A gas turbine fault diagnosis method based on a multi-modal large model

Abstract

The application provides a kind of gas turbine fault diagnosis method based on multi-modal large model, belongs to gas turbine fault diagnosis technical field, the present application carries out wavelet packet decomposition to acoustic emission signal and extracts frequency band energy feature, carries out fourier transform to vibration acceleration signal and extracts amplitude spectrum and phase spectrum, extracts temperature gradient distribution feature to thermogram sequence and organizes as four-order tensor, adopts tensor ring decomposition and decomposes four-order tensor into ring product structure of time kernel tensor frequency kernel tensor space kernel tensor and mode kernel tensor, extracts principal component of time kernel tensor, principal component of frequency kernel tensor and coupling coefficient of mode kernel tensor as multi-modal fusion feature vector, input fault recognition model to carry out fault type discrimination and adjust operating parameter according to residual life prediction value, solve the technical problem that high-order coupling feature of multi-modal sensor data is difficult to effectively extract, resulting in insufficient accuracy of gas turbine fault diagnosis.

Description

Technical Field

[0001] This invention belongs to the field of gas turbine fault diagnosis technology, and more specifically, relates to a gas turbine fault diagnosis method based on a multimodal large model. Background Technology

[0002] Gas turbine fault diagnosis technology identifies faults by collecting multimodal data such as acoustic emission signals, vibration signals, and thermograms. Traditional methods use a simple concatenation of extracted features to construct feature vectors for inputting into a classifier. However, existing technologies suffer from complex coupling relationships between different physical modes (acoustic emission, vibration, and thermograms) in terms of temporal evolution and frequency distribution. Traditional independent processing methods lose high-order interaction information between modes, resulting in extracted feature vectors that cannot fully characterize the essential characteristics of the fault state. Existing methods often employ principal component analysis or canonical correlation analysis for multimodal fusion, but these linear methods struggle to capture nonlinear coupling relationships and lack sufficient feature discrimination capability when dealing with complex fault modes such as compressor surge and turbine blade cracks. In other words, existing technologies suffer from the technical problem of insufficient accuracy in gas turbine fault diagnosis due to the difficulty in effectively extracting high-order coupling features from multimodal sensor data. Summary of the Invention

[0003] In view of this, the present invention provides a gas turbine fault diagnosis method based on a multimodal large model, which can solve the technical problem that the high-order coupling features of multimodal sensor data are difficult to extract effectively, resulting in insufficient accuracy of gas turbine fault diagnosis.

[0004] This invention is implemented as follows: It provides a gas turbine fault diagnosis method based on a multimodal large model. This method acquires acoustic emission signals, vibration acceleration signals, and thermographic sequences during gas turbine operation. Wavelet packet decomposition is performed on the acoustic emission signals to extract frequency band energy features. Fourier transform is performed on the vibration acceleration signals to extract amplitude and phase spectra. Temperature gradient distribution features are extracted from the thermographic sequences. The frequency band energy features (amplitude and phase spectra) and temperature gradient distribution features are organized into fourth-order tensors. Tensor ring decomposition is performed on these fourth-order tensors, decomposing them into a ring-shaped product structure of time kernel tensors, frequency kernel tensors, spatial kernel tensors, and modal kernel tensors. Principal components of the time kernel tensor, principal components of the frequency kernel tensor, and coupling coefficients of the modal kernel tensors are extracted as multimodal fusion feature vectors. These multimodal fusion feature vectors are input into a fault identification model for fault type discrimination. For gas turbines identified as faulty, a fractional-order performance degradation differential equation model is established to predict remaining life, and the fault type discrimination result and remaining life are output.

[0005] The wavelet packet decomposition uses the Daubechies wavelet basis function, with a decomposition layer of 3. After each layer of decomposition, the energy of each sub-band signal is calculated and normalized to obtain the energy characteristics of the frequency band.

[0006] The Fourier transform uses the Fast Fourier Transform algorithm with 2048 transform points, extracting the amplitude spectrum and phase spectrum in the frequency range of 0Hz to 25kHz as frequency domain features.

[0007] The temperature gradient distribution feature is obtained by calculating the temperature difference between each pixel and its eight neighboring pixels in the thermal image sequence. The absolute value of the temperature difference is taken and max pooling is performed in the spatial dimension to obtain the dimensionality-reduced temperature gradient distribution feature.

