A machine learning based rotor tip clearance data analysis method

Abstract

The application provides a rotor tip clearance data analysis method based on machine learning, belonging to the technical field of rotor state monitoring. Time series operation data of rotor tip clearance, rotating speed and other working condition parameters are collected, a mechanism simplified model is constructed relying on rotor structure and material thermal physical property parameters, and an ideal clearance change base value caused by centrifugal force and thermal deformation is accurately calculated. Taking working condition parameters and ideal clearance base value as input, clearance deviation residual is fitted through residual learning branch, and tip clearance prediction value is obtained by superimposing the base value. During model training, only the residual learning branch is optimized. After online deployment of the model, ideal clearance base value and deviation residual can be decoupled and output, dynamic control limit is set through residual statistical characteristics, and rotor structure fault is determined according to residual amplitude and trend anomaly, so that working condition transient fluctuation and rotor structure anomaly can be effectively distinguished, the problems of high false alarm rate and poor interpretability of traditional monitoring are solved, and the monitoring and fault early warning precision of rotor tip clearance is improved.

Description

Technical Field

[0001] This invention belongs to the field of data analysis technology of machine learning, and specifically relates to a method for analyzing rotor blade tip clearance data based on machine learning. Background Technology

[0002] Rotor blade tip clearance is a key parameter for assessing the operational health of rotating machinery such as aero-engines and gas turbines. Insufficient clearance can easily lead to friction between the blade tip and the casing, or even blade breakage, while excessive clearance significantly reduces aerodynamic efficiency and increases fuel consumption. Currently, online blade tip clearance monitoring systems based on eddy current, capacitance, or fiber optic sensors are widely used in industry. Their data analysis methods are mainly divided into two categories: one is the fixed threshold alarm method, which sets a static clearance safety threshold and triggers an alarm when the measured clearance exceeds the threshold; the other is the trend analysis method based on simple statistical regression, which fits historical data to empirical curves showing the change of clearance with single factors such as rotational speed and temperature, thereby determining whether the current clearance deviates from expectations.

[0003] However, the above methods face a problem in practical engineering applications: under complex transient and steady-state conditions involving variable rotor speed, load, and thermodynamic state, the blade tip clearance itself exhibits large and nonlinear dynamic fluctuations due to the coupling effects of centrifugal force, thermal expansion, and aerodynamic loads. Existing fixed threshold or single-factor empirical curve models cannot accurately establish the normal clearance benchmark that should be presented under the current specific operating condition combination in real time, resulting in a large number of normal operating condition fluctuations being misjudged as abnormal (high false alarm rate), while the subtle clearance changes caused by early faults such as blade wear, coating peeling, and uneven thermal deformation are difficult to detect in time because they are submerged in the operating condition fluctuations (high false alarm rate). Summary of the Invention In view of the above-mentioned defects or deficiencies in the prior art, a machine learning-based method for rotor blade tip clearance data analysis is provided, comprising the following steps: Obtain the rotor's historical operating dataset, which includes tip clearance time data, rotational speed time data, and time data of at least one other operating condition parameter; Based on the rotor's structural dimensions and material thermophysical parameters, a simplified mechanism model of the rotor is established. Using the simplified mechanism model, the baseline value of the ideal gap change caused by centrifugal force deformation and thermal deformation is calculated in real time according to the rotational speed and other operating parameters. A physics-data hybrid driven machine learning model is constructed, wherein the machine learning model receives the rotational speed, the other operating condition parameters, and the ideal clearance variation baseline value as input, and fits the clearance deviation residuals in the historical operating dataset other than the ideal clearance variation baseline value through a residual learning branch, and adds the ideal clearance variation baseline value to the fitted clearance deviation residuals to obtain the blade tip clearance prediction value. The machine learning model is trained using the historical running dataset. During the training process, the internal parameters of the simplified mechanism model are kept fixed or only subject to restrictive fine-tuning. The parameters of the residual learning branch are updated with the goal of minimizing the error between the predicted tip clearance and the historical measured tip clearance. The trained machine learning model is deployed to the online monitoring system. Rotation speed and operating parameters are input to the machine learning model in real time to obtain the predicted value of the blade tip clearance online. The ideal clearance change baseline value and the clearance deviation residual are decoupled and output simultaneously. The clearance deviation residual is continuously monitored, and a dynamic control limit is set based on the statistical characteristics of the clearance deviation residual under normal conditions. When the amplitude or change trend of the clearance deviation residual exceeds the dynamic control limit, a structural abnormality warning signal is generated to indicate the change in rotor structural state, so as to distinguish it from normal clearance fluctuations caused by transient changes in operating conditions.

[0004] According to the technical solution provided in this application, before training the machine learning model, the time series data of the decoupled gap deviation residual is subjected to signal transformation processing to construct an enhanced residual feature set. The signal transformation processing includes: taking the time-series data of the gap deviation residual as a time-domain signal, converting it into an order spectrum with rotational speed as the horizontal axis based on the synchronously acquired key phase signal, and / or converting the time-series data of the gap deviation residual into a time-spectrum graph. The feature representation of the transformed order spectrum or time spectrum is used as one of the input features of the residual learning branch.

[0005] According to the technical solution provided in this application, the residual learning branch is implemented using a dual-channel neural network, including the following steps: Construct a first channel, input the rotational speed, the other operating parameters, and the ideal clearance variation base value into the first channel, and extract the operating feature vector; Construct a second channel, which is a convolutional neural network. Input the feature representation of the order spectrum or time spectrum of the gap deviation residual into the second channel, and extract the residual spectrum feature vector. The fusion module fuses the operating condition feature vector with the residual spectrum feature vector, and outputs the predicted value of the gap deviation residual based on the fused features.

[0006] According to the technical solution provided in this application, generating a structural anomaly early warning signal to indicate a change in the rotor structural state includes the following steps: Statistical analysis is performed on the gap deviation residuals output by online decoupling, and dynamic control limits are constructed based on the mean and standard deviation of the gap deviation residuals under historical normal conditions. When the instantaneous amplitude of the gap deviation residual exceeds the dynamic control limit, or the trend slope within its sliding window exceeds a preset threshold, the structural anomaly warning signal is generated. The time-domain and frequency-domain patterns of the gap deviation residuals are identified, and the identified patterns are compared with a preset fault residual pattern library to provide an indication of the associated potential fault type while generating the structural anomaly warning signal.

[0007] According to the technical solution provided in this application, the method further includes a step of quantifying the uncertainty of the predicted blade tip clearance value: Obtain the ideal gap variation base value output by the simplified mechanism model, and calculate the prior distribution of the ideal gap variation base value based on the perturbation range of the material thermophysical parameters of the rotor; The residual learning branch outputs the predicted value of the gap deviation residual, and at the same time outputs the residual prediction variance of the predicted value; The prior distribution of the ideal gap variation baseline value is fused with the predicted value of the gap deviation residual and its predicted variance using Bayesian fusion to obtain the posterior distribution of the predicted blade tip gap value. Wherein, the mean of the posterior distribution is the final predicted tip clearance value, and the variance interval of the posterior distribution is the prediction uncertainty interval.

[0008] According to the technical solution provided in this application, generating a structural anomaly early warning signal to indicate a change in the rotor structural state includes the following steps: Obtain the predicted value of the gap deviation residual and the corresponding portion of the residual in the prediction uncertainty interval; The degree to which the gap deviation residual obtained from online decoupling deviates from its predicted value is calculated and used as the standardized residual deviation. Based on the magnitude of the residual prediction variance, the adaptive warning threshold for the current time is dynamically calculated and set; wherein, the larger the residual prediction variance, the higher the adaptive warning threshold is set. The standardized residual deviation is compared with the adaptive early warning threshold. If the standardized residual deviation continues to exceed the adaptive early warning threshold, a structural anomaly early warning signal based on uncertainty perception is generated.

[0009] According to the technical solution provided in this application, the method further includes the following steps: Retrieve the fault feature database, which contains a mapping relationship between multiple preset fault modes and a set of standard fault feature vectors. The standard fault feature vector corresponding to each fault mode is described by the time domain statistical features and frequency domain features of the gap deviation residual under the corresponding fault state. For real-time data, extract the real-time residual feature vector from the gap deviation residual output by online decoupling; The real-time residual feature vector is matched with all standard fault feature vectors in the fault feature database, and its static matching degree score with each fault mode is calculated. Based on the static matching score and the prediction uncertainty interval, a multi-dimensional fault contribution vector is calculated, where each dimension corresponds to a probability score of a fault mode. By combining the fault contribution vector with the structural anomaly warning signal, a diagnostic report is generated that includes the specific suspected fault mode, severity, and confidence level.

[0010] According to the technical solution provided in this application, some fault modes in the fault feature database have corresponding time-series evolution prototype vectors, which describe the typical pattern of the evolution of the characteristics of the gap deviation residual under the corresponding fault mode over time.

[0011] According to the technical solution provided in this application, after calculating the static matching score with each failure mode, the following steps are also included: Based on the predicted uncertainty interval, the confidence coefficient of the static matching score is calculated; The calculation of a multi-dimensional fault contribution vector based on the static matching score and the prediction uncertainty interval includes the following steps: If the credibility coefficient is greater than or equal to the first preset threshold, a multi-dimensional fault contribution vector is calculated based on the static matching score and the prediction uncertainty interval.

