Turbine shafting vibration intelligent diagnosis and edge early warning method based on multi-source data fusion

Abstract

The invention relates to the technical field of steam turbine shafting vibration monitoring, and discloses a steam turbine shafting vibration intelligent diagnosis and edge early warning method based on multi-source data fusion, and the method comprises the following steps: deploying multiple types of sensors at key measurement points of a steam turbine shafting, and constructing a monitoring network with complementary spatial dimensions; according to the method, multiple types of sensors are deployed at key measuring points to collect multi-source data, time sequence and frequency domain features are extracted, feature layer fusion is carried out, the weight is dynamically adjusted in combination with real-time working conditions, the limitation of a single data source is broken through, and the shafting state is comprehensively reflected; local real-time diagnosis is realized by deploying a lightweight model through edge nodes, and meanwhile, the model is updated by incremental learning to improve the generalization ability of complex working conditions, so that the problem of high cloud processing delay is solved, and the real-time early warning requirement is met; the false alarm rate is reduced by combining a dynamic threshold adjustment mechanism with equipment health baseline and real-time working condition deviation correction, and performing graded early warning on the overrun amplitude and confidence of a reference vibration predicted value.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of steam turbine shafting vibration monitoring, in particular to a steam turbine shafting vibration intelligent diagnosis and edge early warning method based on multi-source data fusion. BACKGROUND

[0002] As the core power transmission component of the steam turbine, the steam turbine shafting is composed of key components such as rotors, shaft couplings, bearings, etc., and its vibration state directly reflects the stability of the unit operation - vibration amplitude exceeding the limit will aggravate the bearing bush wear, rotor fatigue and seal damage, and in severe cases, it will cause shafting rubbing, large shaft bending and other malignant accidents. Therefore, the shafting vibration parameter is regarded as a core index to measure the safe operation state of the unit, which has a decisive significance for preventing unplanned shutdown and ensuring power generation efficiency. Although the traditional diagnosis method has realized vibration monitoring to a certain extent, there are still some problems under complex working conditions:

[0003] 1. The existing technology relies on single vibration sensor data and cannot fully reflect the shafting state; for example, the announcement number: CN112611564B discloses a steam turbine support bearing stiffness discrimination method and device, which only calculates the resonance amplification factor through vibration data without fusing multi-dimensional information such as temperature and pressure;

[0004] 2. The traditional method needs to upload data to the cloud for analysis and processing, and the deep learning model has weak generalization ability under complex working conditions, resulting in high diagnosis delay and inability to meet the real-time early warning demand;

[0005] 3. The existing early warning method relies on preset thresholds and cannot adapt to changes in working conditions, is easily disturbed by working condition fluctuations, and has a high false alarm rate;

[0006] Therefore, a steam turbine shafting vibration intelligent diagnosis and edge early warning method based on multi-source data fusion is proposed. SUMMARY

[0007] The purpose of the present application is to provide a steam turbine shafting vibration intelligent diagnosis and edge early warning method based on multi-source data fusion to solve the problems of the existing technology that rely on single vibration sensor data to cause the shafting state to be not fully reflected, data uploading to the cloud for analysis to cause diagnosis delay and inability to meet the real-time early warning demand, and relying on preset thresholds to make it difficult to adapt to changes in working conditions and have a high false alarm rate.

[0008] To achieve the above purpose, the present application provides the following technical scheme: a steam turbine shafting vibration intelligent diagnosis and edge early warning method based on multi-source data fusion, comprising the following steps:

[0009] Step 1: Deploy multiple types of sensors at key measurement points of the steam turbine shafting to build a monitoring network with complementary spatial dimensions;

[0010] Step two, the monitoring network based on spatial dimension complement collects multi-source data of the steam turbine shaft system, and pre-processes the collected multi-source data;

[0011] Step three, time trend features and frequency domain fault features are extracted from the pre-processed multi-source data to construct a multi-dimensional feature vector matrix;

[0012] Step four, based on the multi-dimensional feature vector matrix, a multi-source data fusion algorithm based on feature layer fusion is used for feature fusion, and the fusion weights of different types of sensor data are dynamically adjusted according to the real-time working condition of the steam turbine to generate a fusion feature set;

[0013] Step five, input the fusion feature set into the lightweight model deployed in the edge node to output the fault type, confidence and vibration prediction value in real time;

[0014] Step six, based on the fault type, confidence and vibration prediction value, the dynamic threshold adjustment mechanism is used to monitor the data distribution, trigger the hierarchical early warning, and synchronously associate the cloud analysis;

[0015] Step seven, based on the hierarchical early warning information and the log accumulation sample library uploaded by the edge node, the lightweight model deployed in the edge node is updated, and the digital twin analysis is combined to predict the service life of the steam turbine shaft system components.

[0016] It is further preferred as the technical solution to provide that in step seven, the digital twin analysis is based on a shaft dynamics model, and the real-time monitoring data are fused to calculate the fatigue damage of the steam turbine shaft system components.

[0017] It is further preferred as the technical solution to provide that in step four, the multi-source data fusion algorithm uses D-S evidence theory to generate a basic probability assignment matrix by establishing a mapping relationship between the feature vector and the fault mode; and the real-time working condition of the steam turbine includes unit load, steam parameter and rotating speed.

[0018] It is further preferred as the technical solution to provide that in step two, the multi-source data includes vibration signal, temperature signal, pressure signal and rotating speed signal; and the pre-processing includes signal conditioning and noise purification.

[0019] It is further preferred as the technical solution to provide that in step three, the method of constructing the multi-dimensional feature vector matrix includes the following steps:

[0020] Step 301, for the vibration signal, the peak value, kurtosis, skewness, pulse index and root mean square value are extracted as time domain statistical features;

[0021] Step 302, for the temperature signal, pressure signal and rotating speed signal, the mean value, range, change rate, fluctuation amplitude and steady-state deviation are extracted as parameter time domain features;

[0022] Step 303: Perform frequency domain transformation on the vibration signal, obtain the frequency domain spectrum through Fourier transform, and extract the amplitude and phase of the power frequency, second harmonic, 0.5 harmonic and higher harmonics to form frequency domain fault characteristics;

[0023] Step 304: Concatenate the time-domain statistical features, parameter time-domain features, and frequency-domain fault features in an orderly manner according to the signal source and feature attributes to construct a multi-dimensional feature vector matrix. The row dimension of the matrix corresponds to the sampling time sequence, and the column dimension corresponds to the feature category.

[0024] As a further preferred embodiment of this technical solution, in step five, the lightweight model is a neural network based on deep separable convolution, which has the functions of fault type identification, confidence output and vibration trend prediction. The architecture of the lightweight model includes an input layer, a feature extraction layer, a multi-task fusion layer and a multi-branch output layer.

[0025] Input layer: Used to standardize the input fused feature set to obtain a 32-dimensional standardized feature vector;

[0026] Feature extraction layer: used to perform high-order feature extraction and dimensionality compression on the 32-dimensional standardized feature vector to obtain a 1D high-order feature map;

[0027] Multi-task fusion layer: used to perform channel compression and feature integration processing on the 1D high-order feature map to obtain a 32-dimensional shared feature vector;

[0028] Multi-branch output layer: including fault classification branch, confidence branch and vibration prediction branch;

[0029] The fault classification branch is used to process the 32-dimensional shared feature vector using a multi-task loss function to obtain a 5-dimensional probability vector.

[0030] The confidence branch reuses the output of the fault classification branch and takes the probability value corresponding to the fault category with the highest probability in the results as the confidence level.

[0031] The vibration prediction branch includes an LSTM layer, a fully connected layer, and a linear activation function, which are used to obtain a vibration prediction value representing the future trend of vibration amplitude change based on the 32-dimensional shared feature vector.

[0032] As a further preferred option of this technical solution, the multi-branch model is trained using a multi-task joint training mode. By integrating classification loss and regression loss, it can optimize three branches simultaneously during one training session. The fault classification loss and vibration prediction loss are weighted and summed to obtain the total model loss. The total loss formula is as follows: Total loss = α × classification loss + β × regression loss.

[0033] The classification loss uses the cross-entropy loss function;

[0034] The regression loss is the mean squared error loss;

[0035] Weight settings: α=0.6, β=0.4.