[0008] The fourth-order tensor has the following dimensions: time step multiplied by frequency dimension multiplied by number of sensors multiplied by number of modes. The time step is 1000 consecutively acquired time points, the frequency dimension has 128 discrete frequency components, the number of sensors is 16 sensor positions, and the number of modes is three physical modes: acoustic emission vibration and thermography.

[0009] Tensor ring decomposition decomposes the fourth-order tensor into a ring-connected sequence of kernel tensors. Each kernel tensor is a third-order tensor. Adjacent kernel tensors are connected through tensor contraction operations to form a closed-loop structure. The parameters of each kernel tensor are optimized through alternating least squares iterations until the reconstruction error is less than 0.001.

[0010] The coupling coefficient of the modal kernel tensor characterizes the interaction strength between different physical modes. The coupling coefficient is obtained by performing singular value decomposition on the modal kernel tensor along the modal dimension. The magnitude of the singular values ​​in the singular value decomposition reflects the contribution of the corresponding mode combination.

[0011] The input layer of the fault identification model receives a 256-dimensional multimodal fusion feature vector, which is then mapped to a 512-dimensional latent space through a feature embedding layer. After the feature embedding layer, four cross-modal attention modules are connected. The cross-modal attention modules include a multi-head self-attention mechanism and a feedforward neural network.

[0012] Among them, the fault identification model utilizes a low-rank approximation optimization mechanism based on matrix factorization to decompose the weight matrix in the cross-modal attention module into the product of the query low-rank matrix, the key low-rank matrix, and the value low-rank matrix, and introduces a nuclear norm regularization term to constrain the sum of the singular values ​​of the query low-rank matrix, the key low-rank matrix, and the value low-rank matrix.

[0013] Among them, the fault identification model utilizes a sparse connection learning framework based on dynamic topology reconstruction. During training, it calculates the mutual information between neuron pairs to evaluate the information transmission contribution of the connection. When the mutual information of the connection is lower than the dynamic threshold, the connection is pruned, and highly important connections are regrown at the pruned connection positions according to the gradient magnitude.

[0014] In the cross-modal attention module, the number of heads for the multi-head self-attention mechanism is determined based on the coupling coefficient of the modal kernel tensor. When the ratio of the maximum to the minimum coupling coefficient of the modal kernel tensor is greater than 5, the number of heads is set to 16. When the ratio of the maximum to the minimum is between 2 and 5, the number of heads is set to 8. When the ratio of the maximum to the minimum is less than 2, the number of heads is set to 4.

[0015] The fault identification model outputs the probability distribution of each fault type. When the maximum probability value exceeds 0.85, it is determined to be the corresponding fault type. When the maximum probability value is lower than 0.85 but higher than 0.60, the secondary diagnostic process is initiated. When the maximum probability value is lower than 0.60, it is determined to be in normal operating condition.

[0016] The secondary diagnostic process includes increasing the sampling time window to three times the original time window to re-acquire signals and extracting the resampled multimodal fusion feature vector. The resampled multimodal fusion feature vector is then concatenated with the multimodal fusion feature vector to form an extended feature vector, which is input into the fault identification model for further discrimination.

[0017] The fractional order of the fractional performance degradation differential equation model is identified from historical monitoring data using the particle swarm optimization algorithm, and the remaining lifetime prediction is obtained by solving the fractional performance degradation differential equation model using the Green's function method.

[0018] Among them, the particle swarm algorithm initializes the particle swarm to be randomly distributed in the fractional order and model parameter space, calculates the mean square error between the model output corresponding to each particle and the historical monitoring data as the fitness function, and updates the particle velocity and position iteratively searches according to the historical best position and the global best position of the particle until the fitness function converges.

[0019] Among them, adjusting the operating parameters of the gas turbine according to the remaining life prediction value means that when the remaining life prediction value is greater than 500 hours, the current power output is maintained; when the remaining life prediction value is between 200 hours and 500 hours, the power output is reduced to 80% of the rated power; and when the remaining life prediction value is less than 200 hours, the shutdown protection action is triggered.