[0012] According to the technical solution provided in this application, after calculating the credibility coefficient of the static matching score, the method further includes the following steps: If the confidence coefficient is less than the first preset threshold, then within a continuous monitoring time window, multiple sets of the real-time residual feature vectors are extracted and cached in sequence to form a residual feature evolution sequence. The residual feature evolution sequence is matched with the time-series evolution prototype vector in the fault feature database to calculate the dynamic matching degree score. The dynamic matching score and the static matching score are fused to update the fault contribution vector, and the diagnostic report is regenerated based on the updated fault contribution vector.

[0013] Compared with the prior art, the beneficial effects of this application are as follows: I. Significantly Reduced False Alarm Rate. By establishing a simplified rotor mechanism model, the baseline value of ideal clearance variation caused by centrifugal force deformation and thermal deformation is calculated in real time. Based on this, the total clearance, including fluctuations under normal operating conditions, is decoupled into a mechanistically explainable normal component and the deviation residual. The monitoring object changes from the total clearance dominated by operating conditions to the clearance deviation residual after removing normal fluctuations, fundamentally avoiding the problem of large dynamic changes in normal clearance under varying operating conditions being misjudged as abnormal, thus significantly reducing the false alarm rate. II. Effectively reduce the false negative rate. By explicitly separating the ideal gap change baseline, the subtle gap changes caused by early faults such as blade wear, coating peeling, and uneven thermal deformation are independently extracted and focused for analysis in the form of gap deviation residuals. These residuals are no longer submerged in large operating condition fluctuations, thereby significantly improving the sensitivity and timeliness of capturing early structural anomalies and solving the problem of high false negative rates.

[0014] III. Achieving Physically Interpretable Gap Monitoring. A physics-data hybrid driven machine learning model is constructed. The simplified mechanism model provides a baseline value for gap changes with clear physical meaning, and the residual learning branch only needs to fit the deviation that the mechanism model cannot explain. The model structure itself conforms to the physical laws of rotor gap changes. Compared with pure black-box data-driven methods, the prediction results have stronger physical interpretability and extrapolation generalization ability. Attached Figure Description

[0015] Other features, objects, and advantages of this application will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings: Figure 1 A flowchart illustrating the steps of the machine learning-based rotor blade tip clearance data analysis method provided in this application. Detailed Implementation

[0016] The present application will now be described in further detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and not intended to limit it. Furthermore, it should be noted that, for ease of description, only the parts relevant to the invention are shown in the accompanying drawings.

[0017] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0018] As mentioned in the background section, this application proposes a machine learning-based method for analyzing rotor blade tip clearance data, such as... Figure 1 As shown, it includes the following steps: S1. Obtain the historical operating dataset of the rotor, which includes tip clearance time data, rotational speed time data, and time data of at least one other operating condition parameter. S2. Based on the rotor's structural dimensions and material thermophysical parameters, establish a simplified mechanism model of the rotor. Using the simplified mechanism model, calculate in real time the base value of the ideal gap change caused by centrifugal force deformation and thermal deformation according to the rotational speed and other operating parameters. S3. Construct a physical-data hybrid driven machine learning model, wherein the machine learning model receives the rotational speed, the other operating condition parameters, and the ideal clearance variation base value as input, fits the clearance deviation residuals in the historical operating dataset except for the ideal clearance variation base value through a residual learning branch, and adds the ideal clearance variation base value to the fitted clearance deviation residuals to obtain the blade tip clearance prediction value. S4. The machine learning model is trained using the historical running dataset. During the training process, the internal parameters of the simplified mechanism model are kept fixed or only subject to restrictive fine-tuning. The parameters of the residual learning branch are updated with the goal of minimizing the error between the predicted tip clearance and the historical measured tip clearance. S5. Deploy the trained machine learning model to the online monitoring system, input the rotational speed and operating parameters into the machine learning model in real time, obtain the predicted value of the blade tip clearance online, and simultaneously decouple and output the base value of the ideal clearance change and the clearance deviation residual. S6. Continuously monitor the clearance deviation residual, set a dynamic control limit based on the statistical characteristics of the clearance deviation residual under normal conditions, and generate a structural abnormality warning signal to indicate the change of rotor structure state when the amplitude or change trend of the clearance deviation residual exceeds the dynamic control limit, so as to distinguish it from the normal clearance fluctuation caused by the transient change of operating conditions.

[0019] Specifically, the historical operating dataset refers to a multi-dimensional time series collection recorded by a synchronous data acquisition system within a continuous historical operating cycle of the rotor equipment. This collection contains at least three types of sensor data: tip clearance time series data collected by capacitive or eddy current clearance sensors mounted on the casing, the physical quantity of which is the instantaneous distance from the blade tip to the inner wall of the casing, measured in micrometers. Sampling is typically performed by triggering acquisition at equal angles per revolution to obtain an accurate full-revolution blade clearance value; rotational speed time series data collected by magnetoelectric or Hall effect speed sensors mounted on the shaft, recording the actual number of revolutions per minute of the rotor; and "time series data of at least one other operating condition parameter," such as exhaust temperature collected by a thermocouple, compressor outlet pressure collected by a pressure transmitter, or fuel flow rate collected by a flow meter. These parameters reflect the rotor's thermodynamic and aerodynamic load state. All data is aligned and stored with a unified timestamp. The simplified mechanistic model refers to a simplified mathematical model established based on the rotor component's geometric dimensions and material properties, using analytical formulas or reduced-order numerical methods, capable of quickly calculating the main components causing clearance changes due to centrifugal force and thermal deformation. The required "structural dimensions" for this model include blade length, inner and outer diameters of the rotor disk, and inner diameter of the casing, all obtained from the design drawings. Material thermophysical parameters, including material density, elastic modulus, Poisson's ratio, and coefficient of linear expansion, can be found in material handbooks or standard specimen test reports. The ideal clearance variation baseline value is defined as the theoretically calculated value obtained by linearly superimposing the radial displacement of the blades and rotor disk caused by the centrifugal force generated by rotor rotation at a single moment, and the clearance variation caused by thermal expansion or contraction due to component temperature changes. This constitutes the main physical trend of clearance variation. The clearance deviation residual specifically refers to the sequence of differences obtained at the same moment, obtained by subtracting the ideal clearance variation baseline value calculated by the simplified mechanistic model from the tip clearance value measured by the sensor. This difference reflects the abnormal components caused by structural state changes (such as misalignment, rubbing, loosening, etc.) that the mechanistic model fails to explain.

[0020] The specific implementation methods for each step are described in detail below.

[0021] The first step is to acquire the historical operating dataset of the rotor. On a typical multi-stage axial compressor or gas turbine test bench, the sensors are first installed and calibrated. A capacitive gap sensor with a range of 0 to 2 mm and a resolution better than 0.5 micrometers is flush-mounted on the inner wall of the casing above a rotor blade of a certain stage, and its signal line is connected to a high-speed synchronous acquisition board. A key-phase sensor is installed at the shaft, using a groove to provide a reference pulse per revolution. Simultaneously, a type K thermocouple and a pressure sensor are installed at the inlet and outlet, respectively. The test bench is started and operated under multiple conditions, ranging from idle at 30% of rated speed to maximum load at 100% of rated speed. The acquisition system acquires the blade tip clearance at an equal angle sampling rate of 128 pulses per revolution, while continuously acquiring speed, all temperature, and pressure parameters at a frequency of 1000 Hz. Continuous operation and recording are performed for approximately 100 hours, covering multiple cold starts, shutdowns, and steady-state and transient variable-condition processes. After the data collection is completed, all data is exported as a timestamped matrix format and data cleaning is performed. Specifically, this includes: using the Laida criterion to remove singular outliers in the gap signals and interpolating with the mean before and after; using moving average filtering to remove high-frequency electrical noise; and using a cross-correlation algorithm to align the micro-hour shifts between channels caused by sensor response delays, ultimately generating a clean and aligned historical operation dataset.

[0022] The second step involves establishing a simplified mechanistic model of the rotor based on its structural dimensions and material thermophysical properties. Taking a single-stage rotor as an example, the effective blade length L, the radius R_disk of the disk from the center of rotation to the blade root, and the inner wall radius R_casing of the casing are measured and recorded. The density ρ, elastic modulus E, and linear expansion coefficient α of the blade and disk materials are then obtained. The centrifugal deformation is calculated using the equivalent rotating beam theory: the blade is considered as a cantilever beam with a uniform cross-sectional area A, fixed at one end to the disk and extending radially. Its radial centrifugal elongation Δr_centrifugal = (ρ * A * ω² * L³) / (3 * E * A) is simplified to this value. The radial expansion of the disk is calculated using the formula Δr_disk = (ρ * ω² * R_disk³) / (4 * E) based on the rotating disk theory. The thermal deformation calculation employs the lumped parameter method: the weighted average of the inlet total temperature T_inlet and the exhaust temperature T_outlet, T_bulk = 0.3*T_inlet + 0.7*T_outlet, is used as the approximate average temperature of the blade, and the casing temperature is assumed to be T_casing = T_outlet. Then, the blade thermal elongation Δr_thermal = α * L* (T_bulk - T_ref), and the casing thermal expansion Δr_casing = α * R_casing * (T_casing - T_ref), where T_ref is the reference temperature at assembly. Finally, the ideal clearance variation base value Clearance_base = (Δr_thermal + Δr_centrifugal) - Δr_casing. All the above calculations are written into a highly efficient function module, with real-time rotational speed ω, inlet temperature T_inlet, and exhaust temperature T_outlet as inputs, and the output being the real-time varying Clearance_base value.