[0036] As a further preferred embodiment of this technical solution, in step six, the dynamic threshold adjustment mechanism is used to correct the deviation between the equipment health baseline and the real-time operating condition; the graded early warning is divided into four levels based on the vibration prediction value exceeding the limit and the confidence level, and the four levels correspond to different response measures.

[0037] As a further preferred embodiment of this technical solution, in step seven, the log accumulation sample library is stored according to fault type and operating condition; the lightweight model deployed on the updated edge nodes adopts an incremental learning approach, optimizing local parameters only for newly added samples.

[0038] As a further preferred embodiment of this technical solution: In step one, the multi-type sensors include a triaxial vibration acceleration sensor, a bearing temperature sensor, a lubricating oil pressure sensor, and a speed sensor; the key measuring points include the turbine bearing housing, coupling, journal, and lubricating oil pipeline; at least one triaxial vibration acceleration sensor is deployed in each bearing housing along the X, Y, and Z directions to form a spatial three-dimensional monitoring array.

[0039] As a further preferred embodiment of this technical solution, the edge node and the cloud transmit data through an encrypted protocol, the edge node performs real-time diagnostics, and the cloud performs offline analysis and lightweight model updates.

[0040] Compared with the prior art, the present invention, employing the above technical solution, has the following technical effects:

[0041] 1. This invention collects multi-source data such as vibration, temperature, pressure, and rotational speed by deploying multiple types of sensors at key measuring points, extracts time-series and frequency-domain features and performs feature layer fusion, and dynamically adjusts weights in combination with real-time operating conditions, thus breaking through the limitations of a single data source and comprehensively reflecting the state of the shaft system.

[0042] 2. This invention achieves local real-time diagnosis by deploying a lightweight model on edge nodes, only synchronizing logs to the cloud, and uses incremental learning to update the model to improve the generalization ability of complex working conditions, solving the problem of high latency in cloud processing and meeting the needs of real-time early warning.

[0043] 3. This invention combines a dynamic threshold adjustment mechanism with equipment health baseline and real-time operating condition deviation correction to provide graded early warning of the over-limit amplitude and confidence level of the reference vibration prediction value, adapting to changes in operating conditions and reducing the false alarm rate. Attached Figure Description

[0044] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0045] Figure 1 This is a flowchart illustrating a method for intelligent diagnosis and edge warning of turbine shaft vibration based on multi-source data fusion according to the present invention.

[0046] Figure 2 This is a flowchart illustrating the method for constructing a multi-dimensional feature vector matrix according to the present invention. Detailed Implementation

[0047] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0048] It should be noted that the structures, proportions, sizes, etc., shown in the accompanying drawings of this specification are only for the purpose of assisting those skilled in the art in understanding and reading the content disclosed in the specification, and are not intended to limit the conditions under which this application can be implemented. Therefore, they have no substantial technical significance. Any modifications to the structure, changes in the proportions, or adjustments to the size should still fall within the scope of the technical content disclosed in this application, provided that they do not affect the effects and purposes that this application can produce.

[0049] Example 1

[0050] In existing technologies, relying on data from a single vibration sensor results in an incomplete reflection of the shaft system status, data upload to the cloud for analysis causes diagnostic delays that fail to meet real-time early warning requirements, and relying on preset thresholds makes it difficult to adapt to changes in operating conditions and results in a high false alarm rate.

[0051] Figure 1 This is a flowchart illustrating an intelligent diagnosis and edge warning method for turbine shaft system vibration based on multi-source data fusion, according to an embodiment of the present invention. It should be noted that if substantially the same result is obtained, the method of this application is not necessarily identical. Figure 1 The illustrated process sequence is limited. For example... Figures 1-2 As shown: A method for intelligent diagnosis and edge warning of turbine shaft system vibration based on multi-source data fusion includes the following steps:

[0052] Step 1: Deploy multiple types of sensors at key measuring points in the turbine shaft system to construct a monitoring network with complementary spatial dimensions;

[0053] Specifically, first, the various types of sensors and key measuring points to be deployed are identified. The various types of sensors include triaxial vibration acceleration sensors, bearing temperature sensors, lubricating oil pressure sensors, and speed sensors. The key measuring points cover the turbine's bearing housings, couplings, journals, and lubricating oil pipelines.

[0054] Then, sensors are deployed at key measuring points, with at least one triaxial vibration acceleration sensor deployed along the X, Y, and Z directions in each bearing housing to form a three-dimensional spatial monitoring array to comprehensively capture the vibration state of the bearing housing.

[0055] Next, a bearing temperature sensor is deployed at the bearing location to monitor the bearing operating temperature in real time; a lubricating oil pressure sensor is deployed in the lubricating oil pipeline to collect lubricating oil pressure data; and a speed sensor is deployed near the coupling or journal to obtain unit speed information.

[0056] Finally, by distributing the aforementioned sensors at different key measuring points, spatial complementary monitoring of multi-dimensional parameters such as shaft vibration, temperature, pressure, and rotational speed is achieved, thus constructing a spatially complementary monitoring network.

[0057] Step 2: Collect multi-source data of the turbine shaft system based on a monitoring network with complementary spatial dimensions, and preprocess the collected multi-source data;

[0058] Specifically, firstly, a monitoring network with complementary spatial dimensions is activated. Through the deployment of triaxial vibration acceleration sensors, bearing temperature sensors, lubricating oil pressure sensors, and speed sensors, vibration signals, temperature signals, pressure signals, and speed signals of the turbine shaft system are collected simultaneously to form initial multi-source data.

[0059] Then, the collected multi-source data is subjected to signal conditioning. The specific processing procedure is as follows:

[0060] The vibration signal is quantized using a 16-bit analog-to-digital converter with a sampling rate of 8-12kHz, converting the analog vibration signal into a digital signal; the temperature and pressure signals are converted using a 12-bit analog-to-digital converter with a sampling rate of 1-2kHz, completing the conversion from analog to digital; the rotational speed signal is pulse shaped to extract the effective frequency information.

[0061] Finally, the processed multi-source data is subjected to noise reduction and purification. The specific processing steps are as follows:

[0062] The vibration signal is decomposed using a 5-level db4 wavelet base combined with a soft threshold function to suppress high-frequency noise while retaining effective vibration characteristics. The temperature and pressure signals are processed by a differential amplifier circuit to eliminate common-mode interference and improve signal stability. The speed signal is processed by a moving average filter to remove instantaneous fluctuations and obtain smooth speed data, thus completing the preprocessing process.

[0063] Step 3: Extract time-series trend features and frequency domain fault features from the preprocessed multi-source data, and construct a multi-dimensional feature vector matrix;

[0064] Specifically, first, the data types of the preprocessed multi-source signals are determined, including vibration signals, temperature signals, pressure signals, and rotational speed signals. The feature categories to be extracted are time-series trend features (including time-domain statistical features and parameter time-domain features) and frequency-domain fault features.

[0065] Then, time-domain statistical features are extracted from the vibration signal: its peak value, kurtosis, skewness, impulse index and root mean square value are calculated as components of the time-series trend features of the vibration signal;

[0066] Next, time-domain features of parameters are extracted for temperature, pressure and speed signals: the mean, range, rate of change, fluctuation amplitude and steady-state deviation of the three types of signals are calculated respectively to form the time-series trend features of the process parameters;

[0067] Next, frequency domain transformation is performed on the vibration signal: the vibration signal is transformed from the time domain to the frequency domain through Fourier transform to obtain the frequency domain spectrum, from which the amplitude and phase of the power frequency, second harmonic, 0.5 harmonic and higher harmonics are extracted to form frequency domain fault characteristics;

[0068] Finally, the extracted time-domain statistical features, parameter time-domain features, and frequency-domain fault features are sequentially concatenated according to signal source (such as vibration, temperature, pressure, and rotational speed) and feature attributes (time-series trend, frequency-domain fault) to construct a multi-dimensional feature vector matrix. The row dimension of the matrix corresponds to the data sampling time sequence, and the column dimension corresponds to different feature categories, thus completing the matrix construction.