[0020] This invention organizes multimodal sensor data into fourth-order tensors using tensor ring decomposition technology, further decomposing them into a ring-shaped product structure of time kernel tensors, frequency kernel tensors, spatial kernel tensors, and modal kernel tensors. This preserves the high-order coupling relationships between time, frequency, and space modes, extracting compact feature representations rich in fault discrimination information, thus overcoming the deficiency of traditional independent feature extraction methods that lose intermodal interaction information. The ring-shaped structure of tensor ring decomposition reduces the storage complexity of high-order tensors from exponential to polynomial level, avoiding the problem of the core tensor dimension growing exponentially with order in traditional Tucker decomposition, enabling the algorithm to handle large-scale multimodal data. By extracting the coupling coefficients of the modal kernel tensors, the interaction strength between different physical modes is quantified, and the constructed multimodal fusion feature vector fully characterizes the multidimensional physical mechanism characteristics of faults, improving the diagnostic accuracy of the fault identification model under small sample conditions. In summary, this invention solves the technical problem mentioned in the background art, which is that the high-order coupling features of multimodal sensor data are difficult to extract effectively, leading to insufficient accuracy in gas turbine fault diagnosis. Attached Figure Description

[0021] Figure 1 This is a flowchart of the method of the present invention.

[0022] Figure 2 This is a schematic diagram of the sensor arrangement and multimodal signal acquisition for a gas turbine.

[0023] Figure 3 This is a flowchart of the tensor structure and feature extraction process for tensor ring decomposition kernel tensor.

[0024] Figure 4 The time-frequency spatial distribution diagram of multimodal characteristics of turbine blade crack faults.

[0025] Figure 5 The curves for fitting the fractional-order performance degradation model and predicting the remaining lifetime are shown. Detailed Implementation

[0026] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below.

[0027] like Figure 1 The diagram shown is a flowchart of a gas turbine fault diagnosis method based on a multimodal large model provided by the present invention. This method includes the following steps:

[0028] S01. Acquire acoustic emission signals, vibration acceleration signals and thermal image sequences during the operation of the gas turbine, wherein the acoustic emission sensor sampling frequency is 500kHz, the vibration sensor sampling frequency is 50kHz, and the thermal imager frame rate is 60Hz. The acoustic emission sensor is arranged at the root of the turbine blade and the combustion chamber wall, and the vibration sensor is arranged at the bearing housing and the casing shell.

[0029] S02. Perform wavelet packet decomposition on the acoustic emission signal to extract energy features of 8 frequency bands, perform Fourier transform on the vibration acceleration signal to extract amplitude spectrum and phase spectrum, extract temperature gradient distribution features from the thermal image sequence, and organize the energy features of the 8 frequency bands, the amplitude spectrum, the phase spectrum and the temperature gradient distribution features into a fourth-order tensor.

[0030] S03. Perform tensor ring decomposition on the fourth-order tensor, decomposing the fourth-order tensor into a ring-shaped product structure of time kernel tensor, frequency kernel tensor, spatial kernel tensor and modal kernel tensor. Optimize the parameters of each kernel tensor through alternating least squares iteration until the reconstruction error is less than 0.001. Extract the first principal component of the time kernel tensor, the first three principal components of the frequency kernel tensor and the coupling coefficient of the modal kernel tensor as multimodal fusion feature vectors.

[0031] S04. Input the multimodal fusion feature vector into the fault identification model to determine the fault type. The fault identification model outputs the probability distribution of each fault type. When the maximum probability value exceeds 0.85, it is determined to be the corresponding fault type. When the maximum probability value is lower than 0.85 but higher than 0.60, the secondary diagnosis process is started. When the maximum probability value is lower than 0.60, it is determined to be in normal operation state.

[0032] S05. Establish a fractional-order performance degradation differential equation model for the gas turbine that is determined to be in a fault state. The fractional order of the fractional-order performance degradation differential equation model is identified from historical monitoring data by the particle swarm algorithm. The remaining life prediction value is obtained by solving the fractional-order performance degradation differential equation model using the Green's function method.

[0033] S06. Adjust the operating parameters of the gas turbine according to the remaining life prediction value. When the remaining life prediction value is greater than 500 hours, maintain the current power output. When the remaining life prediction value is between 200 hours and 500 hours, reduce the power output to 80% of the rated power. When the remaining life prediction value is less than 200 hours, trigger the shutdown protection action and issue a maintenance warning signal.

[0034] The wavelet packet decomposition uses the Daubechies wavelet basis function with three decomposition layers. After each layer of decomposition, the energy of each sub-band signal is calculated and normalized to obtain the energy features of the eight frequency bands. The Fourier transform uses the Fast Fourier Transform algorithm with 2048 transform points, and the amplitude spectrum and phase spectrum in the frequency range of 0Hz to 25kHz are extracted as frequency domain features.

[0035] The temperature gradient distribution feature is obtained by calculating the temperature difference between each pixel and its eight neighboring pixels in the thermal image sequence. The absolute value of the temperature difference is taken and max pooling is performed in the spatial dimension. The pooling window size is 8×8 pixels to obtain the dimensionality-reduced temperature gradient distribution feature.