[0023] The third step involves constructing a physics-data hybrid driven machine learning model. A neural network structure is designed with a six-dimensional input: [current engine speed, current exhaust temperature, current inlet pressure, previous engine speed, previous exhaust temperature, and the ideal clearance change baseline value (Clearance_base) calculated in real-time by the mechanism model in the second step]. This network contains a "residual learning branch" as its main component, specifically a three-layer fully connected feedforward network. The input layer has six neurons, hidden layer 1 contains 64 neurons using the ReLU activation function, and hidden layer 2 contains 32 neurons. The output layer is a single neuron, whose output is the "fitted clearance deviation residual (Residual_pred)". At the end of the network, an additive fusion layer is set up to algebraically add the neural network output (Residual_pred) to the ideal clearance change baseline value (Clearance_base) in the input, obtaining the final "tip clearance prediction value (Clearance_pred)" at that moment. The network weights are initialized using the Xavier method.

[0024] The fourth step involves training the machine learning model using the historical dataset. The dataset prepared in the first step is randomly divided into training, validation, and test sets at a ratio of 70%, 15%, and 15%, respectively. The loss function is defined as Mean Squared Error (MSE) = (1 / N)*Σ(Clearance_pred_i - Clearance_measured_i)², where Clearance_measured_i is the historical true gap value. The Adam optimizer is used, with an initial learning rate of 0.001. In the training loop, a batch of data from the training set is input into the model each time, and Clearance_pred is calculated through forward propagation. Crucially, during backpropagation to calculate gradients and update parameters, only the gradients are passed to the neuron weights and biases of the "residual learning branch" constructed in the third and fourth steps. All physical formulas and parameters (such as L, E, α) within the "mechanism-simplified model" established in the second step are set to an untrainable state, and the gradient flow is truncated here. This ensures that physical knowledge serves as invariant prior knowledge throughout the process. Training continues until the MSE loss on the validation set no longer decreases within 20 consecutive iterations, at which point the model is saved.

[0025] The fifth step is to deploy the trained machine learning model to the online monitoring system. After compiling the saved model file and mechanistic model function, it is deployed to an edge computing gateway in the industrial site. This gateway receives sensor data streams from the data acquisition system in real time. During online operation, in each sampling cycle, the gateway program first calls the mechanistic model function to calculate the current ideal clearance change baseline value using real-time rotational speed and operating parameters. Then, it feeds this value, along with the normalized rotational speed and operating parameters, into the machine learning model. The model performs a forward calculation and outputs the tip clearance prediction value at the additive fusion layer. Simultaneously, the program timestamps the calculated ideal clearance change baseline value and the clearance deviation residual prediction value output by the model branch, publishes them to the monitoring data bus, and plots them as two trend curves on the interface.

[0026] The sixth step involves continuously monitoring the gap deviation residuals and generating structural anomaly early warning signals. During the first 200 hours of stable operation in the initial phase of system deployment, the program automatically collects the gap deviation residual sequence from the online decoupling output and establishes its statistical baseline. The calculation employs an exponentially weighted moving average method. For each newly arrived residual value x_t, the mean μ_t = α*x_t + (1-α)*μ_{t-1} and variance σ²_t = α*(x_t - μ_t)² + (1-α)*σ²_{t-1} are updated, with a smoothing coefficient α of 0.1. The upper limit for dynamic control is set to μ_t + 3σ_t, and the lower limit is set to μ_t - 3σ_t. Simultaneously, a sliding window with a length of 100 sampling points is maintained, and the least-squares linear fitting slope k of the residual sequence within the window is calculated in real time. The system employs a two-level early warning logic: a yellow warning is immediately triggered when any instantaneous residual value exceeds the dynamic control limit; if the absolute value of the trend slope k exceeds a preset threshold (e.g., 0.01 micrometers / minute) for five consecutive windows, a red warning is triggered, generating a "structural anomaly warning signal." Under this mechanism, most normal clearance fluctuations caused by rapid acceleration and deceleration are compensated by the base values ​​calculated in real time by the mechanism model, resulting in minimal changes in the residuals. This effectively avoids false alarms caused by rapid changes in operating conditions, ensuring that the warnings point to genuine changes in the rotor structure, such as early rubbing or uneven wear.

[0027] The technical principle behind this solution lies in embedding immutable physical formulas into a neural network, forcing the model's learning objective to shift from fitting the entire data distribution to fitting only the residuals that the physical model cannot explain. This solves the problem of poor extrapolation ability of pure data models in unseen high-speed regions. Decoupling the output baseline and residuals naturally separates the influence of operating conditions from the influence of structural state, allowing subsequent dynamic control limits to be directly based on the residuals characterizing structural health. The technical effect is a significant improvement in the accuracy of clearance monitoring and the reliability of anomaly warning under varying operating conditions, achieving a balance between high sensitivity and low false alarm rate.

[0028] In a preferred embodiment, before training the machine learning model, the time-series data of the decoupled gap deviation residuals are subjected to signal transformation processing to construct an enhanced residual feature set; The signal transformation processing includes: taking the time-series data of the gap deviation residual as a time-domain signal, converting it into an order spectrum with rotational speed as the horizontal axis based on the synchronously acquired key phase signal, and / or converting the time-series data of the gap deviation residual into a time-spectrum graph. The feature representation of the transformed order spectrum or time spectrum is used as one of the input features of the residual learning branch.

[0029] Specifically, this section describes the methods for signal transformation of data to construct enhanced feature sets before model training. First, the technical terms involved are clarified. Signal transformation processing specifically refers to mapping time-domain gap deviation residual time-series data to other representation domains through mathematical transformations to reveal its structured patterns in the frequency or time-frequency domains. The key phase signal is a binary signal output by a key phase sensor mounted on the shaft, generating a voltage pulse per revolution. Its pulse rising edge precisely corresponds to a fixed phase angle of the rotor shaft system, serving as a reference for equal-angle sampling. The order spectrum is the result of a special spectral analysis method for rotating machinery signals. Its processing involves: using the key phase signal as a trigger, resampling the time-domain signal sampled at equal time intervals through interpolation into a signal with equal angular intervals; then performing a Fast Fourier Transform (FFT) on the resampled angular domain sequence. The horizontal axis of the spectrum represents the order, i.e., the frequency multiple of the current rotational frequency. The time-frequency spectrum is a signal representation combining time and frequency dimensions, calculated through a Short-Time Fourier Transform (STFT). It shows how the frequency components of a signal evolve over time. The enhanced residual feature set refers to the digital representation of the order spectrum or time spectrum obtained by the above transformation as a new feature channel, which is then combined with the original input features such as rotational speed and operating parameters to form a feature set with richer information dimensions and greater discriminative power.

[0030] The specific implementation steps are as follows. Before starting to train the machine learning model, an important data preprocessing and feature engineering process needs to be completed. The input is the decoupled historical gap deviation residual sequence and the synchronized bond phase pulse sequence.

[0031] For constructing the order spectrum: First, the bond phase pulse signal is detected, and the time interval sequence between the rising edges of every two adjacent pulses is recorded, thus calculating the precise duration of each rotation cycle. Then, an angular resolution is set, for example, 64 equal-angle resampling operations per rotation, achieving an order resolution of 32. For the original continuous time-domain gap deviation residual segment, equal-angle interpolation resampling is performed using the time information per rotation provided by the bond phase pulse. Specifically, for each rotation, based on the time consumed, the theoretical timestamps corresponding to the 64 equal-angle points to be interpolated are calculated. Then, using the original unequal-interval sampling data, a cubic spline interpolation algorithm is used to calculate the gap deviation residual values ​​at these 64 angle points. In this way, the residual signal for the entire time period is converted into a matrix arranged by rotation cycle and angle. Next, a one-dimensional fast Fourier transform is performed on each column of this matrix (i.e., data from different rotations at the same angle) to obtain an order spectrum matrix, where the horizontal axis represents the order (from 0 to 32) and the vertical axis represents the spectral amplitude. Finally, extract the amplitudes of the top N (e.g., the top 16) strongest order spectral lines, or directly take the entire order spectral vector as a feature representation.

[0032] For the construction of the time-frequency spectrogram: a short-time Fourier transform is directly applied to the time-domain gap deviation residual sequence. A Hamming window is chosen as the window function, with a window width set to cover the signal length of approximately 64 rotational speed cycles to ensure sufficient frequency resolution. The sliding step size is set to approximately the signal length of one rotational speed cycle to ensure sufficient time resolution. An FFT is performed on the signal segment within each window to obtain the instantaneous spectrum at that moment. The spectra of all windows are concatenated in chronological order to obtain a two-dimensional time-frequency spectrogram matrix. The horizontal axis of this matrix represents the timestamp of the center point of the time window, the vertical axis represents the frequency, and the matrix element value is the power spectral density at that time-frequency point. Subsequently, the amplitude of the time-frequency spectrogram is logarithmically taken, and it is downsampled to a fixed size, such as a 128x128 pixel grayscale image. The grayscale value of each pixel is used as the feature representation of that time-frequency unit. After the above transformation is completed, the calculated order spectral feature vector or time-frequency spectrogram matrix is ​​associated and aligned with the corresponding rotational speed, operating parameters, and ideal gap variation baseline value at the same moment. For example, for each time window in the time-spectrum graph, its time center point is taken, and the rotational speed and operating condition data at that moment are found and combined into a sample. These samples containing spectral information constitute the "enhanced residual feature set". In subsequent machine learning model training, the order spectral vector or time-spectrum graph matrix in this feature set will be fed into the residual learning branch as an independent input stream.