[0069] Step 4: Based on the multi-dimensional feature vector matrix, a multi-source data fusion algorithm based on feature layer fusion is used for feature fusion. At the same time, the fusion weights of different types of sensor data are dynamically adjusted in combination with the real-time operating conditions of the steam turbine to generate a fused feature set.

[0070] Specifically, first, the multi-dimensional feature vector matrix constructed in step three is obtained, and real-time operating data of the steam turbine is collected, including unit load, steam parameters and speed.

[0071] Then, the DS evidence theory is used as the multi-source data fusion algorithm for feature layer fusion. By establishing the mapping relationship between feature vectors and fault modes, a basic probability assignment matrix is ​​generated for the multi-dimensional feature vector matrix, which provides a basis for feature fusion.

[0072] Next, the fusion weights of different types of sensor data are dynamically adjusted based on real-time operating conditions. According to the real-time changes in unit load, steam parameters and speed, the weights of vibration signal characteristics, temperature signal characteristics, pressure signal characteristics and speed signal characteristics are adaptively allocated to ensure that the weight ratio of key features is reasonable under different operating conditions.

[0073] Finally, based on the generated basic probability assignment matrix and the dynamically adjusted weights, feature layer fusion operation is performed using DS evidence theory to integrate multi-dimensional features into a unified fusion feature set, thus completing the feature fusion process.

[0074] Step 5: Input the fused feature set into the lightweight model deployed at the edge nodes, and output the fault type, confidence level and vibration prediction value in real time;

[0075] Specifically, first, confirm that the edge nodes have deployed a lightweight neural network model based on depthwise separable convolution and complete the model initialization to ensure that the model has the functions of fault type identification, confidence output and vibration trend prediction.

[0076] Then, the fusion feature set generated in step four is standardized according to the model input format and converted into input data that is suitable for the lightweight model.

[0077] Next, the standardized fused feature set is input into the lightweight model of the edge nodes to start the real-time inference process of the model and quickly process the feature data by utilizing the efficient computing power of depthwise separable convolution.

[0078] Finally, after the model inference is completed, the diagnostic results are output in real time, including the specific fault type (such as imbalance, misalignment, oil film eddy, etc.), the confidence level (probability value) of the fault type, and the vibration prediction value for a future period of time (such as 5-10 minutes), thus completing the processing in step five.

[0079] Step 6: Based on the fault type, confidence level, and vibration prediction value, monitor the data distribution using a dynamic threshold adjustment mechanism, trigger tiered early warnings, and simultaneously link them to cloud-based analysis;

[0080] Specifically, first, obtain the fault type, confidence level and vibration prediction value output in step five, and at the same time retrieve the equipment health baseline data (established based on historical normal operation data) and the real-time operating parameters of the turbine (such as unit load, speed, etc.).

[0081] Then, the dynamic threshold adjustment mechanism is activated. Based on the equipment health baseline, the correction coefficient is calculated in combination with real-time operating condition deviations (such as load fluctuations and speed deviations) to dynamically correct the baseline threshold and obtain a dynamic threshold that adapts to the current operating condition.

[0082] Next, monitor the distribution of vibration prediction values ​​relative to the dynamic threshold, calculate the over-limit amplitude of vibration prediction values ​​(the proportion exceeding the dynamic threshold), and combine the severity of the fault type and the corresponding confidence level to classify the warning level (four levels in total) according to preset rules.

[0083] Then, corresponding response measures are triggered according to the determined warning level: such as triggering local log recording for a level 1 warning, activating audible and visual alarms for a level 2 warning, pushing information to the operation and maintenance terminal for a level 3 warning, and linking emergency shutdown prompts for a level 4 warning.

[0084] Finally, through the encryption protocol between the edge nodes and the cloud, the warning level, fault type, confidence level, vibration prediction value and real-time operating condition parameters are synchronously uploaded to the cloud, triggering the cloud offline analysis process (such as fault case matching, in-depth tracing, etc.) to complete the synchronous association.

[0085] Step 7: Based on the hierarchical early warning information and the accumulated sample library of logs uploaded by edge nodes, update the lightweight model deployed on the edge nodes, and combine digital twin analysis to predict the life of turbine shaft system components.

[0086] Specifically, firstly, the hierarchical early warning information generated in step six (including early warning level, fault type, confidence level, etc.) and the operation logs uploaded by edge nodes (including multi-source data, feature vectors, operating parameters, etc.) are collected. The data are then classified and organized according to fault type (such as imbalance, misalignment) and real-time operating conditions (such as unit load, speed range) and stored in the sample library to complete sample accumulation.

[0087] Then, based on the new samples in the sample library, the lightweight model deployed on the edge nodes is updated using an incremental learning approach; only the local parameters of the model (such as the weights of the fully connected layers of the neural network) are fine-tuned for the new sample data, avoiding full retraining, reducing the computing power consumption of the edge nodes, and ensuring the efficiency of model updates;

[0088] Next, the digital twin system constructed by the shaft system dynamics model is invoked, and real-time monitoring data (such as vibration amplitude, temperature, pressure, etc.) is input into the digital twin model. The data is then fused to calculate the cumulative fatigue damage of key components of the turbine shaft system (bearings, couplings, rotors, etc.).

[0089] Finally, based on the cumulative fatigue damage and component material properties, the remaining life of the shaft components is predicted using a digital twin system, and the prediction results are synchronized to the equipment management system to complete the life prediction process.

[0090] The lightweight model architecture includes an input layer, a feature extraction layer, a multi-task fusion layer, and a multi-branch output layer, while also incorporating lightweight design details to adapt to edge devices.

[0091] Input layer: fuses feature set dimensions and input format;

[0092] Feature extraction layer: 3 depthwise separable convolutional blocks, 2 max pooling layers, and convolutional kernels;

[0093] Multi-task fusion layer: one 1×1 convolutional layer, global average pooling; function: to integrate the output of the feature extraction layer and share features across multiple branches;

[0094] Multi-branch output layer: fault classification branch, confidence branch: reuse the Softmax output of the classification branch and the vibration prediction branch;

[0095] In one specific embodiment, the model building method is as follows:

[0096] Traditional 1D convolution is split into single-channel independent convolutions to extract spatial features and fuse channel features;

[0097] Taking an input channel of 16, an output channel of 32, and a convolution kernel of 3×1 as an example, the computational cost is only 1 / 16 of that of traditional convolution.

[0098] Fault classification and vibration prediction share the same feature extraction layer, with only the output layer branch, avoiding the need to train two separate models and reducing the storage footprint of edge nodes.

[0099] First, the feature extraction layer is pre-trained on the public steam turbine fault dataset, and then the output layer is fine-tuned using historical data of the target unit to solve the problem of insufficient target samples.

[0100] Fine-tuning the parameters of the fully connected layer and LSTM layer, freezing the feature extraction layer, and reducing the update time from hours to minutes to adapt to edge node requirements.

[0101] The model weights were quantized from 32-bit floating-point to 8-bit integers and optimized using TensorRT, resulting in a 3-4x increase in inference speed, a reduction in model size, and adaptation to storage limitations on edge devices.

[0102] Cache three consecutive identical fault diagnosis results to skip repeated inference and reduce computing power consumption;

[0103] Input layer: The input data is a fused feature set, which is used to standardize the input data to obtain the output data, which is a 32-dimensional standardized feature vector.

[0104] Feature extraction layer: The input data is a 32-dimensional standardized feature vector output by the input layer, which is used to perform high-order feature extraction and dimensionality compression on the input data to obtain the output data, which is a 1×8×64 1D high-order feature map;

[0105] Multi-task fusion layer: The input data is a 1×8×64 1D high-order feature map output by the feature extraction layer, which is used to perform channel compression and feature integration processing on the input data to obtain the output data, which is a 32-dimensional shared feature vector.

[0106] Multi-branch output layer: The input data is a 32-dimensional shared feature vector output by the multi-task fusion layer, which is used to process the input data using a multi-task loss function to obtain the output data, which is a 5-dimensional probability vector, confidence value and vibration prediction value.