[0036] The fourth-order tensor has the following dimensions: time step × frequency dimension × number of sensors × number of modes. The time step is 1000 continuously acquired time points, the frequency dimension has 128 discrete frequency components, the number of sensors is 16 sensor positions, and the number of modes includes three physical modes: acoustic emission, vibration, and thermal imaging.

[0037] The tensor ring decomposition decomposes the fourth-order tensor into a ring-connected sequence of kernel tensors, each of which is a third-order tensor. Adjacent kernel tensors are connected through tensor contraction operations to form a closed-loop structure. The tensor ring decomposition preserves the multilinear structure of the original tensor and achieves data compression. The alternating least squares iterative optimization process fixes the parameters of other kernel tensors and optimizes only the parameters of the current kernel tensor, cyclically updating all kernel tensors until the objective function converges.

[0038] Compared to traditional Tucker decomposition, the proposed tensor ring decomposition avoids the problem of the core tensor dimension growing exponentially with the order. By using a ring structure, the storage complexity of high-order tensors is reduced from exponential to polynomial level, while preserving the high-order coupling relationship between time-frequency and space-time modes. The extracted compact feature representation contains richer fault discrimination information, reduces the training data requirements of subsequent classifiers, and improves the diagnostic accuracy under small sample conditions.

[0039] The coupling coefficient of the modal kernel tensor characterizes the interaction strength between different physical modes. The coupling coefficient is obtained by performing singular value decomposition on the modal kernel tensor along the modal dimension. The magnitude of the singular values ​​in the singular value decomposition reflects the contribution of the corresponding mode combination.

[0040] The specific structure of the fault identification model is as follows: the input layer receives the 256-dimensional multimodal fusion feature vector, and the feature embedding layer maps the multimodal fusion feature vector to a 512-dimensional latent space. After the feature embedding layer, four cross-modal attention modules are connected. Each cross-modal attention module includes a multi-head self-attention mechanism and a feedforward neural network. The multi-head self-attention mechanism has 8 heads, and the hidden layer dimension of the feedforward neural network is 2048. After the cross-modal attention modules, a global average pooling layer is connected. The output of the global average pooling layer is mapped to the dimension of the number of fault types through two fully connected layers. Finally, the probability distribution of each fault type is output through the Softmax activation function.

[0041] The steps for establishing the training dataset of the fault identification model specifically include collecting normal operation data and various fault data, using a digital twin system based on physical simulation to generate synthetic fault samples to expand scarce fault categories, downsampling normal samples to make the ratio of normal to fault samples reach 3:1, extracting the multimodal fusion feature vector from all samples and labeling the fault type, and dividing the dataset into training set and validation set according to an 8:2 ratio.

[0042] The specific steps for training the fault identification model include initializing model parameters using a Xavier uniform distribution, setting the learning rate to 0.0001 and employing the Adam optimizer, setting the batch size to 32, using the cross-entropy loss function combined with label smoothing regularization during training, evaluating the accuracy on the validation set every 10 training cycles, reducing the learning rate to one-tenth of the original value when the validation set accuracy does not improve for 5 consecutive cycles, and terminating training and saving the optimal model parameters when the validation set accuracy does not improve for 15 consecutive cycles.

[0043] The fault identification model utilizes a low-rank approximation optimization mechanism based on matrix factorization to decompose the weight matrix in the cross-modal attention module into a product of a query low-rank matrix, a key low-rank matrix, and a value low-rank matrix. Singular value decomposition determines that the rank of each low-rank matrix is ​​one-quarter of the original matrix rank. A nuclear norm regularization term is introduced to constrain the sum of the singular values ​​of the query low-rank matrix, the key low-rank matrix, and the value low-rank matrix. An alternating minimization strategy is used to optimize the current low-rank matrix parameters while fixing other matrix parameters. All low-rank matrices are iteratively updated until the reconstruction error stabilizes within 0.01.

[0044] The matrix factorization-based low-rank approximation optimization mechanism significantly reduces the number of model parameters and computational complexity by decomposing the high-dimensional weight matrix into the product of multiple low-dimensional matrices. This reduces the number of floating-point operations in a single forward propagation and lowers memory usage. Meanwhile, the kernel norm regularization constraint ensures that the decomposed query low-rank matrix, key low-rank matrix, and value low-rank matrix retain the main information of the original matrix, avoiding performance loss due to over-compression. The alternating minimization strategy ensures that the parameters of each low-rank matrix are collaboratively optimized to reach a globally approximate optimal solution. The matrix factorization-based low-rank approximation optimization mechanism achieves lightweight model deployment while maintaining fault diagnosis accuracy, meeting the computing power constraints of edge computing devices and accelerating inference speed to meet the millisecond-level fault response requirements of gas turbines.