[0033] The technical principle of this solution is based on order tracking theory and time-frequency analysis theory in rotating machinery fault diagnosis. Since the vibration and gap fluctuation signals of a rotor system are inherently non-stationary during variable speed operation, traditional spectrum analysis produces a "frequency ambiguity" phenomenon. By resampling the key phase signal at equal angles, the analysis domain is transformed from the time domain to the angle domain, making the vibration signal synchronized with the rotational speed exhibit a constant order spectrum independent of the rotational speed. This greatly improves the stability and separability of fault features (such as frequency-division harmonics caused by rubbing). The time-spectrum diagram provides a dynamic perspective on the occurrence and development of the fault. Injecting these physically meaningful transform domain features as prior knowledge into the model is equivalent to guiding the data-driven model to focus on regions in the feature space that are truly related to the fault mechanism. This fundamentally improves the model's feature extraction capability and robustness under strong noise and multi-condition coupling, making subsequent anomaly detection more sensitive.

[0034] In a preferred embodiment, the residual learning branch is implemented using a dual-channel neural network, including the following steps: Construct a first channel, input the rotational speed, the other operating parameters, and the ideal clearance variation base value into the first channel, and extract the operating feature vector; Construct a second channel, which is a convolutional neural network. Input the feature representation of the order spectrum or time spectrum of the gap deviation residual into the second channel, and extract the residual spectrum feature vector. The fusion module fuses the operating condition feature vector with the residual spectrum feature vector, and outputs the predicted value of the gap deviation residual based on the fused features.

[0035] Specifically, a dual-channel neural network refers to a composite neural network architecture that includes two parallel, structurally dissimilar input and feature extraction paths, which are fused at the feature map or feature vector level. The first channel is designed as a dedicated channel for processing operating condition-related parameters. Its input is structured data in scalar or low-dimensional vector form, and its output is an operating condition feature vector that globally describes the current operating state of the equipment. The second channel is a convolutional neural network (CNN) channel, specifically used to process high-dimensional structured data with spatial or local correlation, i.e., the feature representation of the generated "order spectrum" or "time spectrum", and outputs a "residual spectrum feature vector" that captures local frequency domain or time-frequency domain patterns. The "fusion module" is the structure for information integration at the higher levels of the network. Specifically, it concatenates the feature vectors from the two sources, then performs nonlinear combination and dimensionality reduction through several fully connected layers, finally outputting the required scalar prediction value.

[0036] The specific steps for network construction and implementation are as follows.

[0037] Construction of the First Channel: This channel aims to extract the operating context from macroscopic operating parameters. First, the shape of the input layer is defined as a one-dimensional vector containing four elements: the normalized current engine speed, the normalized exhaust temperature, the normalized compressor outlet pressure, and the ideal gap variation baseline value output by the mechanistic model. Next, two fully connected hidden layers are constructed. The first hidden layer contains 64 neurons with a weight matrix of W1 and a bias of b1. After a linear transformation of the input vector x, the ReLU activation function is applied, outputting FeatureMap1 = max(0, W1*x + b1). The second hidden layer contains 32 neurons with the same structure, outputting a 32-dimensional vector, which we name "Operating Condition Feature Vector F_condition". This vector highly abstractly encodes the combined state of the current rotor operating speed, thermal load, and physical baseline. To prevent overfitting, a Dropout layer is added after each of these two fully connected layers, randomly discarding neurons with a probability of 0.5 during training.

[0038] The second channel is constructed using a convolutional neural network to process spectral features. Assuming we use a 128x128 time-spectrum map as input, we design a feature extraction network consisting of three stacked convolutional blocks. The first block contains a convolutional layer with 32 kernels of size 3x3, a stride of 1, and padding using a "Same" pattern to maintain the feature map size. This is followed by a batch normalization layer and a ReLU activation layer, and finally a 2x2 max-pooling layer, reducing the feature map size to 64x64x32. The second block increases the number of kernels to 64, and similarly performs convolution, normalization, activation, and pooling, outputting a 32x32x64 feature map. The third block follows the same process, with 128 kernels, ultimately outputting a 16x16x128 feature map. Next, this 3D feature map is flattened into a 1D long vector and connected to a fully connected layer containing 128 neurons. After ReLU activation, it outputs a 128-dimensional vector, which we call the "residual spectral feature vector F_spectrum". This vector is automatically learned by the network from the time-spectrum map and contains high-order features such as frequency band energy distribution and time-varying modes related to structural faults.

[0039] Construction and training of the fusion module: First, the 32-dimensional "condition feature vector F_condition" from the first channel output and the 128-dimensional "residual spectrum feature vector F_spectrum" from the second channel output are concatenated to obtain a 160-dimensional fusion feature vector F_fusion = [F_condition, F_spectrum]. Then, F_fusion is fed into a fusion sub-network consisting of two fully connected layers. The first fusion layer contains 32 neurons activated by ReLU, and the second fusion layer is a single linear output neuron whose activation value is the current time step's "predicted value of gap deviation residual, Residual_pred". The training of the entire dual-channel network model is consistent with the previously described process, with the loss function being the mean squared error between Residual_pred and the actual gap deviation residual obtained from historical decoupling. The optimizer simultaneously updates all trainable weights and biases of the first channel, the second channel, and the fusion module.

[0040] The technical principle of this solution lies in the effective fusion of multi-source heterogeneous data in the feature space. The fully connected network of the first channel is naturally suitable for processing scalar operating parameters with independent physical meanings and different dimensions. Through deep connections, it learns the comprehensive modulation law of the gap caused by their complex nonlinear combinations. The convolutional neural network of the second channel, utilizing its local connectivity and weight sharing characteristics, can effectively capture local time-frequency patterns representing specific faults (such as frequency friction caused by rubbing, subharmonic clusters) in the time-frequency spectrum. It is also insensitive to the position of these patterns in the spectrum and has strong translation invariance. The fusion module couples two highly complementary feature vectors, "macroscopic operating state" and "microscopic frequency dynamics," at the decision layer. This allows the model to predict the residual not only based on the current operating condition amplitude (dominated by F_condition) but also to correct this prediction based on abnormal patterns in the spectrum (dominated by F_spectrum), thereby achieving prediction accuracy and generalization ability far exceeding that of a single-channel model. The ultimate technical effect is to provide a highly accurate dynamic benchmark for gap deviation residuals, so that any structural anomalies that deviate from this benchmark can be identified more clearly and earlier, greatly improving the early warning capability of the online monitoring system.

[0041] In a preferred embodiment, generating a structural anomaly early warning signal to indicate a change in the rotor's structural state includes the following steps: Statistical analysis is performed on the gap deviation residuals output by online decoupling, and dynamic control limits are constructed based on the mean and standard deviation of the gap deviation residuals under historical normal conditions. When the instantaneous amplitude of the gap deviation residual exceeds the dynamic control limit, or the trend slope within its sliding window exceeds a preset threshold, the structural anomaly warning signal is generated. The time-domain and frequency-domain patterns of the gap deviation residuals are identified, and the identified patterns are compared with a preset fault residual pattern library to provide an indication of the associated potential fault type while generating the structural anomaly warning signal.

[0042] Specifically, the dynamic control limit refers to a normal fluctuation range that adapts in real time to the equipment's operating status and the external environment. It is not a fixed value but is calculated online based on the statistical characteristics of the residual data stream, forming a band-shaped interval consisting of an upper and lower limit. "The mean and standard deviation of the gap deviation residuals under historical normal conditions" refers to the arithmetic mean and standard deviation obtained by statistically calculating all online decoupled gap deviation residual values ​​during a reference time period after confirming the integrity of the equipment structure during the initial stage of system operation. This arithmetic mean and standard deviation serve as the benchmark statistics for subsequent monitoring. "Trend slope" specifically refers to the slope of the fitted straight line obtained by performing a univariate linear regression on the gap deviation residual time series contained within a fixed-length sliding observation window. The unit is micrometers per minute or micrometers per hour, used to quantify the rate and direction of residual change. The "Fault Residual Pattern Library" is a pre-built database that stores the unique time-domain waveform patterns and frequency-domain characteristic patterns of various typical structural faults (such as blade cracks, wheel loosening, rubbing, bearing wear, etc.) in the gap deviation residual signals when they occur. For example, rubbing often manifests as the appearance of fractional harmonics such as 0.5 harmonics and 1.5 harmonics.

[0043] The specific implementation steps are as follows: Step 1: Construction of dynamic control limits. After the system is deployed and it is confirmed through trial operation that the rotor is in good condition, it automatically enters a baseline learning period, for example, running continuously for 24 hours. During this period, each gap deviation residual value r_t of the online decoupled output is continuously received, and the mean μ_0 and standard deviation σ_0 of all residuals in the entire baseline period are calculated. Then, the system switches to the online monitoring mode and enables the exponentially weighted moving average algorithm to dynamically update the statistics to adapt to the slow changes in the environment. For each newly arrived residual value r_t, the following update calculations are performed: mean μ_t = β * r_t + (1 - β) * μ_{t - 1}, variance σ²_t = β * (r_t - μ_t)² + (1 - β) * σ²_{t - 1}, where the smoothing factor β is taken as 0.05. The upper control limit UCL_t at the current time t = μ_t + k * σ_t, and the lower control limit LCL_t = μ_t - k * σ_t. The coefficient k is usually taken as 3, representing the three - standard - deviation principle. When the operating point changes significantly, the model may jump from one steady state to another. To avoid false alarms, when a large - amplitude step change in speed or load exceeding the preset threshold is detected, the update of the statistics is paused for a short delay to wait for the operating conditions to stabilize.