[0107] The fault classification branch has the following processing logic: Fully connected layer 1: 32 dimensions reduced to 16 dimensions; Fully connected layer 2: 16 dimensions reduced to 5 dimensions; 3. Softmax activation function;

[0108] The confidence branch reuses the Softmax output of the fault classification branch and takes the probability value corresponding to the fault category with the highest probability as the confidence score.

[0109] Vibration prediction branch, LSTM layer, fully connected layer and linear activation function.

[0110] The multi-branch model adopts a multi-task joint training mode, which integrates classification loss and regression loss to optimize three branches at the same time during training; the fault classification loss and vibration prediction loss are weighted and summed as the total loss of the model, as shown in the following formula: Total loss = α × classification loss + β × regression loss;

[0111] Classification loss: Cross-entropy loss function is used;

[0112] Regression loss: Mean squared error loss is used;

[0113] Weight settings: α=0.6, β=0.4;

[0114] Training process:

[0115] Dataset preparation: Labeled multi-task samples are used, each sample contains 32-dimensional fusion features, 1 real fault label, and 1 real vibration value for the next 5 minutes;

[0116] Model initialization: The parameters of the feature extraction layer, multi-task fusion layer, and multi-branch output layer are all randomly initialized. The LSTM layer uses orthogonal initialization to avoid gradient vanishing.

[0117] Joint training:

[0118] Each iteration inputs a batch of samples, and the model simultaneously propagates forward to output the fault probability, confidence level, and vibration prediction value.

[0119] Calculate the classification loss and regression loss, and sum them by weight to obtain the total loss;

[0120] During backpropagation, the gradient of the total loss updates the parameters of the feature extraction layer, the multi-task fusion layer, and the three branches simultaneously.

[0121] Incremental update: When adding new samples, only the parameters of the fully connected layer and LSTM layer are fine-tuned, and the same multi-task loss function is used for training. The update time is ≤5 minutes.

[0122] Once the trained model is deployed on an edge node, it can output three types of results simultaneously with a single input, without the need for branching calls. The specific usage process is as follows:

[0123] Input preparation: Edge nodes receive the 32-dimensional fusion feature set generated in step four in real time and process it according to the standardization rules of the input layer;

[0124] Model inference: Input the standardized 32-dimensional feature vector into the model and propagate it forward once;

[0125] Results and Analysis:

[0126] Fault type: Select the category with the highest probability from the 5-dimensional probability vector of the fault classification branch;

[0127] Confidence level: Directly extract the highest probability value mentioned above;

[0128] Vibration prediction value: Take the 1D output value of the vibration prediction branch;

[0129] Early warning linkage: The three types of results are simultaneously transmitted to the dynamic threshold adjustment mechanism to trigger tiered early warnings.

[0130] In one embodiment, specifically: in step seven, the digital twin analysis is based on the shaft dynamics model and integrates real-time monitoring data to calculate the fatigue damage of the turbine shaft components;

[0131] Specifically, firstly, the core foundation of digital twin analysis is the shaft dynamics model, which is a digital replica of the turbine shaft system (including key components such as rotors, bearings, and couplings). This model includes the physical properties of each component (such as material elastic modulus, density, and fatigue limit), structural parameters (such as rotor diameter and bearing clearance), and dynamic characteristics (such as stiffness, damping, and critical speed). It can accurately simulate the motion state of the shaft system under different operating conditions (such as vibration response and force transmission path).

[0132] Secondly, the integration of real-time monitoring data is a key step. The real-time monitoring data comes from the multi-source data collected and preprocessed in step two (effective features extracted after fusion in step four), including: vibration amplitude and frequency at each measuring point of the shaft system (reflecting dynamic load), bearing temperature (affecting material fatigue performance), lubricating oil pressure (related to the influence of lubrication state on friction and wear), and unit speed (determining stress cycle frequency), etc. These data dynamically correct the input parameters of the shaft system dynamic model through real-time interaction between edge nodes and the digital twin system, so that the virtual model is consistent with the operating state of the physical shaft system.

[0133] Finally, fatigue damage is calculated based on the theory of cumulative fatigue damage in materials (such as Miner's rule), as follows:

[0134] The shaft dynamics model uses fused real-time data to simulate the dynamic stress (such as alternating stress amplitude and stress cycle number) on each component (such as bearing rollers and rotor journal) under the current working conditions.

[0135] By combining the SN curve (stress-life curve) of the component material, the stress cycle is converted into the corresponding fatigue damage value (single cycle damage).

[0136] By accumulating damage values ​​at different operating stages, the total fatigue damage of each component is obtained, thereby quantifying and assessing the degree of fatigue deterioration of the component due to long-term alternating loads, providing a core basis for subsequent "prediction of shaft component life".

[0137] By deeply coupling "virtual model + real-time data", accurate and dynamic calculation of fatigue damage of shaft components is achieved, breaking through the limitations of traditional damage assessment based on experience or static models.

[0138] In one embodiment, specifically: in step four, the multi-source data fusion algorithm adopts DS evidence theory to generate a basic probability assignment matrix by establishing a mapping relationship between feature vectors and fault modes; the real-time operating conditions of the steam turbine include unit load, steam parameters, and speed;

[0139] Specifically, firstly, common failure modes of turbine shaft systems are identified (such as rotor imbalance, shaft misalignment, oil film whirl, bearing wear, etc.), and a mapping relationship library between feature vectors and these failure modes is established. By analyzing the correlation between each feature in the multi-dimensional feature vector matrix (such as the power frequency amplitude of vibration signals, the steady-state deviation of temperature signals, etc.) and different failure modes, the indicative weight of features on failures is determined (for example, excessive power frequency amplitude is strongly correlated with rotor imbalance, and abnormal second harmonic amplitude is strongly correlated with shaft misalignment).

[0140] Then, based on the DS evidence theory, a basic probability assignment matrix is ​​generated. Taking the multi-dimensional feature vector matrix as input, for each feature term in the feature vector, a basic probability value (i.e., the confidence level of a certain feature supporting a certain fault mode) is assigned according to its mapping relationship with the fault mode, forming a basic probability assignment matrix (the matrix rows correspond to feature vector samples, the columns correspond to fault modes, and the elements are the probability values ​​of the sample features supporting the corresponding fault).

[0141] Simultaneously, real-time operating data of the steam turbine is collected, including unit load (such as MW-level output), steam parameters (such as steam pressure and temperature), and speed (such as r / min), and association rules between operating conditions and feature weights are established; for example, under high load conditions, the influence weight of vibration signal features on fault diagnosis is increased; under low speed conditions, the weights of temperature and pressure signal features are increased accordingly.

[0142] Finally, by dynamically adjusting the fusion weights of data features from different types of sensors in real-time operating conditions, evidence synthesis is performed on the basic probability assignment matrix based on the adjusted weights (by fusing the probability values ​​of multiple features through the DS combination rule), eliminating conflicts between features and generating a fusion feature set that comprehensively reflects the state of the shaft system.

[0143] In one embodiment, specifically: in step two, the multi-source data includes vibration signals, temperature signals, pressure signals, and rotational speed signals;

[0144] Among them, vibration signals are collected by triaxial vibration acceleration sensors deployed at key measuring points such as bearing housings, couplings, and journals. They reflect the vibration amplitude, frequency, and dynamic response of the shaft system in the X, Y, and Z directions, and are the core basis for judging shaft system imbalance, misalignment, rubbing and other faults.

[0145] Temperature signal: mainly from the bearing temperature sensor, which collects the bearing operating temperature and is used to monitor the friction status of the shaft system (such as oil film failure, which can cause a sudden temperature rise).

[0146] Pressure signal: Collected by pressure sensors deployed in the lubricating oil pipeline, reflecting the oil supply pressure of the lubricating oil system and related to the lubrication status of the shaft (such as poor lubrication due to low pressure, which aggravates vibration).

[0147] Speed ​​signal: Collected by speed sensor, it reflects the unit's operating speed and is a key parameter for analyzing the correlation between vibration and speed (such as the direct correlation between power frequency vibration and speed).