[0045] The fault identification model utilizes a sparse connection learning framework based on dynamic topology reconstruction. During training, it statistically analyzes the activation frequency and intensity of each neuron across all training samples, calculates the mutual information between neuron pairs to evaluate the information transmission contribution of the connection, and prunes the connection in a probabilistic manner when the mutual information of the connection is lower than a dynamic threshold. The dynamic threshold is adaptively adjusted based on the current number of training epochs and the accuracy of the validation set. At the same time, L1 regularization and group Lasso regularization of the connection weights are introduced to promote sparsity. For the pruned connection positions, highly important connections are regrown based on the gradient magnitude and the curvature of the loss function.

[0046] The sparse connection learning framework based on dynamic topology reconstruction dynamically optimizes the network topology by adaptively adding and deleting neurons. It uses information theory criteria to identify redundant connections and prune them to reduce network density and computational overhead. The grouped Lasso regularization constraint ensures that the pruning process does not destroy the network's hierarchical feature extraction ability and maintains the model's expressive power. The connection regeneration mechanism avoids irreversible information loss caused by over-pruning and explores new feature propagation paths. During training, the sparse connection learning framework based on dynamic topology reconstruction gradually evolves the dense network into a sparse network, achieving model compression while improving generalization performance. The sparse network structure reduces the number of memory accesses and multiply-accumulate operations during inference, accelerating the diagnosis speed. The dynamically adjusted sparse mode enables the fault identification model to adapt to the feature distribution of different fault modes, improving the diagnostic robustness for new models of units and undetected faults.

[0047] In the cross-modal attention module, the number of heads in the multi-head self-attention mechanism is determined based on the coupling coefficient of the modal kernel tensor. When the ratio of the maximum to the minimum coupling coefficient of the modal kernel tensor is greater than 5, the number of heads in the multi-head self-attention mechanism is set to 16 to capture fine-grained interactions between strongly coupled modalities. When the ratio of the maximum to the minimum coupling coefficient of the modal kernel tensor is between 2 and 5, the number of heads in the multi-head self-attention mechanism is set to 8 to balance computational efficiency and expressive power. When the ratio of the maximum to the minimum coupling coefficient of the modal kernel tensor is less than 2, the number of heads in the multi-head self-attention mechanism is set to 4 to reduce computational complexity.

[0048] The secondary diagnostic process includes: increasing the sampling time window to three times the original time window; re-acquiring the acoustic emission signal, the vibration acceleration signal, and the thermogram sequence; performing wavelet packet decomposition on the re-acquisition acoustic emission signal to extract resampled frequency band energy features; performing Fourier transform on the re-acquisition vibration acceleration signal to extract resampled amplitude and phase spectra; extracting resampled temperature gradient distribution features from the re-acquisitioned thermogram sequence; and combining the resampled frequency band energy features, the resampled amplitude spectrum, and the resampled phase spectrum. The position spectrum and the resampled temperature gradient distribution features are organized into a resampled fourth-order tensor. The resampled fourth-order tensor is decomposed into tensor loops to extract a resampled multimodal fusion feature vector. The resampled multimodal fusion feature vector is concatenated with the multimodal fusion feature vector to form an extended feature vector. The extended feature vector is input into the fault identification model for re-discrimination. When the maximum probability value of the re-discrimination exceeds 0.75, it is determined to be the corresponding fault type. When the maximum probability value of the re-discrimination is lower than 0.75, it is determined to be a suspected fault state and manual diagnostic intervention is triggered.

[0049] Among them, the types of gas turbine failures include compressor surge failure, turbine blade crack failure, combustion chamber hot spot failure, bearing wear failure, and seal leakage failure.

[0050] The compressor surge fault manifests as periodic large fluctuations in the intake air flow, leading to violent oscillations in the compressor outlet pressure. The acoustic emission signal exhibits low-frequency, high-amplitude impact characteristics. The vibration acceleration signal shows a significant peak at the subharmonic of the rotational frequency. The thermal image sequence shows that the compressor inlet temperature distribution is uneven and there are local high-temperature areas.