[0044] Step 2: Two - dimensional warning logic based on amplitude and trend. For each newly generated residual r_t, two independent judgments are made. The first is the instantaneous over - limit judgment, which directly compares r_t with UCL_t and LCL_t. If r_t > UCL_t or r_t < LCL_t, an "amplitude over - limit event" is immediately recorded. The second is the trend over - speed judgment. The system maintains a first - in - first - out queue with a fixed length N (for example, N = 120 points, representing 2 - hour data). When the queue is full, each time new data enters, the least - squares method is used to calculate the slope of the straight line fitted by the N data points in the queue. The specific formula is slope = Σ[(x_i - ) * (y_i - )] / Σ[(x_i - )²], where x_i is the time index and y_i is the residual value. If the absolute value of this slope exceeds the preset threshold continuously for several times, for example, greater than 0.1 μm / hour in three consecutive calculations, a "trend anomaly event" is recorded. When the "amplitude over - limit event" occurs frequently in a short period or the "trend anomaly event" is triggered, the system generates a "structural anomaly warning signal" and pushes this warning information to the monitoring interface.

[0045] The third step involves time-domain and frequency-domain pattern recognition and fault comparison. Once a structural anomaly warning signal is generated, the system automatically initiates an in-depth pattern analysis. The program extracts the gap deviation residual signal segment within a time window preceding the warning time (e.g., from 10 minutes prior to the current time) and performs time-domain and frequency-domain feature extraction. Time-domain features include peak value, RMS value, skewness, and kurtosis; frequency-domain features are obtained by performing a Fast Fourier Transform on this signal segment, extracting the amplitude of the dominant frequency and the amplitudes of its harmonics and sub-frequency divisions to form a fault feature vector. Subsequently, the system iterates through a pre-set "fault residual pattern library." This library exists in the form of a database table, with each record containing a fault name and a standardized fault feature vector. The matching algorithm calculates the Euclidean distance or cosine similarity between the currently extracted feature vector and each standard vector in the library. The fault type with the highest similarity exceeding a certain threshold (e.g., 0.8) is selected, and the warning signal and the associated potential fault type indication "possibly: radial rubbing" are pushed together to the diagnostic report, providing maintenance personnel with direct decision-making clues.

[0046] The technical principle of this solution lies in its dynamic control limit, which perfectly adapts to the long-term, slow changes in the residual baseline caused by environmental temperature and sensor drift, thus avoiding the failure of fixed thresholds after long-term operation. It incorporates two abnormal behaviors—abrupt amplitude changes and continuous trend changes—into independent judgment logic, ensuring efficient capture of both sudden faults (such as foreign object damage) and gradual faults (such as wear). The integrated pattern recognition function solidifies expert experience and knowledge in the fault mode library, providing qualitative diagnosis simultaneously with early warning, effectively integrating monitoring and diagnosis. The ultimate effect is to advance the average warning time, significantly reduce false alarms caused by operating condition fluctuations, and provide targeted diagnostic information, significantly improving the system's intelligent operation and maintenance level.

[0047] In a preferred embodiment, the method further includes the step of quantifying the uncertainty of the predicted tip clearance value: Obtain the ideal gap variation base value output by the simplified mechanism model, and calculate the prior distribution of the ideal gap variation base value based on the perturbation range of the material thermophysical parameters of the rotor; The residual learning branch outputs the predicted value of the gap deviation residual, and at the same time outputs the residual prediction variance of the predicted value; The prior distribution of the ideal gap variation baseline value is fused with the predicted value of the gap deviation residual and its predicted variance using Bayesian fusion to obtain the posterior distribution of the predicted blade tip gap value. Wherein, the mean of the posterior distribution is the final predicted tip clearance value, and the variance interval of the posterior distribution is the prediction uncertainty interval.

[0048] Specifically, deterministic quantification refers to quantitatively assessing the possible range of variation or confidence level of a given predicted value, typically expressed as a confidence interval or probability distribution. "The perturbation range of material thermophysical parameters" refers to the fact that, in reality, parameters such as the elastic modulus and coefficient of linear expansion of a material are not constant values, but rather fluctuate within a known small range due to factors such as manufacturing batch and operating temperature. For example, the coefficient of linear expansion of a certain alloy is 12.0 × 10⁻⁶. -6 / °C to 12.5×10 -6 Between / °C. "Prior distribution" specifically refers to the probability distribution of the ideal gap variation baseline value derived from the inherent parameter uncertainties of the physical model, without considering current sensor measurement data; it is usually assumed to follow a Gaussian distribution. "Residual prediction variance" is a measure of the uncertainty of the prediction itself, output simultaneously by the residual learning branch in a machine learning model when making point predictions; it is the model's intrinsic assessment of the reliability of its predictions. "Posterior distribution" is a probabilistic description of the final state of the blade tip gap value after obtaining actual measurement information or fusing more information sources; its mean is the most reliable prediction value, and its variance interval gives the range of prediction uncertainty.

[0049] The specific implementation steps are as follows: The first step is to calculate the prior distribution of the ideal clearance variation base value. First, identify the main parameters exhibiting perturbations in the simplified mechanistic model. For example, the linear expansion coefficient α and elastic modulus E of the blade and disk materials. By consulting material handbooks, the nominal value of α is found to be 12.3 × 10⁻⁶. -6 / °C, its total perturbation range due to compositional fluctuations and temperature changes is ±4%, i.e., standard deviation σ_α = (12.3 × 10) / °C. -6 * 0.04) / 3, here is an estimate. The perturbation range of the elastic modulus E is obtained similarly. A Monte Carlo sampling method is used to randomly sample a set of parameter values ​​from the Gaussian distributions of α and E simultaneously, for example, 1000 sets. Each set of parameters, along with current operating conditions such as rotational speed and temperature, is input into the simplified mechanistic model to calculate 1000 samples of ideal gap variation base values. Then, the mean and standard deviation of these 1000 samples are calculated to model the prior distribution of the ideal gap variation base values ​​as a Gaussian distribution N(μ_base, σ²_base).

[0050] The second step is to make the residual learning branch output the residual prediction variance. The output layer of the dual-channel neural network is modified. The original residual learning branch had only one output neuron, outputting the residual prediction value μ_res. Now we change it to a structure with two parallel output neurons, one outputting μ_res and the other outputting a parameter representing precision, log_precision = log(1 / σ²_res). During training, instead of using the standard mean squared error loss, a loss function that can handle heteroscedastic noise, namely the negative log-likelihood loss, is adopted. For a batch of training data, the loss function is calculated as: L = 0.5 * Σ [log(2π) - log_precision_i + (exp(log_precision_i) * (residual_true_i - μ_res_i)²)]. By minimizing this loss, the network learns to predict the residuals while also learning the inherent, non-eliminable randomness variance σ²_res = exp(-log_precision) of the residual value under the current input combination. After training, the residual learning branch can simultaneously provide μ_res and σ²_res during inference.

[0051] The third step involves Bayesian fusion to obtain the posterior distribution. During the online deployment phase, for each new time step, we possess information about the prior distribution N(μ_base, σ²_base) and the likelihood component, i.e., the output N(μ_res, σ²_res) of the residual learning branch. According to Bayes' theorem, the final tip gap prediction also follows a Gaussian posterior distribution. Since we assume that the prior and the predicted residuals are independent, the mean μ_post of the posterior distribution can be obtained by simply adding the base prediction and the residual prediction: μ_post = μ_base + μ_res. The update of the posterior distribution variance follows the inverse variance weighting principle, i.e., σ²_post = σ²_base + σ²_res, because the gap prediction is a linear sum of the base value and the residual, and its total variance is the sum of their individual variances. Finally, the monitoring system displays the real-time tip gap value as μ_post, and renders an uncertainty interval band around it, such as [μ_post - 2*σ_post, μ_post + 2*σ_post]. This band visually represents the reliability of the current prediction; the more uncertain the model, the wider the band.

[0052] The technical principle of this solution lies in its first formal unification and quantification of the uncertainty of physical model parameters and the cognitive uncertainty of data-driven models in complex nonlinear fitting within the field of blade tip clearance monitoring. For the physical model, the introduction of prior distributions acknowledges the imperfections of our physical knowledge. For the data model, allowing it to output variance and using negative log-likelihood for training is essentially a mechanism of "knowing when not knowing." When the input data deviates significantly from its training distribution or is extremely noisy, the model automatically generates a large σ²_res, indicating that its predictions are unreliable. Through Bayesian fusion, this measure of unreliability is transmitted to the final decision variables. The technical effect is that it assigns a confidence level to each prediction, allowing maintenance personnel to make risk-controlled decisions. For example, when the uncertainty interval is wide, the system can automatically suppress warnings or suggest more cautious manual review, greatly improving the system's robustness and security in complex real-world environments.

[0053] In a preferred embodiment, generating a structural anomaly early warning signal to indicate a change in the rotor's structural state includes the following steps: Obtain the predicted value of the gap deviation residual and the corresponding portion of the residual in the prediction uncertainty interval; The degree to which the gap deviation residual obtained from online decoupling deviates from its predicted value is calculated and used as the standardized residual deviation. Based on the magnitude of the residual prediction variance, the adaptive warning threshold for the current time is dynamically calculated and set; wherein, the larger the residual prediction variance, the higher the adaptive warning threshold is set. The standardized residual deviation is compared with the adaptive early warning threshold. If the standardized residual deviation continues to exceed the adaptive early warning threshold, a structural anomaly early warning signal based on uncertainty perception is generated.