[0148] Preprocessing includes signal conditioning and noise reduction / cleaning;

[0149] Signal conditioning converts the original analog signal into a digital signal that meets the analysis requirements through hardware or software means, adapting it to the characteristics of different signals. The specific processing steps are as follows:

[0150] Vibration signal conditioning uses a 16-bit analog-to-digital converter (ADC) to convert analog vibration signals into digital signals, and selects a sampling rate of 8-12kHz—a high sampling rate can preserve high-frequency vibration characteristics (such as high-order harmonics), and 16-bit precision ensures that the signal quantization error is extremely small, meeting the accuracy requirements for fault feature extraction.

[0151] Temperature / pressure signal conditioning uses 12-bit analog-to-digital conversion with a sampling rate of 1-2kHz. Temperature and pressure signals are slowly varying parameters, so there is no need for an excessively high sampling rate. 12-bit precision is sufficient to meet the quantization requirements of their variation range (such as temperature, which is usually between 0-100℃, and pressure, which is between 0-1MPa), thus reducing data redundancy.

[0152] Speed ​​signal conditioning uses a pulse shaping circuit to filter and shape the original speed pulse signal, remove glitches and interference, and extract the effective frequency (such as the pulse frequency corresponding to the speed) to ensure the stability of the speed data.

[0153] Noise reduction and purification are used to eliminate noise components in different signals while retaining effective features. The specific processing procedure is as follows:

[0154] Vibration signal noise reduction uses the db4 wavelet basis for 5-level decomposition to decompose the signal into different frequency components. Combined with a soft threshold function, high-frequency noise components are suppressed. The db4 wavelet basis has a good ability to capture the high-frequency abrupt changes of non-stationary vibration signals (such as impact vibration). The 5-level decomposition can separate fault-related characteristic frequencies (such as power frequency and second harmonic frequency) from irrelevant high-frequency noise.

[0155] Temperature / pressure signal noise reduction eliminates common-mode interference through differential amplifier circuits. Temperature and pressure sensors are susceptible to environmental electromagnetic interference (such as motor magnetic fields). Common-mode interference can also affect the signal input. Differential amplification improves the signal-to-noise ratio by amplifying the differential signal (real signal) and suppressing the common-mode signal (noise).

[0156] The speed signal noise reduction uses a moving average filter (such as taking the average of the most recent 5-10 sampling points) to filter out instantaneous pulse interference (such as instantaneous jumps caused by poor sensor contact), resulting in a smooth speed curve and ensuring that the speed data is consistent with the actual operating state.

[0157] In one embodiment, specifically: the method for constructing the multi-dimensional feature vector matrix in step three includes the following steps:

[0158] Step 301: For the vibration signal, extract the peak value, kurtosis, skewness, impulse index and root mean square value as time-domain statistical features;

[0159] Specifically, firstly, the vibration signal after preprocessing in step two is acquired. This signal is a discrete-time data sequence, denoted as... Where: t represents the sampling point index (t=1,2,…,N), and N represents the total number of sampling points. This represents the vibration amplitude at the t-th sampling point (unit: usually m / s). 2 (or μm);

[0160] Then, the peak value of the vibration signal is calculated by traversing the entire vibration data sequence. Extract the maximum value using the following formula:

[0161] ;

[0162] In the formula, It represents the peak value of the vibration signal, reflecting the maximum instantaneous amplitude that occurs during the vibration process;

[0163] Next, the root mean square (RMS) value is calculated: first, the vibration amplitude at each sampling point is squared; then, the average of the squared sequence is calculated; finally, the square root of the average is taken. The formula is as follows:

[0164] ;

[0165] In the formula, RMS represents the root mean square value, which reflects the energy level of the vibration signal and is directly related to the vibration intensity.

[0166] Next, calculate the pulse index: First, calculate the average absolute value of the vibration signal, using the formula:

[0167] ;

[0168] In the formula, This represents the average absolute value of the vibration signal;

[0169] Divide the peak value by the average absolute value to obtain the impulse index:

[0170] ;

[0171] This indicator is sensitive to the instantaneous impact characteristics in the signal and can highlight early impact failures (such as bearing ball spalling).

[0172] Next, calculate the skewness:

[0173] The first step is to calculate the mean value μ of the vibration signal:

[0174] ;

[0175] In the formula, μ represents the arithmetic mean of the vibration signal, reflecting the central tendency of the signal;

[0176] The second step is to calculate the third-order central moments. :

[0177] ;

[0178] In the formula, It represents the third-order central moment, reflecting the degree of asymmetry in the signal distribution;

[0179] The third step is to calculate the standard deviation σ:

[0180] ;

[0181] In the formula, σ represents the standard deviation, which reflects the degree of dispersion of the signal;

[0182] Step 4: Calculate the skewness:

[0183] ;

[0184] The further the skewness value deviates from 0, the more asymmetrical the signal distribution is, which can reflect the signal shape distortion caused by faults (such as signal skew caused by rotor friction).

[0185] Finally, calculate the kurtosis:

[0186] The first step is to calculate the fourth-order central moments. :

[0187] ;

[0188] In the formula, It represents the fourth-order central moment, reflecting the density of spike pulses in the signal;

[0189] The second step is to calculate the kurtosis:

[0190] ;

[0191] The kurtosis of a normal Gaussian distribution signal is 3. When the kurtosis value is greater than 3, it indicates that there are many spike pulses in the signal, which are highly sensitive to early fatigue, wear and other faults (such as the small impacts in the early stage of shaft cracks).

[0192] The peak value obtained from the above calculation By integrating kurtosis, skewness, impulse index, and root mean square (RMS) value, the time-domain statistical characteristics of the vibration signal are obtained.

[0193] Step 302: Extract the mean, range, rate of change, fluctuation amplitude, and steady-state deviation as time-domain features of the temperature, pressure, and speed signals.

[0194] Specifically, first, the temperature signal, pressure signal, and rotational speed signal, which have been preprocessed in step two, are acquired. These three types of signals are all time-domain discrete data sequences, denoted as follows:

[0195] Temperature signal: (t=1,2,…,N), where t is the index of the sampling point and N is the total number of sampling points. This represents the temperature value (in °C) at the t-th sampling point.

[0196] Pressure signal: (t=1,2,…,N), where, This represents the pressure value at the t-th sampling point (unit: MPa).

[0197] Speed ​​signal: (t=1,2,…,N), where, This represents the rotational speed value at the t-th sampling point (unit: r / min);

[0198] Then, the time-domain features of the parameters are extracted for the above three types of signals, specifically including:

[0199] Calculate the mean: The mean reflects the average level of the signal during the sampling period. The formula is standardized as follows:

[0200] Average temperature signal: ;

[0201] Average pressure signal: ;

[0202] Average speed signal: ;

[0203] Calculate the range: The range reflects the maximum fluctuation range of the signal within the sampling period, that is, the difference between the maximum and minimum values. The formula is:

[0204] Temperature signal extreme difference: ;

[0205] in, This represents the maximum temperature. This is the minimum temperature value;

[0206] Extremely poor pressure signal: ;

[0207] in, This represents the maximum pressure. This represents the minimum pressure value;

[0208] Extremely poor speed signal: ;

[0209] in, This is the maximum rotational speed. This is the minimum rotational speed;

[0210] Calculate the rate of change: The rate of change reflects the average speed at which the signal changes over time. It is calculated by the difference between adjacent sampling points, using the following formula:

[0211] Rate of change of temperature signal: ;

[0212] in, This represents the absolute difference between adjacent temperature sampling points;

[0213] Pressure signal change rate: ;

[0214] in, This represents the absolute difference between adjacent pressure sampling points;

[0215] Rate of change of rotational speed signal: ;

[0216] in, This represents the absolute difference between adjacent speed sampling points;

[0217] Calculate the fluctuation amplitude: The fluctuation amplitude reflects the maximum deviation of the signal from its mean. The formula is:

[0218] Temperature signal fluctuation amplitude: ;

[0219] That is, the larger of the absolute values ​​of the differences between the maximum temperature, minimum temperature, and mean temperature;

[0220] Pressure signal fluctuation amplitude: ;

[0221] That is, the larger of the absolute values ​​of the differences between the maximum, minimum and mean pressure values;

[0222] Speed ​​signal fluctuation amplitude: ;

[0223] That is, the larger of the absolute values ​​of the differences between the maximum, minimum and average speeds;

[0224] Calculate steady-state deviation: Steady-state deviation reflects the deviation of the signal from the equipment's health baseline (preset normal steady-state value) during steady-state operation. The specific steps are as follows:

[0225] Step 1: Determine the steady-state range of the signal (by determining if the signal fluctuation amplitude A is less than or equal to a steady-state threshold, such as setting the steady-state threshold for temperature to 2℃), and extract the data sequence within the steady-state range. (temperature), (pressure), (Rotation speed), where s is the index of the steady-state sampling point;

[0226] Step 2: Calculate the mean value over the steady-state interval:

[0227] ;

[0228] Where M is the number of steady-state sampling points; similarly, we get , ;

[0229] Step 3: Integrate with the equipment health baseline (temperature baseline) Pressure baseline Speed ​​baseline ), calculate the deviation:

[0230] Temperature steady-state deviation: ;

[0231] Pressure steady-state deviation: ;

[0232] Steady-state deviation of rotational speed: ;

[0233] Finally, the mean obtained from the above calculations is... Range rate of change Fluctuation Amplitude and steady-state deviation They are integrated separately and used as time-domain features of the temperature, pressure, and speed signals.