[0051] Turbine blade crack failure is caused by thermal stress and centrifugal stress under high temperature and high pressure environment, which leads to material fatigue crack initiation and propagation. The acoustic emission signal generates continuous high-frequency burst signals during crack propagation. The high-frequency component of the vibration acceleration signal increases in energy and is accompanied by modulation. The thermal image sequence shows abnormal temperature gradient and concentrated heat flux density at the crack location.

[0052] Combustion chamber hot spot faults are caused by uneven fuel injection or failure of cooling film, resulting in local overheating. The acoustic emission signal is characterized by a wideband energy increase in the combustion noise spectrum. The vibration acceleration signal is affected by thermal deformation and produces low-frequency drift. The thermal image sequence shows that there are obvious high-temperature spots in the combustion chamber outlet temperature field and the temperature peak exceeds the design limit by more than 50°C.

[0053] Bearing wear failure is caused by poor lubrication or particulate contamination, resulting in material loss on the rolling elements and raceway surfaces. The acoustic emission signal contains the bearing characteristic frequency and its harmonics. The vibration acceleration signal forms periodic impact pulses at the bearing passing frequency. The thermal image sequence reflects a continuous increase in bearing housing temperature and a temperature rise rate greater than three times that under normal operating conditions.

[0054] The sealing leakage fault is caused by the wear of the sealing ring or excessive installation gap, resulting in the leakage of high-pressure gas to the low-pressure side. The acoustic emission signal exhibits broadband noise characteristics generated by airflow turbulence. The random component of the vibration acceleration signal is enhanced and the root mean square value of the amplitude increases. The thermal image sequence shows an abnormal temperature gradient and a high-temperature airflow trajectory at the location of the leakage path.

[0055] The fractional-order performance degradation differential equation model describes the degradation evolution of gas turbine performance parameters over time. The order of the fractional derivative in the fractional-order performance degradation differential equation model reflects the memory effect and nonlocal interaction characteristics of the degradation process.

[0056] The particle swarm optimization algorithm initializes a swarm of particles and randomly distributes them in the fractional order and model parameter space. Each particle represents a set of candidate parameters. The mean square error between the model output corresponding to each particle and the historical monitoring data is calculated as the fitness function. The particle velocity and position are updated according to the historical best position and the global best position of the particles. The search is iterated until the fitness function converges to the minimum value, thus obtaining the optimal fractional order and model parameters.

[0057] The Green's function method for solving the fractional-order performance degradation differential equation model transforms the time-domain problem into a frequency-domain problem, uses the Laplace transform to convert the fractional-order derivative into an algebraic equation, solves the algebraic equation to obtain the frequency-domain solution, and then uses the inverse Laplace transform to obtain the predicted remaining lifetime value in the time domain.

[0058] Compared to integer-order models, the fractional-order performance degradation differential equation model can describe the long-range dependence of material aging and performance degradation. The non-integer values ​​of the fractional-order order characterize the nonlinear variation of the degradation rate over time, capturing the slow evolution in the early degradation stage and the dynamic process of accelerated failure in the later stage. This provides more accurate remaining lifetime predictions to support predictive maintenance decisions, avoiding economic losses caused by premature downtime and the risk of catastrophic failures due to delayed maintenance.

[0059] The historical monitoring data includes the acoustic emission signals, vibration acceleration signals, and thermal image sequences collected over the past six consecutive months, as well as the power output, exhaust temperature, compressor efficiency, and turbine inlet temperature at the corresponding times, as observed values ​​of performance parameter degradation.

[0060] Alternatively, the present invention also provides a gas turbine fault diagnosis system implemented by a computer, wherein the computer is provided with a readable storage medium, the readable storage medium storing program instructions, and the program instructions executing the above-described gas turbine fault diagnosis method based on a multimodal large model when running in the computer.

[0061] The specific implementation methods of the above steps are described in detail below.