[0054] The specific implementation steps are as follows: The first step is to calculate the standardized residual bias online. At each sampling time, the monitoring system simultaneously acquires three key pieces of information: the current true gap bias residual value r_actual obtained by directly decoupling the mechanistic model and measurement data; the residual prediction mean μ_res output by the residual learning branch in the two-channel neural network; and the residual prediction variance σ²_res output by the same branch, and calculates the standard deviation σ_res = sqrt(σ²_res). Standardization is then immediately performed: Z_score = (r_actual - μ_res) / σ_res. This Z_score is the standardized score that measures whether the current residual is an anomaly or a normal fluctuation. When the system is functioning normally, the model prediction is accurate, and the uncertainty is very low, the Z_score should approximately follow a standard normal distribution, mostly falling between -3 and 3.

[0055] The second step is to dynamically calculate the adaptive warning threshold. A base warning threshold T_base is set, for example, T_base = 3.0, corresponding to the conventional 3σ principle. Now, we design a threshold adjustment function such that the threshold T_adaptive = T_base + f(σ_res). f(σ_res) is a positive correlation function with respect to σ_res. For example, it can be simply defined as f(σ_res) = γ * σ_res, where γ is a configurable sensitivity coefficient. This means that when σ_res is large, i.e., the model's prediction is extremely uncertain at this moment, T_adaptive will be significantly increased, for example, from 3.0 to 5.0 or higher. A smoother design is T_adaptive = T_base * (1 + σ_res / σ_res_nominal), where σ_res_nominal is the model's average prediction standard deviation under typical conditions. In this way, the threshold can continuously and smoothly increase proportionally with the model's uncertainty.

[0056] The third step is to generate a structural anomaly warning based on uncertainty perception. The absolute value of Z_score calculated in the first step is continuously compared with the adaptive threshold T_adaptive calculated in the second step. To avoid single-point noise, a continuous judgment logic is adopted. A counter is maintained; when |Z_score|>T_adaptive, the counter is incremented by 1; when the condition is not met, the counter is divided by 2 or reset to zero. When the counter value exceeds a set number of consecutive exceedances, such as 5 consecutive sampling points exceeding the limit, the system determines to generate a "structural anomaly warning signal." This mechanism ensures that when the model faces high noise or unseen uncertain conditions (such as severe intake distortion) causing σ_res to spike, even if the actual residual value fluctuates significantly, the alarm will not be easily triggered because the threshold also increases synchronously. Conversely, under steady-state conditions, the model is extremely confident, σ_res is small, and the threshold is close to the baseline value. At this time, any small structural anomaly will cause Z_score to increase sharply and be quickly detected.

[0057] In a preferred embodiment, the method further includes the following steps: Retrieve the fault feature database, which contains a mapping relationship between multiple preset fault modes and a set of standard fault feature vectors. The standard fault feature vector corresponding to each fault mode is described by the time domain statistical features and frequency domain features of the gap deviation residual under the corresponding fault state. For real-time data, extract the real-time residual feature vector from the gap deviation residual output by online decoupling; The real-time residual feature vector is matched with all standard fault feature vectors in the fault feature database, and its static matching degree score with each fault mode is calculated. Based on the static matching score and the prediction uncertainty interval, a multi-dimensional fault contribution vector is calculated, where each dimension corresponds to a probability score of a fault mode. By combining the fault contribution vector with the structural anomaly warning signal, a diagnostic report is generated that includes the specific suspected fault mode, severity, and confidence level.

[0058] Specifically, a refined fault diagnosis system combining static feature matching and uncertainty scoring was constructed, aiming to generate diagnostic reports containing specific fault modes, severity, and confidence levels. The fault feature database is a pre-built relational or key-value database. Its core structure stores a mapping relationship from a multidimensional vector composed of time-domain statistical features and frequency-domain features to the name of a specific fault mode. For example, for a "leaf tip rubbing" fault, the kurtosis index in its standard fault feature vector will be high, and the amplitude proportions of the 0.5 and 1.5 harmonics in the spectrum will be significant. The "static matching score" is a score between 0 and 1, obtained by calculating the similarity (such as cosine similarity) between the real-time extracted feature vector and each standard vector in the database, without considering any uncertainties. A higher score indicates a more similar mode. The "fault contribution vector" is a multidimensional vector whose dimension equals the number of fault modes in the database. Each element in the vector represents the probability score of the corresponding fault mode occurring under all current evidence; it is a final score that integrates static feature matching and prediction confidence.

[0059] The specific implementation steps are as follows: The first step is to retrieve and structure the fault feature database. This database needs to be built first. For the M common fault modes of rotor systems, such as "uniform wear," "localized rubbing," "blade cracks," and "casing deformation," a standard fault feature vector V_std_j is defined for each fault mode j, using historical fault cases, simulation data, or domain expert knowledge. This vector is composed of two parts: time-domain statistical features [RMS value, peak factor, skewness, kurtosis] and frequency-domain features [1st harmonic amplitude, 2nd harmonic amplitude, ..., 0.5th harmonic amplitude, ...]. After establishing this M×N dimensional (N is the number of features) database, it is serialized and loaded into the memory of the online monitoring system.

[0060] The second step is to extract the real-time residual feature vector. When the diagnostic process is triggered by an early warning signal, the system acquires the gap deviation residual signal window for the current time and a period of time prior. For this signal window, N features that are completely consistent with the definitions in the database are calculated to form a real-time residual feature vector V_real_time. The calculation process is standardized; for example, for frequency domain features, the signal must first be windowed before performing FFT and amplitude normalization to ensure strict consistency with the processing method used when building the database.

[0061] The third step is to calculate the static matching score. For M fault modes, the cosine similarity between V_real_time and each V_std_j is calculated iteratively. The cosine similarity calculation formula is Sim_j = (V_real_time · V_std_j) / (||V_real_time|| * ||V_std_j||). Since all vector elements are positive or have undergone mean removal processing, the value of Sim_j is usually between 0 and 1. This Sim_j is the static matching score for the j-th fault. We can obtain a set of score vectors S_static = [Sim_1, Sim_2, ..., Sim_M].

[0062] The fourth step is to fuse uncertainties to calculate the fault contribution vector. The prediction uncertainty interval for the current moment is obtained from the output, specifically the variance σ²_res corresponding to the residual. This variance reflects the reliability of the current data and model state. We convert σ²_res into a confidence coefficient, for example, Confidence = exp(-λ * σ²_res). The greater the uncertainty, the lower the confidence. Then, the static matching score vector S_static is multiplied by this scalar Confidence to obtain the final fault contribution vector Contribution = Confidence * S_static. The physical meaning of this step is: even if a residual pattern is very similar to the "bump and rub" pattern in the database, if the system is currently in an extremely uncertain operating condition (σ²_res is extremely large), then the confidence of this pattern match is low, and the final score for the probability of the fault will be lowered to avoid giving an overconfident erroneous diagnosis under noisy or unknown conditions.

[0063] The fifth step is to generate a diagnostic report. Based on the warning signals and contribution vectors from the preceding steps, a structured diagnostic report is generated. The report includes: basic information such as trigger time and abnormal residual amplitude; a list of suspected fault modes sorted by the contribution vector, for example, "localized impact" is ranked first with a probability score of 0.85 (out of 1); the severity is assessed by the percentage of abnormal residual amplitude exceeding the threshold and signal characteristics (such as the number of impacts), categorized as "mild, moderate, or severe"; the confidence level is directly given by the current overall confidence level, such as Confidence = 0.9, which indicates a 90% certainty about the diagnosis.

[0064] In a preferred embodiment, some fault modes in the fault feature database have corresponding temporal evolution prototype vectors, which describe the typical pattern of the evolution of the characteristics of the gap deviation residual under the corresponding fault mode over time.

[0065] Specifically, the temporal evolution prototype vector is a sequence of multidimensional feature vectors arranged chronologically. It comprehensively depicts the typical trajectory of the gap deviation residual in the time and frequency domains throughout the entire lifecycle of a specific fault, from its inception to its development and deterioration. Unlike static fault feature vectors, which only describe the mature state of a fault after it has fully developed, the temporal evolution prototype vector captures the dynamic evolution of the fault. For example, a rubbing fault may initially manifest as a slight scraping pulse per revolution. As wear intensifies, the amplitude of this pulse gradually increases, and frequency sub-components begin to appear in the spectrum and gradually strengthen. Preset fault modes refer to typical fault types with clear physical mechanisms that have been confirmed and labeled by experts in the fault feature database, such as radial rubbing, axial rubbing, loose blade tenons, and disc cracks. Typical patterns refer to trajectory templates obtained through statistical analysis of a large number of historical fault cases, fault injection experiments, or high-fidelity simulations, representing the common evolutionary characteristics of the fault in most cases.

[0066] The specific implementation steps are as follows.

[0067] The first step is to identify which failure modes require the construction of temporal evolution prototype vectors. Not all failure modes exhibit clear evolutionary patterns; for example, sudden blade fracture lacks a continuous evolutionary process. Therefore, domain experts and equipment engineers first screen out several failure modes with progressive evolution characteristics based on the rotor system's failure mode impact and hazard analysis report. These include wear-related failures (uniform blade tip wear, seal tooth wear), fatigue propagation-related failures (blade root crack propagation, disk crack propagation), and gradual rubbing failures. These failure modes are then marked as those capable of dynamic evolution matching.