[0234] Step 303: Perform frequency domain transformation on the vibration signal, obtain the frequency domain spectrum through Fourier transform, and extract the amplitude and phase of the power frequency, second harmonic, 0.5 harmonic and higher harmonics to form frequency domain fault characteristics;

[0235] Specifically, firstly, the vibration signal after preprocessing in step two is acquired. This signal is a discrete-time data sequence, denoted as... (t=1,2,…,N), where t is the index of the sampling point and N is the total number of sampling points. The vibration amplitude at the t-th sampling point (unit: m / s) 2 (or μm), while recording the sampling frequency. (Unit: Hz, i.e., number of sampling points per second);

[0236] Then, frequency domain transformation is performed on the vibration signal: the time domain signal is transformed using Fast Fourier Transform (FFT). Convert to frequency domain complex sequence (k=0,1,…,N-1), the formula is:

[0237] ;

[0238] In the formula, j is the imaginary unit (j 2 =-1), k is the frequency domain index. The complex amplitude at the k-th frequency point includes amplitude and phase information; calculated using the formula:

[0239] ;

[0240] The amplitude spectrum (reflecting the vibration intensity of each frequency component) is obtained.

[0241] Calculation formula:

[0242] ;

[0243] The phase spectrum (reflecting the vibration initiation phase of each frequency component) is obtained, and a frequency domain spectrum is constructed by combining it with the frequency axis;

[0244] Next, determine the target frequency:

[0245] Power frequency f0: The corresponding unit rotation frequency, which is calculated from the speed signal. The formula is f0=n / 60 (n is the unit speed, unit: r / min, f0 unit: Hz).

[0246] 2 times the frequency f2 = 2f0, 0.5 times the frequency f 0.5 =0.5f0, higher harmonics (such as 3rd harmonic f3=3f0, 4th harmonic f4=4f0, etc., select typical higher harmonics according to actual fault characteristics);

[0247] Next, extract the amplitude and phase of the target frequency: according to the frequency axis scale (frequency resolution Δf=f s / N, the k-th frequency point corresponds to the frequency f k =k×Δf, located in the frequency domain spectrum at f0, f2, f 0.5 and the closest frequencies of higher harmonics k0, k2, k 0.5 k m (where m is the order of higher harmonics), extract the corresponding amplitudes |X(k0)|, |X(k2)|, and |X(kk3)|. 0.5 )|、|X(k m )| and phase φ(k0), φ(k2), φ(k 0.5 ), φ(k m );

[0248] Finally, the extracted amplitudes and phases of the main frequency, second harmonic, 0.5 harmonic, and higher harmonics are integrated to form frequency domain fault characteristics (for example, abnormal main frequency amplitude corresponds to rotor imbalance, increased second harmonic amplitude corresponds to shaft misalignment, and 0.5 harmonic is associated with faults such as oil film whirl).

[0249] Step 304: Concatenate the time-domain statistical features, parameter time-domain features, and frequency-domain fault features in an orderly manner according to the signal source and feature attributes to construct a multi-dimensional feature vector matrix. The row dimension of the matrix corresponds to the sampling time sequence, and the column dimension corresponds to the feature category.

[0250] Specifically, firstly, the time-domain statistical features of the vibration signal extracted in step 301, the time-domain features of the temperature / pressure / speed parameters extracted in step 302, and the frequency-domain fault features of the vibration signal extracted in step 303 are classified and organized:

[0251] Based on the "signal source", it is divided into four categories: vibration signal characteristics (including time-domain statistical characteristics and frequency-domain fault characteristics), temperature signal characteristics (parameter time-domain characteristics), pressure signal characteristics (parameter time-domain characteristics), and rotational speed signal characteristics (parameter time-domain characteristics).

[0252] Subdivided by "characteristic attributes": In vibration signal characteristics, time-domain statistical characteristics (peak value, kurtosis, etc.) and frequency-domain fault characteristics (power frequency amplitude / phase, etc.) are separated; temperature / pressure / speed characteristics are all parametric time-domain characteristics (mean, range, etc.).

[0253] Then, determine the feature concatenation order: sort by "signal source priority + feature attribute hierarchy", specifically:

[0254] Vibration signal characteristics: First, time-domain statistical characteristics (5 items in step 301), then frequency-domain fault characteristics (amplitude and phase of power frequency, second harmonic, 0.5 harmonic and higher harmonics in step 303, each frequency item contains 2 characteristics, and a total of 8 items are set for 4 frequencies).

[0255] Temperature signal characteristics: Time-domain characteristics of the five parameters in step 302 (mean, range, etc.);

[0256] Pressure signal characteristics: Time-domain characteristics of the five parameters in step 302;

[0257] Rotational speed signal characteristics: Time-domain characteristics of the five parameters in step 302;

[0258] Next, construct a matrix row vector according to the sampling time sequence: for each sampling time t (t=1,2,…,N), extract all the above-mentioned feature values ​​corresponding to that time, and concatenate them into a row vector in a preset order; for example, the row vector at time t is:

[0259] ;

[0260] Finally, the row vectors of all sampling times are arranged in chronological order to form a multi-dimensional feature vector matrix F: the row dimension of the matrix is ​​the number of sampling times N (each row corresponds to 1 sampling time), the column dimension is the total number of features (the sum of all the feature terms mentioned above), and the matrix elements... Let i represent the i-th eigenvalue at time t, and complete the matrix construction.

[0261] In one embodiment, specifically: in step five, the lightweight model is a neural network based on deep separable convolution, which has the functions of fault type identification, confidence output and vibration trend prediction.

[0262] Specifically, the lightweight model addresses the limited computing power of edge nodes by using depthwise separable convolutions instead of traditional convolution operations. This significantly reduces the parameter size and computational load while maintaining feature extraction capabilities, thus achieving "lightweight" performance.

[0263] Depthwise separable convolution principle: It breaks down traditional convolution (which simultaneously performs channel fusion and spatial feature extraction) into two steps:

[0264] Depthwise convolution: Convolution is performed on each channel of the input feature map individually (each channel corresponds to 1 convolution kernel), extracting only the spatial features within a single channel;

[0265] Pointwise convolution: Uses a 1×1 convolution kernel to perform cross-channel fusion on the output of depthwise convolution, integrating features from different channels;

[0266] Compared to traditional convolution, this structure can reduce the computational cost to 1 / (number of input channels) of the original, significantly reduce the parameter size, and adapt to the real-time inference needs of edge nodes;

[0267] The lightweight model's fault type identification function can accurately identify common shaft system faults (such as rotor imbalance, shaft misalignment, oil film whirl, bearing wear, etc.); the specific implementation method is as follows:

[0268] The lightweight model input receives the fusion feature set (including multi-dimensional features such as vibration, temperature, and pressure) generated in step four.

[0269] After extracting high-order features step by step through depthwise separable convolutional layers and pooling layers, the features are connected to the classification output layer through a fully connected layer.

[0270] The classification output layer uses the softmax activation function to output the probability distribution of each fault type (such as "imbalance" probability 0.92, "misalignment" probability 0.05, etc.). The type with the highest probability is the diagnosed fault type.