[0062] The specific implementation of step S01 involves installing acoustic emission sensors at the root of the turbine blades and on the combustion chamber wall of the gas turbine, respectively; installing vibration sensors on the bearing housing and casing; and installing a thermal imager outside the gas turbine. The acoustic emission sensors are broadband sensors made of piezoelectric ceramic material with a frequency response range of 100kHz to 600kHz, and the sampling frequency is set to 500kHz to meet the requirements of the Nyquist sampling theorem. The vibration sensors are accelerometers with a sensitivity of 100mV / g and a measurement range of ±50g, and the sampling frequency is set to 50kHz to capture the high-frequency vibration characteristics of the turbine rotor. The thermal imager is an uncooled focal plane type. The surface detector has a temperature resolution of 0.05℃ and a frame rate of 60Hz to record the dynamic changes in the temperature field. The acoustic emission sensor is in close contact with the blade root and combustion chamber wall through a coupling agent to reduce signal transmission loss. The vibration sensor is fixed to the bearing housing and casing surface by bolts to ensure accurate transmission of vibration signals. The optical axis of the thermal imager is perpendicular to the gas turbine casing surface to obtain a clear image of the temperature distribution. The signal output terminals of all sensors are connected to the data acquisition system, which converts analog signals into digital signals and stores them in the computer hard drive. The continuous acquisition time is 10 minutes to obtain sufficient sample data for subsequent analysis.

[0063] The specific implementation of step S02 involves using a three-level wavelet packet decomposition on the acquired acoustic emission signal. The Daubechies wavelet basis function is selected as the decomposition basis function. The Daubechies wavelet basis function possesses compact support and orthogonality, providing good localization characteristics in both the time and frequency domains. The first level of decomposition decomposes the acoustic emission signal into low-frequency approximation coefficients and high-frequency detail coefficients. The second and third levels of decomposition continue to decompose the coefficients of the previous level, ultimately obtaining eight sub-band signals of different frequency bands. The energy value of each sub-band signal is calculated by integrating the sum of the squares of the sub-band signals. The eight energy values ​​are normalized so that their sum equals 1, yielding the energy characteristics of the eight frequency bands. A Fast Fourier Transform (FFT) algorithm is applied to the vibration acceleration signal, converting the time-domain signal to a frequency-domain signal. The number of transformation points is set to 2048, resulting in 1024 frequency points. The corresponding amplitude and phase spectra are extracted from the frequency range of 0Hz to 25kHz as frequency domain features. The temperature gradient distribution features are calculated for each frame of the thermal image sequence. The temperature difference between each pixel and its eight neighboring pixels is calculated. The absolute value of the temperature difference is taken to obtain the gradient amplitude. Max pooling is performed using an 8×8 pixel window. The maximum gradient amplitude within the window is selected as the output value to reduce the spatial dimension of the temperature gradient distribution features. The energy features, amplitude spectrum, phase spectrum, and temperature gradient distribution features of the eight frequency bands are organized into a fourth-order tensor according to four dimensions: time step, frequency dimension, number of sensors, and number of modes. The first dimension of the fourth-order tensor corresponds to 1000 time points, the second dimension corresponds to 128 discrete frequency components, the third dimension corresponds to 16 sensor positions, and the fourth dimension corresponds to three physical modes: acoustic emission, vibration, and thermal imaging.

[0064] The specific implementation of step S03 involves performing tensor ring decomposition on the fourth-order tensor. Tensor ring decomposition represents the higher-order tensor as a ring product of a series of third-order kernel tensors. Each kernel tensor captures local structural information of the tensor along different dimensions. The elements of the time kernel tensor, frequency kernel tensor, spatial kernel tensor, and modal kernel tensor are initialized as random numbers following a standard normal distribution. The alternating least squares algorithm is used to iteratively optimize the parameters of each kernel tensor. In each iteration, the alternating least squares algorithm fixes the parameters of other kernel tensors and optimizes only the parameters of the current kernel tensor. The kernel tensor parameters are updated by minimizing the squared Frobenius norm between the original fourth-order tensor and the reconstructed fourth-order tensor. The squared Frobenius norm measures the difference between the two tensors. The iteration process continues until the reconstruction error is less than 0.001 or the maximum number of iterations (100) is reached. Singular value decomposition (SVD) is performed on the time kernel tensor along the time dimension, and the left singular vector corresponding to the largest singular value is extracted as the first principal component. Singular value decomposition is performed on the frequency kernel tensor along the frequency dimension, and the left singular vectors corresponding to the first three largest singular values ​​are extracted as the first three principal components. Singular value decomposition is performed on the modality kernel tensor along the modality dimension, and the singular values ​​are extracted as coupling coefficients. These coupling coefficients reflect the interaction strength between different physical modes. The first principal component of the time kernel tensor, the first three principal components of the frequency kernel tensor, and the coupling coefficients of the modality kernel tensor are sequentially concatenated to form a 256-dimensional multimodal fusion feature vector.