[0068] The second step is to construct a time-series evolution prototype vector for each selected failure mode. Taking uniform tip wear as an example, by reviewing the maintenance records and borehole inspection data of this rotor model, the wear state and corresponding clearance characteristics under different operating hours can be compiled. These features are arranged in chronological order of wear degree, forming a sequence of K stages. For each stage i, a feature vector V_i with the same dimension and definition as the standard failure feature vector is extracted. These vectors contain the typical residual effective value, kurtosis value, and energy proportion of each order harmonic of the wear stage. Arranging these K vectors in time sequence constitutes the time-series evolution prototype vector Sequence_Wear = [V_1, V_2,..., V_K] for this failure mode. For prototypes obtained through simulation, the gradual deterioration process of the failure is simulated on a high-fidelity finite element model. For example, the contact stiffness at the tenon is gradually released to simulate the intensification of loosening. A transient dynamic simulation is run once at each deterioration step, and feature vectors are extracted from the output clearance signal to form the prototype sequence.

[0069] The third step is to store the constructed time-series evolution prototype vectors in the fault feature database. The database table structure needs to be expanded; in addition to the existing table storing static fault feature vectors, a new time-series prototype table is added. Each record in this table corresponds to a fault mode, containing a fault mode name field and a large object field or structured array field for storing the vector sequence. Each feature vector in the sequence also includes a normalized time or fault severity label; for example, 0% represents a healthy state, 100% represents the fault reaching the alarm limit, with intermediate divisions based on linear or non-linear scales. The database is indexed to enable rapid retrieval and comparison during the subsequent online matching phase.

[0070] The fourth step is the logical preparation for online matching. During the initialization of the online diagnostic program, not only are the static fault feature vectors loaded into memory, but also all fault modes with temporal evolution prototype vectors and their complete sequences are loaded simultaneously. The computational interfaces for the dynamic time warping or sequence similarity algorithms to be used subsequently are defined, preparing for dynamic matching.

[0071] The technical principle of this solution lies in advancing fault diagnosis from single-point snapshot-style pattern recognition to process trajectory-style behavior recognition. Many mechanical faults, in their early stages, may exhibit gap residual characteristics that deviate from a healthy state but do not conform to the static vector description of any mature fault, resulting in low static matching scores and missed detections. By introducing a temporal evolution prototype, the system can identify which fault evolution path the feature sequence is following, thus capturing the fault in its very early stages. The technical effect is to achieve early trend warnings for progressive faults, enabling diagnostic reports to not only include the most likely fault type but also information on the current stage of development, providing a more forward-looking and detailed basis for predictive maintenance decisions.

[0072] In a preferred embodiment, after calculating its static matching score with each failure mode, the following step is further included: Based on the predicted uncertainty interval, the confidence coefficient of the static matching score is calculated; The calculation of a multi-dimensional fault contribution vector based on the static matching score and the prediction uncertainty interval includes the following steps: If the credibility coefficient is greater than or equal to the first preset threshold, a multi-dimensional fault contribution vector is calculated based on the static matching score and the prediction uncertainty interval.

[0073] Specifically, a key calculation and judgment step is introduced: before calculating the multi-dimensional fault contribution vector, the credibility of the current static matching score is assessed, and the branching direction of the subsequent diagnostic strategy is determined based on the credibility level. First, let's explain the core terminology. The credibility coefficient is a scalar index derived from the obtained prediction uncertainty interval, used to quantify the reliability of the entire prediction and feature extraction process at the current moment; its value range is typically mapped to 0 to 1. The first preset threshold is an empirical threshold value set by the system administrator based on the analysis of the subject operating characteristic curve of the system in the calibration experiment, used to distinguish between high and low credibility states; a typical setting value can be 0.7 or 0.8. If the credibility coefficient is greater than or equal to this threshold, it indicates that the current sensor signal is clear, the operating condition is stable, the model is in its preferred interpolation region, and the confidence of the diagnostic conclusion is very high; conversely, it indicates that there is significant uncertainty in the signal or model, requiring a more cautious diagnostic strategy.

[0074] The specific implementation steps are as follows.

[0075] The first step is to calculate the credibility coefficient of the static matching score. From the uncertainty quantification pipeline, the prediction variance σ²_res corresponding to the gap deviation residual at the current moment is obtained in real time. First, a basic credibility metric is calculated, mapped to a coefficient C_raw = exp(-α * σ²_res) that is inversely proportional to the variance using an exponential decay function. Here, α is a scaling factor, calibrated based on the average residual variance σ²_nominal under normal operating conditions on the training set, and is set to α = 1 / σ²_nominal. In addition, a condition coverage factor is introduced as a multiplier. The minimum Mahalanobis distance from the current condition point (a vector composed of speed, temperature, etc.) to the condition point cloud of the model training set is calculated to determine whether the model is extrapolating. If the distance is too large, an additional penalty factor less than 1 is applied. Finally, the credibility coefficient C_trust = C_raw * extrapolation penalty factor. A C_trust close to 1 indicates a highly reliable prediction, while a C_trust close to 0 indicates an extremely unreliable prediction.

[0076] The second step is to compare the credibility coefficient with a first preset threshold. After the online diagnostic program completes the static matching score calculation, it obtains a static matching score vector S_static. Before using S_static to further calculate the fault contribution vector, a branch decision statement is inserted. The conditional expression is: if C_trust>=Threshold_1, then execute branch A; otherwise, execute branch B. Branch A is the conventional and direct calculation path, i.e., adopting the current static matching result.

[0077] The third step, execution branch A: Based on the static matching score and the prediction uncertainty interval, directly calculate the multi-dimensional fault contribution vector. This step uses the confidence coefficient C_trust as a global scaling factor, multiplying it with the static matching score vector S_static to obtain the final fault contribution vector Contribution = C_trust * S_static. Because C_trust is relatively high at this point, this multiplication mainly fine-tunes the score, maintaining the dominance of the static matching result while assigning it a slightly adjusted high confidence score.

[0078] The fourth step is to generate a diagnostic report. Based on the fault contribution vector calculated from branch A, a diagnostic report is generated. The report will indicate that this diagnostic mode is a high-confidence static matching mode, and will provide the main suspected fault modes and their scores.

[0079] The technical principle of this solution lies in introducing an uncertainty gating mechanism at the decision-making level. In condition monitoring and fault diagnosis, data quality at all times does not support equally strong judgments. When the system is in a highly uncertain state, pattern matching results based on point features may have a high degree of randomness; a coincidental similarity may be entirely due to noise and prediction errors. By setting a confidence threshold, the system can clearly identify the current cognitive boundaries. Its technical effect is to ensure that within a window of excellent data quality and model status, it provides rapid, clear, and highly confident diagnostic conclusions. For industrial equipment that operates routinely for the majority of the time, this guarantees the timeliness and directness of the diagnosis, avoiding ambiguous conclusions even when everything is normal.

[0080] In a preferred embodiment, after calculating the credibility coefficient of the static matching score, the method further includes the following steps: If the confidence coefficient is less than the first preset threshold, then within a continuous monitoring time window, multiple sets of the real-time residual feature vectors are extracted and cached in sequence to form a residual feature evolution sequence. The residual feature evolution sequence is matched with the time-series evolution prototype vector in the fault feature database to calculate the dynamic matching degree score. The dynamic matching score and the static matching score are fused to update the fault contribution vector, and the diagnostic report is regenerated based on the updated fault contribution vector.

[0081] Specifically, the residual feature evolution sequence refers to a set of real-time residual feature vectors extracted and cached chronologically across multiple consecutive monitoring time windows. It is not a snapshot at a single moment, but rather a stream of video evidence recording how fault features dynamically change over time. The dynamic matching score is obtained by calculating the sequence-to-sequence similarity between the real-time extracted residual feature evolution sequence and the time-series evolution prototype vector of a fault mode pre-stored in the fault feature database. A commonly used algorithm is dynamic time warping, which can tolerate the nonlinear time scaling differences between the actual fault evolution rate and the prototype. Fusion refers to integrating this new trend-based evidence—the dynamic matching score—with the previously calculated but less reliable static matching score through weighted summation, Bayesian update, or other information fusion algorithms into an updated and more comprehensive fault contribution vector.

[0082] The specific implementation steps are as follows.

[0083] The first step is to monitor the execution of the window and feature cache. When the conditional judgment result indicates that the confidence coefficient C_trust is less than the first preset threshold Threshold_1, the system will not directly discard the current data or issue a report based on the low-confidence static result. Instead, it will enter a caching and observation mode. The system creates a fixed-length, first-in-first-out feature vector cache queue in memory. The queue length corresponds to a preset observation window duration, for example, covering the time span equivalent to 10% of the process in the early stage of fault evolution. In each subsequent sampling period, the system continues to extract real-time residual feature vectors and add them to the end of the cache queue. When the queue is full, it constitutes a residual feature evolution sequence S_real = [V_t, V_{t+1}, ..., V_{t+N}], where N is the queue length.

[0084] The second step involves calculating the dynamic matching score for sequence matching. For each fault mode j with a temporal evolution prototype vector in the fault feature database, its prototype sequence S_proto_j is extracted. The dynamic time warping algorithm is called to calculate the dynamic time warping distance DTW_distance_j between S_real and S_proto_j. The core of the dynamic time warping algorithm is to construct an N×M cumulative distance matrix, where N and M are the lengths of the sequence to be matched S_real and the prototype sequence S_proto_j, respectively. The element d(i, k) of the matrix is ​​the Euclidean distance between the i-th vector in S_real and the k-th vector in S_proto_j. Through dynamic programming, a curved path from the starting point to the ending point of the matrix that minimizes the cumulative distance is found; this minimum cumulative distance is DTW_distance_j. To convert this into a score, the dynamic matching score Score_DTW_j = exp(-β * DTW_distance_j / N) is defined and normalized to the interval 0 to 1. This score quantifies how similar the evolution trajectory of the real-time feature is to the typical evolution trajectory of fault j.