[0271] The confidence output function of the lightweight model is used to quantify the reliability of the model's fault type identification results, and to assist in judging the reliability of the diagnosis; the specific implementation method is as follows:

[0272] The softmax probability value of the classification output layer is directly used, and the probability value corresponding to the fault type identification result is used as the confidence level (such as the 0.92 of the "imbalance" mentioned above, which is the confidence level of the diagnosis result).

[0273] A higher confidence level (usually ≥0.7) indicates that the model is more reliable in identifying the fault type, providing a basis for subsequent graded early warning (low confidence results may trigger manual review).

[0274] The lightweight model's vibration trend prediction function is used to predict the vibration amplitude change trend in the short term (e.g., 5-10 minutes) based on current and historical fusion characteristics, thus identifying potential risks in advance. The specific implementation method is as follows:

[0275] After feature extraction, the model connects to the regression output layer through a branch network (which shares the preceding feature extraction layer with fault identification to reduce redundant computation);

[0276] The regression output layer uses a linear activation function and combines the temporal trends in the fusion features (such as the historical rate of change of vibration amplitude, working condition stability, etc.) to output vibration prediction values ​​for multiple future moments (such as vibration amplitude at the 1st minute and the 5th minute in the future), providing forward-looking data for dynamic threshold adjustment and early warning triggering.

[0277] This lightweight model achieves efficient feature processing through depthwise separable convolution, while integrating classification and regression branches to complete fault diagnosis, reliability quantification, and trend prediction locally in real time at edge nodes, meeting the real-time and accuracy requirements of shaft vibration monitoring.

[0278] In one embodiment, specifically: in step six, the dynamic threshold adjustment mechanism is corrected by combining the equipment health baseline with the real-time operating condition deviation; the graded early warning is divided into four levels according to the vibration prediction value exceeding the limit and the confidence level, and the four levels correspond to different response measures.

[0279] Specifically, first, a health baseline for the equipment is established by extracting multi-source data (steady-state values ​​of vibration, temperature, pressure, and rotational speed) from steps one through four during the equipment's historical normal operation periods. For example:

[0280] Vibration amplitude baseline V0: Take the average root mean square value of vibration during normal operation (combined with the 3σ principle, covering 99.7% of normal fluctuations).

[0281] Temperature baseline T0: steady-state average temperature of the bearing bush (e.g., if the normal operating temperature range of the bearing is 25~50℃, the baseline is 35℃).

[0282] By using statistical clustering (such as K-means to divide historical data into "normal" and "abnormal" clusters) or physical modeling (such as the linear relationship between bearing temperature and load), the normal range benchmarks for each parameter are determined as initial references for the thresholds.

[0283] Then, real-time operating conditions are obtained and deviations are calculated: real-time operating parameters of the steam turbine (including unit load, steam parameters, speed, etc.) are collected, and the deviations from the rated operating conditions (such as speed deviation, load deviation, etc.) are calculated to quantify the degree of operating condition fluctuations.

[0284] Next, dynamic thresholds are generated based on the deviation correction baseline: a correlation model between operating condition deviation and threshold is established, and the real-time operating condition deviation is substituted into the model through the correction coefficient (obtained by fitting historical data) to dynamically correct the equipment health baseline and obtain dynamic thresholds (such as vibration dynamic threshold, temperature dynamic threshold, etc.) that are adapted to the current operating conditions.

[0285] Next, the classification dimensions for graded early warning were determined: the excess amplitude (deviation percentage) of the vibration prediction value (output of step five) relative to the dynamic threshold and the fault identification confidence (probability value output of step five) were used as the core classification criteria.

[0286] Next, four warning levels are defined: the level is determined based on the range of the exceedance and the confidence level. For example: Level 1 warning corresponds to exceedance ≤ 5% and confidence level ≥ 90%; Level 2 warning corresponds to 5% < exceedance ≤ 15% and 80% ≤ confidence level < 90%; Level 3 warning corresponds to 15% < exceedance ≤ 30% and 70% ≤ confidence level < 80%; Level 4 warning corresponds to exceedance > 30% or confidence level < 70% (rapidly rising trend takes priority).

[0287] Finally, implement the corresponding response measures for each level: Level 1 warnings mainly involve local recording and monitoring; Level 2 warnings trigger remote manual inspections; Level 3 warnings push emergency information and reduce load during operation; and Level 4 warnings initiate mandatory emergency shutdowns to ensure that appropriate response actions are taken according to the risk level.

[0288] In one embodiment, specifically: in step seven, the log accumulation sample library is stored according to fault type and operating condition; the lightweight model deployed on the edge node is updated using an incremental learning approach, optimizing local parameters only for newly added samples;

[0289] Specifically, first, the raw data to be stored in the sample library is collected, including the hierarchical early warning information generated in step six (including early warning level, fault type, confidence level, etc.) and the operation logs uploaded by edge nodes (covering multi-source monitoring data, feature vectors, real-time operating parameters, etc.).

[0290] Then, the collected data is classified and stored according to two dimensions: one is classified according to the type of fault (such as rotor imbalance, shaft misalignment, oil film whirl, etc.), and the other is classified according to real-time operating conditions (such as unit load range, speed range, steam parameter level, etc.) to ensure the structure and searchability of the data in the sample library.

[0291] Next, identify new samples in the sample library: compare with historical samples, and select new fault cases that were not included in the model training, and samples with significant differences in data and feature distribution under new working conditions, as input for model update;

[0292] Then, the lightweight model is updated using incremental learning: only for newly added sample data, the local parameters of the model (such as the weights of the fully connected layers of the neural network, the late feature fusion parameters of the depthwise separable convolutional layers, etc.) are fine-tuned and optimized, rather than the entire model is retrained.

[0293] Finally, the model is updated and deployed: the optimized local parameters are synchronized to the lightweight model at the edge nodes to ensure that the model can adapt to new failure modes and changes in operating conditions while retaining historical learning results, thereby improving the accuracy of diagnosis and prediction.

[0294] In one embodiment, specifically: in step one, the multiple types of sensors include a triaxial vibration acceleration sensor, a bearing temperature sensor, a lubricating oil pressure sensor, and a speed sensor;

[0295] Among them, the triaxial vibration acceleration sensor is used to monitor the vibration characteristics of the shaft system;

[0296] The bearing temperature sensor is used to capture the frictional heating state of the bearing.

[0297] Lubricating oil pressure sensors are used to monitor the oil supply pressure of the lubrication system;

[0298] The speed sensor is used to record the operating speed of the unit;

[0299] Key measuring points include the turbine's bearing housing, coupling, journal, and lubrication oil lines;

[0300] Among them, the bearing housing is a key node for vibration transmission, which can reflect the dynamic interaction between the rotor and the bearing.

[0301] Couplings are the weakest link in shaft systems, and their condition is related to the alignment accuracy of the shaft system.

[0302] The journal is in direct contact with the rotor, which can reflect the surface condition and stress of the rotor;

[0303] The lubricating oil pipeline is the key path of the lubrication system, and its pressure changes directly affect the lubrication effect of the shaft system;

[0304] At least one triaxial vibration accelerometer is deployed in each bearing housing along the X, Y, and Z directions to form a three-dimensional spatial monitoring array. This deployment method can comprehensively capture the vibration components of the shaft system in the radial (horizontal and vertical) and axial directions, avoiding the limitations of single-direction monitoring. It ensures that the characteristic differences of faults such as imbalance, misalignment, and rubbing in different directions can be fully extracted, providing complementary spatial information for subsequent multi-source data fusion and fault diagnosis.

[0305] In one embodiment, specifically: edge nodes and the cloud transmit data via an encrypted protocol, edge nodes perform real-time diagnostics, and the cloud performs offline analysis and lightweight model updates;

[0306] Specifically, firstly, edge nodes and the cloud establish a secure data transmission channel through encryption protocols (such as SSL / TLS protocols) to ensure the confidentiality and integrity of hierarchical early warning information, operation logs, and feature data during transmission, and to prevent data leakage or tampering.

