-domain wavelet packet energy entropy, dielectric constant change rate of oil parameters, particle concentration gradient, and dynamic fluctuation variance of temperature and pressure.
[0060] Further, the data preprocessing includes wavelet denoising processing of the vibration signals to eliminate high-frequency interference components; Z-score standardization processing of the oil particle count data; and sliding average filtering processing of the temperature and pressure data.
[0061] S5. The degradation prediction module performs input processing and feature fusion on the turbine lubrication system characteristic parameters obtained in step S4, and then inputs them into a deep neural network model with time series processing capability for training to obtain updated model parameters;
[0062] Further, the specific implementation method of step S5 includes the following steps:
[0063] S5.1. Input processing is performed on the turbine lubrication system characteristic parameters obtained in step S4, including feature alignment and timestamp synchronization processing;
[0064] Further, it also includes missing value filling, which adopts a dynamic interpolation method based on equipment working condition, specifically including:
[0065] When the missing duration , linear interpolation method is used to reconstruct the data; when , the LSTM predictor is started to generate substitute data, and the input is the previous complete cycle data window:
[0066] ;
[0067] S5.2. Feature fusion adopts a gating mechanism to realize multi-source data cross-modal feature interaction, and outputs satisfy the relationship:
[0068]
[0069] Wherein, represents the vibration feature vector, represents the temperature feature vector, and are the first trainable parameter, the second trainable parameter, sigmoid activation function;
[0070] S5.3. Constructing a deep neural network model with time series processing capability includes the following steps: first, establish a feature encoder based on time convolution network , the mathematical expression is:
[0071]
[0072] Wherein, represents sensor data sequence in the time window, encoder parameter; represents the time convolution symbol, represents the implicit feature representation at time t;
[0073] Then design the degradation degree decoder , calculate the difference degree between the current state and the reference state, the expression is:
[0074]
[0075] Wherein, is the weight matrix corresponding to the degradation degree decoder, is the bias term corresponding to the degradation degree decoder, is the degradation degree.
[0076] S6. The federated aggregation module encrypts the updated model parameters obtained in step S5, and then transmits the encrypted updated model parameters to the central server.
[0077] Furthermore, in step S6, the federated aggregation module encrypts the updated model parameters obtained in step S5. The expression for the encryption process is as follows:
[0078]
[0079] in, For the first Model parameters for each local node. Indicates the use of a public key encryption function, For the first The encrypted model parameters of each local node.
[0080] Furthermore, the encryption process includes performing a random permutation operation before uploading the local model parameters; and generating a unique model fingerprint vector for each local node. Malicious model updates are identified based on the cosine similarity index, and the judgment criteria are as follows:
[0081]
[0082] in For the preset threshold, The standard deviation of historical similarity.
[0083] S7. The central server aggregates the updated model parameters from each node, decrypts the data, updates the global model parameters, obtains the updated global model, and distributes it to each local node.
[0084] Furthermore, in step S7, the central server uses a federated averaging algorithm for aggregation, with the aggregation rules as follows:
[0085]
[0086] in, The total number of participating nodes. Indicates the use of private key The decryption function, These are global model parameters;
[0087] A dynamic weighting strategy is adopted, based on node data quality evaluation indicators. Adjusting aggregate weights :
[0088]
[0089] in, For temperature coefficient, The data integrity, sampling frequency and signal-to-noise ratio are comprehensively calculated, The data quality score of the jth node is calculated.
[0090] Further, the dynamic weighted average algorithm further comprises calculating a weight correction coefficient according to the data acquisition integrity index of each node and the noise level index ; generating a performance weight based on the validation set accuracy and the loss value ; and the maximum weight calculation satisfies:
[0091] ;
[0092] Further, the federated training mechanism is implemented through the following process: each edge node maintains a local model , periodically uploads the model parameters to the aggregation server ; the aggregation server performs weighted average calculation:
[0093]
[0094]
[0095] wherein is a dynamic weight coefficient, and the calculation formula is:
[0096]
[0097] In the formula, Qk represents the data quality score of the kth node, is a temperature coefficient; The degradation index calculation module fuses multi-dimensional features and defines the comprehensive degradation degree
[0098] as follows:
[0099] wherein
[0100] represents the i-th sensor feature value output by the k-th local model, is a feature weight coefficient, is a trend adjustment factor; The prediction output layer adopts a sliding window mechanism to generate a degradation trend curve, and a time series prediction model
[0101] is used to calculate the predicted value of the next T time steps:
[0102]
[0103] In the formula, For the length of the history window, These are the parameters for the prediction model. The system's final output includes real-time degradation values, trend prediction curves, and estimated remaining useful life.
[0104] S8. In the local node, based on the updated global model obtained in step S7, the degradation prediction module generates a device degradation trend report, including health status grading warnings and maintenance strategy recommendations.
[0105] Furthermore, the multi-scale health index constructed in step S8 The expression is:
[0106]
[0107] in, For the first One characteristic in time The normalized value, for The feature weight coefficients, Indicates the rate of change of the characteristic. For temperature coefficient, The total number of features;
[0108] Then, the multi-scale health index is predicted using the following prediction function:
[0109]
[0110] in, These are the global model parameters obtained through federated learning. The time window length, This represents the federated learning prediction function. For health index series, This is a multi-scale health index prediction value; the prediction model's network structure includes 3 layers of LSTM units and 1 layer of attention mechanism.
[0111] The federated learning cycle is set to perform model aggregation updates every 24 hours. Under abnormal operating conditions, an adaptive learning mode is activated to shorten the update interval to 1 hour. When the predicted value of the multi-scale health index exceeds the baseline threshold... A primary warning is triggered when the failure threshold is exceeded. Start the shutdown protection when needed.