[0085] The third step is to fuse the static and dynamic scores to update the fault contribution vector. We now have two types of evidence: a static matching score vector S_static_low calculated under low-confidence conditions (with discounted confidence), and a dynamic matching score vector S_DTW calculated at this moment. A fusion weight w is set, its value dynamically determined based on the length of the cached sequence and signal quality; the longer the sequence and the more stable the dynamic score, the larger w becomes. The fusion formula is: Contribution_updated_j = (1 - w) * S_static_low_j + w * Score_DTW_j. This updated contribution vector integrates the shape information of a single-point state and the evolution trend information over a period of time.

[0086] The fourth step involves regenerating the diagnostic report based on the updated fault contribution vector. The system discards the initially generated low-confidence report and, based on the `Contribution_updated` vector, regenerates a new diagnostic report sorted by contribution from highest to lowest. This report specifically notes that the diagnostic mode is the dynamic evolution analysis mode and provides the top-ranked suspected fault modes, the updated confidence level, and the possible fault development stages. For example, the report might state that although the current operating condition has high uncertainty, based on the dynamic evolution pattern of continuously increasing residual kurtosis and gradually strengthening 0.5 octave components over the past 20 minutes, the system highly suspects that it is in the early stage of radial rubbing, and the confidence level is updated from low to above average.

[0087] The effect of this technology is to significantly enhance the accuracy and robustness of the diagnostic system under harsh operating conditions, strong noise interference, or early minor fault stages, and to expand the diagnostic perspective from isolated points in time to continuous time processes, thus achieving true predictive intelligent diagnosis.

[0088] The above description is merely a preferred embodiment of this application and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of the invention involved in this application is not limited to technical solutions formed by specific combinations of the above-described technical features, but should also cover other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the inventive concept. For example, technical solutions formed by substituting the above features with (but not limited to) technical features with similar functions disclosed in this application.

Claims

1. A method for analyzing rotor blade tip clearance data based on machine learning, characterized in that, Includes the following steps: Obtain the rotor's historical operating dataset, which includes tip clearance time data, rotational speed time data, and time data of at least one other operating condition parameter; Based on the rotor's structural dimensions and material thermophysical parameters, a simplified mechanism model of the rotor is established. Using the simplified mechanism model, the baseline value of the ideal gap change caused by centrifugal force deformation and thermal deformation is calculated in real time according to the rotational speed and other operating parameters. A physics-data hybrid driven machine learning model is constructed, wherein the machine learning model receives the rotational speed, the other operating condition parameters, and the ideal clearance variation baseline value as input, and fits the clearance deviation residuals in the historical operating dataset other than the ideal clearance variation baseline value through a residual learning branch, and adds the ideal clearance variation baseline value to the fitted clearance deviation residuals to obtain the blade tip clearance prediction value. The machine learning model is trained using the historical running dataset. During the training process, the internal parameters of the simplified mechanism model are kept fixed or only subject to restrictive fine-tuning. The parameters of the residual learning branch are updated with the goal of minimizing the error between the predicted tip clearance and the historical measured tip clearance. The trained machine learning model is deployed to the online monitoring system. Rotation speed and operating parameters are input to the machine learning model in real time to obtain the predicted value of the blade tip clearance online. The ideal clearance change baseline value and the clearance deviation residual are decoupled and output simultaneously. The clearance deviation residual is continuously monitored, and a dynamic control limit is set based on the statistical characteristics of the clearance deviation residual under normal conditions. When the amplitude or change trend of the clearance deviation residual exceeds the dynamic control limit, a structural abnormality warning signal is generated to indicate the change in rotor structural state, so as to distinguish it from normal clearance fluctuations caused by transient changes in operating conditions.

2. The rotor blade tip clearance data analysis method based on machine learning according to claim 1, characterized in that, Before training the machine learning model, the time-series data of the decoupled gap deviation residuals are subjected to signal transformation processing to construct an enhanced residual feature set. The signal transformation processing includes: taking the time-series data of the gap deviation residual as a time-domain signal, converting it into an order spectrum with rotational speed as the horizontal axis based on the synchronously acquired key phase signal, and / or converting the time-series data of the gap deviation residual into a time-spectrum graph. The feature representation of the transformed order spectrum or time spectrum is used as one of the input features of the residual learning branch.

3. The rotor blade tip clearance data analysis method based on machine learning according to claim 2, characterized in that, The residual learning branch is implemented using a dual-channel neural network, and includes the following steps: Construct a first channel, input the rotational speed, the other operating parameters, and the ideal clearance variation base value into the first channel, and extract the operating feature vector; Construct a second channel, which is a convolutional neural network. Input the feature representation of the order spectrum or time spectrum of the gap deviation residual into the second channel, and extract the residual spectrum feature vector. The fusion module fuses the operating condition feature vector with the residual spectrum feature vector, and outputs the predicted value of the gap deviation residual based on the fused features.

4. The rotor blade tip clearance data analysis method based on machine learning according to claim 1, characterized in that, The generation of the structural anomaly early warning signal for indicating changes in the rotor structural state includes the following steps: Statistical analysis is performed on the gap deviation residuals output by online decoupling, and dynamic control limits are constructed based on the mean and standard deviation of the gap deviation residuals under historical normal conditions. When the instantaneous amplitude of the gap deviation residual exceeds the dynamic control limit, or the trend slope within its sliding window exceeds a preset threshold, the structural anomaly warning signal is generated. The time-domain and frequency-domain patterns of the gap deviation residuals are identified, and the identified patterns are compared with a preset fault residual pattern library to provide an indication of the associated potential fault type while generating the structural anomaly warning signal.

5. The rotor blade tip clearance data analysis method based on machine learning according to claim 3, characterized in that, The method further includes a step of quantifying the uncertainty of the predicted blade tip clearance value: Obtain the ideal gap variation base value output by the simplified mechanism model, and calculate the prior distribution of the ideal gap variation base value based on the perturbation range of the material thermophysical parameters of the rotor; The residual learning branch outputs the predicted value of the gap deviation residual, and at the same time outputs the residual prediction variance of the predicted value; The prior distribution of the ideal gap variation baseline value is fused with the predicted value of the gap deviation residual and its predicted variance using Bayesian fusion to obtain the posterior distribution of the predicted blade tip gap value. Wherein, the mean of the posterior distribution is the final predicted tip clearance value, and the variance interval of the posterior distribution is the prediction uncertainty interval.

6. The rotor blade tip clearance data analysis method based on machine learning according to claim 5, characterized in that, The generation of the structural anomaly early warning signal for indicating changes in the rotor structural state includes the following steps: Obtain the predicted value of the gap deviation residual and the corresponding portion of the residual in the prediction uncertainty interval; The degree to which the gap deviation residual obtained from online decoupling deviates from its predicted value is calculated and used as the standardized residual deviation. Based on the magnitude of the residual prediction variance, the adaptive warning threshold for the current time is dynamically calculated and set; wherein, the larger the residual prediction variance, the higher the adaptive warning threshold is set. The standardized residual deviation is compared with the adaptive early warning threshold. If the standardized residual deviation continues to exceed the adaptive early warning threshold, a structural anomaly early warning signal based on uncertainty perception is generated.

7. The rotor blade tip clearance data analysis method based on machine learning according to claim 5, characterized in that, The method further includes the following steps: Retrieve the fault feature database, which contains a mapping relationship between multiple preset fault modes and a set of standard fault feature vectors. The standard fault feature vector corresponding to each fault mode is described by the time domain statistical features and frequency domain features of the gap deviation residual under the corresponding fault state. For real-time data, extract the real-time residual feature vector from the gap deviation residual output by online decoupling; The real-time residual feature vector is matched with all standard fault feature vectors in the fault feature database, and its static matching degree score with each fault mode is calculated. Based on the static matching score and the prediction uncertainty interval, a multi-dimensional fault contribution vector is calculated, where each dimension corresponds to a probability score of a fault mode. By combining the fault contribution vector with the structural anomaly warning signal, a diagnostic report is generated that includes the specific suspected fault mode, severity, and confidence level.

8. The rotor blade tip clearance data analysis method based on machine learning according to claim 7, characterized in that, The fault feature database has corresponding time-series evolution prototype vectors for some fault modes. These time-series evolution prototype vectors describe the typical pattern of the evolution of the characteristics of the gap deviation residuals under the corresponding fault modes over time.

9. The rotor blade tip clearance data analysis method based on machine learning according to claim 8, characterized in that, After calculating its static fit score with each failure mode, the following steps are also included: Based on the predicted uncertainty interval, the confidence coefficient of the static matching score is calculated; The calculation of a multi-dimensional fault contribution vector based on the static matching score and the prediction uncertainty interval includes the following steps: If the credibility coefficient is greater than or equal to the first preset threshold, a multi-dimensional fault contribution vector is calculated based on the static matching score and the prediction uncertainty interval.

10. The rotor blade tip clearance data analysis method based on machine learning according to claim 9, characterized in that, After calculating the credibility coefficient of the static matching score, the method further includes the following steps: If the confidence coefficient is less than the first preset threshold, then within a continuous monitoring time window, multiple sets of the real-time residual feature vectors are extracted and cached in sequence to form a residual feature evolution sequence. The residual feature evolution sequence is matched with the time-series evolution prototype vector in the fault feature database to calculate the dynamic matching degree score. The dynamic matching score and the static matching score are fused to update the fault contribution vector, and the diagnostic report is regenerated based on the updated fault contribution vector.

Images