[0307] Then, relying on the deployed lightweight model, the edge nodes quickly analyze the multi-source data collected in real time, and immediately output the fault type, confidence level and vibration prediction value to achieve local real-time diagnosis and meet the real-time requirements of turbine shaft vibration monitoring.

[0308] Finally, after receiving the data uploaded by the edge nodes, the cloud performs offline in-depth analysis (such as historical data trend mining, fault mode clustering, cross-unit data comparison, etc.), and optimizes the lightweight model parameters based on the accumulated log sample library. Then, the updated model is synchronized to the edge nodes to realize model iteration and upgrade, and improve diagnostic generalization ability.

[0309] Example 2:

[0310] This embodiment provides a turbine shaft vibration intelligent diagnosis and edge early warning device based on multi-source data fusion, including a processor and a storage medium;

[0311] The storage medium is used to store instructions;

[0312] The processor is configured to operate according to the instructions to perform the steps of the method according to Embodiment 1.

[0313] Example 3:

[0314] This embodiment provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method described in Embodiment 1.

[0315] In summary, the present invention provides a method for intelligent diagnosis and edge warning of turbine shaft vibration based on multi-source data fusion. By deploying multiple types of sensors to collect multi-source data such as shaft vibration, temperature, pressure, and speed, and after preprocessing and feature extraction, a multi-source data fusion algorithm is used to generate a fused feature set, which is input into a lightweight model of edge nodes to achieve real-time fault diagnosis and vibration prediction. Combined with a dynamic threshold adjustment mechanism, a graded warning is triggered, and offline analysis and model updates are performed in conjunction with the cloud, thereby realizing intelligent diagnosis and edge warning of turbine shaft vibration.

[0316] Those skilled in the art will understand that the features described in the various embodiments and / or claims of the present invention can be combined or combined in various ways, even if such combinations or combinations are not explicitly described in the present invention. In particular, the features described in the various embodiments and / or claims of the present invention can be combined or combined in various ways without departing from the spirit and teachings of the present invention. All such combinations and / or combinations fall within the scope of the present invention.

Claims

1. A method for intelligent diagnosis and edge warning of turbine shaft system vibration based on multi-source data fusion, characterized in that, Includes the following steps: Multiple types of sensors are deployed at key measuring points in the turbine shaft system to construct a monitoring network with complementary spatial dimensions; A monitoring network based on complementary spatial dimensions collects multi-source data of the turbine shaft system and preprocesses the collected multi-source data. Extract time-series trend features and frequency domain fault features from the preprocessed multi-source data, and construct a multi-dimensional feature vector matrix; Based on a multi-dimensional feature vector matrix, a multi-source data fusion algorithm based on feature layer fusion is used for feature fusion. At the same time, the fusion weights of different types of sensor data are dynamically adjusted in combination with the real-time operating conditions of the steam turbine to generate a fused feature set. A lightweight model that integrates feature sets is deployed at edge nodes and outputs fault type, confidence level, and vibration prediction value in real time. Based on fault type, confidence level, and vibration prediction value, combined with a dynamic threshold adjustment mechanism, the data distribution is monitored to trigger graded early warnings and simultaneously correlate with cloud analysis. Based on the tiered early warning information and the accumulated sample library of logs uploaded by edge nodes, the lightweight model deployed on the edge nodes is updated, and the lifespan of turbine shaft system components is predicted by combining digital twin analysis.

2. The intelligent diagnosis and edge warning method for turbine shaft vibration based on multi-source data fusion according to claim 1, characterized in that: The digital twin analysis is based on the shaft dynamics model and integrates real-time monitoring data to calculate fatigue damage of turbine shaft components.

3. The intelligent diagnosis and edge warning method for turbine shaft vibration based on multi-source data fusion according to claim 1, characterized in that: The multi-source data fusion algorithm adopts the DS evidence theory and generates a basic probability assignment matrix by establishing a mapping relationship between feature vectors and fault modes; the real-time operating conditions of the steam turbine include unit load, steam parameters and speed.

4. The intelligent diagnosis and edge warning method for turbine shaft vibration based on multi-source data fusion according to claim 1, characterized in that: The multi-source data includes vibration signals, temperature signals, pressure signals, and rotational speed signals; the preprocessing includes signal conditioning and noise reduction.

5. The intelligent diagnosis and edge warning method for turbine shaft vibration based on multi-source data fusion according to claim 4, characterized in that: The method for constructing a multi-dimensional feature vector matrix includes the following steps: Step 301: For the vibration signal, extract the peak value, kurtosis, skewness, impulse index and root mean square value as time-domain statistical features; Step 302: Extract the mean, range, rate of change, fluctuation amplitude, and steady-state deviation as time-domain features of the temperature, pressure, and speed signals. Step 303: Perform frequency domain transformation on the vibration signal, obtain the frequency domain spectrum through Fourier transform, and extract the amplitude and phase of the power frequency, second harmonic, 0.5 harmonic and higher harmonics to form frequency domain fault characteristics; Step 304: Concatenate the time-domain statistical features, parameter time-domain features, and frequency-domain fault features in an orderly manner according to the signal source and feature attributes to construct a multi-dimensional feature vector matrix. The row dimension of the matrix corresponds to the sampling time sequence, and the column dimension corresponds to the feature category.

6. The intelligent diagnosis and edge warning method for turbine shaft vibration based on multi-source data fusion according to claim 1, characterized in that: The architecture of the lightweight model includes an input layer, a feature extraction layer, a multi-task fusion layer, and a multi-branch output layer. Input layer: Used to standardize the input fused feature set to obtain a 32-dimensional standardized feature vector; Feature extraction layer: used to perform high-order feature extraction and dimensionality compression on the 32-dimensional standardized feature vector to obtain a 1D high-order feature map; Multi-task fusion layer: used to perform channel compression and feature integration processing on the 1D high-order feature map to obtain a 32-dimensional shared feature vector; Multi-branch output layer: including fault classification branch, confidence branch and vibration prediction branch; The fault classification branch is used to process the 32-dimensional shared feature vector using a multi-task loss function to obtain a 5-dimensional probability vector. The confidence branch reuses the output of the fault classification branch and takes the probability value corresponding to the fault category with the highest probability in the results as the confidence level. The vibration prediction branch includes an LSTM layer, a fully connected layer, and a linear activation function, which are used to obtain a vibration prediction value representing the future trend of vibration amplitude change based on the 32-dimensional shared feature vector.

7. The intelligent diagnosis and edge warning method for turbine shaft vibration based on multi-source data fusion according to claim 6, characterized in that: The multi-branch model is trained using a multi-task joint training mode. By integrating classification loss and regression loss, it optimizes all three branches simultaneously during a single training session. The total model loss is calculated by weighted summation of the fault classification loss and vibration prediction loss, as shown in the following formula: Total loss = α × classification loss + β × regression loss; The classification loss uses the cross-entropy loss function; The regression loss is the mean squared error loss; Weight settings: α=0.6, β=0.

4.

8. The intelligent diagnosis and edge warning method for turbine shaft vibration based on multi-source data fusion according to claim 1, characterized in that: The dynamic threshold adjustment mechanism is corrected by combining the equipment health baseline with the real-time operating condition deviation; the graded early warning is divided into four levels according to the vibration prediction value exceeding the limit and the confidence level, and the four levels correspond to different response measures.

9. The intelligent diagnosis and edge warning method for turbine shaft vibration based on multi-source data fusion according to claim 1, characterized in that: The log accumulation sample library is stored according to fault type and operating condition; the lightweight model deployed on the updated edge nodes adopts an incremental learning method, and only optimizes local parameters for newly added samples. The edge nodes transmit data to the cloud via an encrypted protocol. The edge nodes perform real-time diagnostics, while the cloud performs offline analysis and lightweight model updates.

10. The intelligent diagnosis and edge warning method for turbine shaft vibration based on multi-source data fusion according to claim 1, characterized in that: The various types of sensors include a triaxial vibration acceleration sensor, a bearing temperature sensor, a lubricating oil pressure sensor, and a speed sensor; the key measuring points include the turbine's bearing housing, coupling, journal, and lubricating oil pipeline; at least one triaxial vibration acceleration sensor is deployed in each bearing housing along the X, Y, and Z directions to form a spatial three-dimensional monitoring array.

Images