[0112] Furthermore, the health status classification early warning includes calculating the current health index of the device. Compared with the benchmark value Deviation:
[0113]
[0114] When a first-level early warning is triggered; when a second-level early warning is triggered and a maintenance work order is generated; and when a third-level early warning is triggered and a device shutdown protection instruction is executed.
[0115] It should be noted that the terms "first", "second", and so on, and the like relational terms merely serve to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "comprise", "include" or any other variant thereof are intended to cover non-exclusive inclusions, so that a process, method, article or device that includes a list of elements does not only include those elements, but also includes other elements not explicitly listed, or further includes elements inherent to such process, method, article or device. Without more limitations, the element defined by the phrase "comprising a" does not exclude the presence of additional identical elements in the process, method, article or device that includes the element.
[0116] Although the present application has been described above with reference to specific embodiments, various modifications can be made to it without departing from the scope of the application, and equivalent components can be substituted therefor. In particular, each feature disclosed in the specific embodiments of the present application can be used in any combination with one another, provided that there is no structural conflict, and the fact that these combinations are not described in the present specification is merely due to the consideration of omitting the description and saving resources. Therefore, the present application is not limited to the specific embodiments disclosed herein, but includes all technical solutions falling within the scope of the claims.
Claims
1. A steam turbine unit lubrication system federated learning degradation prediction method, characterized in that, Comprising the following steps: S1. Establishing a federated learning network topology, configuring a communication protocol between a central server and local nodes of a turbine lubrication system; S2. Building a data acquisition module, a local processing module and a degradation prediction module on the local nodes of the turbine lubrication system, the data acquisition module, the local processing module and the degradation prediction module being connected in sequence, deploying a federated aggregation module on the central server, the federated aggregation module being connected to the local processing module and the degradation prediction module respectively; S3. The data acquisition module acquires lubrication system operating parameters in real time; S4. The local processing module performs data preprocessing and feature extraction on the lubrication system operating parameters obtained in step S3 to obtain turbine lubrication system feature parameters; S5. The degradation prediction module inputs the turbine lubrication system feature parameters obtained in step S4, fuses features, and then inputs them into a deep neural network model with time series processing capability for training to obtain updated model parameters; S6. The federated aggregation module encrypts the updated model parameters obtained in step S5, and then transmits the encrypted updated model parameters to the central server; S7. The central server aggregates the updated model parameters from each node and decrypts the data, updates the global model parameters, obtains an updated global model and distributes it to each local node; S8. In the local node, based on the updated global model obtained in step S7, the degradation prediction module generates a device degradation trend report including health state classification warning and maintenance strategy suggestion.
2. The method of claim 1, wherein the method further comprises: The lubrication system operating parameters in step S3 include oil viscosity sensor data, bearing vibration signals, temperature gradient distribution parameters and pressure pulsation characteristic quantities.
3. The method of claim 2, wherein the method further comprises: The turbine lubrication system feature parameters obtained in step S4 include time domain statistics of vibration signals, frequency domain wavelet packet energy entropy, dielectric constant change rate of oil parameters, particle concentration gradient and dynamic fluctuation variance of temperature and pressure.
4. The method of claim 3, wherein the method further comprises: The specific implementation method of step S5 comprises the following steps: S5.
1. Input processing of the turbine lubrication system feature parameters obtained in step S4, including feature alignment and timestamp synchronization processing; S5.
2. Feature fusion adopts a gating mechanism to realize multi-source data cross-modal feature interaction, and outputs a relationship that satisfies: wherein, represents a vibration feature vector, represents a temperature feature vector, and are a first trainable parameter, a second trainable parameter, is a sigmoid activation function; S5.
3. Building a deep neural network model with time series processing capability includes the following steps: first, a feature encoder based on a time convolution network is established , which is mathematically expressed as: wherein, denotes a sequence of sensor data within a time window, is an encoder parameter; denotes a temporal convolutional symbol, denotes a latent feature representation at time instant t; A degeneration degree decoder is then designed The difference between the current state and the reference state is calculated, expressed as: wherein, is a weight matrix corresponding to the degradation degree decoder, is a bias term corresponding to the degradation degree decoder, is a degradation degree.
5. The method of claim 4, wherein the method further comprises: In step S6, the federated aggregation module encrypts the updated model parameters obtained in step S5, and the expression of the encryption process is: in, For the first Model parameters for each local node. Indicates the use of a public key encryption function, For the first The encrypted model parameters of each local node.
6. The method of claim 5, wherein the method further comprises: In step S7, the central server uses a federated averaging algorithm for aggregation, and the aggregation rule is: in, The total number of participating nodes. Indicates the use of private key The decryption function, These are global model parameters; Adopting dynamic weighting strategy, according to node data quality evaluation index Adjusting aggregation weight : wherein, is the temperature coefficient, is calculated by comprehensive data integrity, sampling frequency and signal-to-noise ratio, is the data quality score of the jth node.
7. The method of claim 6, wherein the method further comprises: The multi-scale health index constructed in step S8 The expression is: wherein, is the normalized value of the th feature at time , is the feature weight coefficient of the th feature, represents the feature change rate, is the temperature coefficient, is the total number of features; Then the multi-scale health index is predicted, and the prediction function is: wherein, global model parameters obtained for federated learning, is a length of a time window, denotes a federated learning prediction function, is a health index sequence, is a multi-scale health index prediction value; The federal learning cycle is set to perform model aggregation updates every 24 hours, in abnormal conditions, the adaptive learning mode is started to shorten the update interval to 1 hour, when the multi-scale health index prediction value exceeds the baseline threshold The primary warning is triggered, and the shutdown protection is started when the failure threshold is exceeded The primary warning is triggered, and the shutdown protection is started when the failure threshold is exceeded