[0065] The specific implementation of step S04 involves inputting the multimodal fusion feature vector into the fault identification model. The input layer of the fault identification model receives a 256-dimensional feature vector. The feature embedding layer contains a fully connected layer with a weight matrix dimension of 256×512 and a bias vector dimension of 512. The ReLU activation function is used to perform a nonlinear transformation on the output of the fully connected layer to obtain a 512-dimensional latent space representation. Four layers of cross-modal attention modules process the latent space representation sequentially. The multi-head self-attention mechanism in each layer of the cross-modal attention module divides the input features into 8 heads, each with a dimension of 64. The query matrix, key matrix, and value matrix are calculated. The query matrix is ​​multiplied by the transpose of the key matrix to obtain the attention weights. The attention weights are normalized by Softmax and then multiplied by the value matrix to obtain the attention output. The attention outputs of the 8 heads are concatenated and subjected to a linear transformation to obtain the final output of the multi-head self-attention mechanism. The feedforward neural network contains two fully connected layers. The first fully connected layer converts the 512-dimensional feature vector into a 256×512-dimensional feature vector. The input is mapped to 2048 dimensions. The second fully connected layer maps the 2048 dimensions back to 512 dimensions. The ReLU activation function is used between the two fully connected layers. The global average pooling layer averages the output of the cross-modal attention module along the feature dimension to obtain a 512-dimensional vector. The two fully connected layers map the 512-dimensional vector to 5 dimensions, which correspond to five fault types: compressor surge fault, turbine blade crack fault, combustion chamber hot spot fault, bearing wear fault, and seal leakage fault. The Softmax activation function converts the 5-dimensional vector into a probability distribution for each fault type. Each element of the probability distribution takes a value between 0 and 1, and the sum is 1. The maximum probability value in the probability distribution is selected. When the maximum probability value exceeds 0.85, the gas turbine is determined to have the corresponding fault type. When the maximum probability value is below 0.85 but above 0.60, the diagnosis result is uncertain and a secondary diagnosis process needs to be initiated. When the maximum probability value is below 0.60, the gas turbine is determined to be in normal operating condition.

[0066] The specific implementation of step S05 involves establishing a fractional-order performance degradation differential equation model for the gas turbine determined to be in a fault state. This model uses fractional derivatives to describe the changes in performance parameters over time. The order of the fractional derivative is a real number between 0 and 1; the closer the order is to 1, the stronger the memory effect of the degradation process. A particle swarm optimization algorithm is used to identify the optimal fractional order and model parameters from historical monitoring data. Fifty particles are initialized and randomly distributed in the search space formed by the fractional order and model parameters. Each particle contains a fractional order and three model parameters. The mean square error between the output of the fractional-order performance degradation differential equation model corresponding to each particle and the observed performance parameters in the historical monitoring data is calculated. This mean square error is used as a fitness function to evaluate the particle's performance. The historical optimal position of each particle is recorded. The particle velocity and position are updated based on the current velocity, historical best position, and global best position. After 200 iterations, the fitness function converges to the minimum value, and the optimal fractional order reference value is obtained as 0.75. The reference values ​​for the three model parameters are 0.02, 0.15, and 0.80, respectively. The fractional performance degradation differential equation model is solved using the Green's function method. The fractional performance degradation differential equation model is transformed into an algebraic equation in the frequency domain by performing a Laplace transform on both sides of the fractional performance degradation differential equation model. The algebraic equation is solved to obtain the frequency domain expression of the performance parameters. The inverse Laplace transform is performed on the frequency domain expression to obtain the remaining lifetime prediction value in the time domain. The remaining lifetime prediction value represents the operating time required for the gas turbine to reduce its performance parameters to the failure threshold from the current moment. The failure threshold is set to 70% of the rated performance.

[0067] The specific implementation of step S06 is to adjust the operating parameters of the gas turbine according to the remaining life prediction value. When the remaining life prediction value is greater than 500 hours, it is determined that the gas turbine still has a long service life. The current power output is maintained to ensure power generation efficiency. When the remaining life prediction value is between 200 hours and 500 hours, it is determined that the gas turbine has entered the mid-stage of performance degradation. The power output is reduced to 80% of the rated power to slow down the degradation rate and extend the service life. When the remaining life prediction value is less than 200 hours, it is determined that the gas turbine is about to fail. The shutdown protection action is triggered to prevent catastrophic failure. At the same time, a maintenance warning signal is sent to the monitoring center through the communication interface. The maintenance warning signal includes the fault type and the remaining life prediction value.

[0068] It should be noted that the key technical ideas of this invention include a multimodal feature fusion mechanism based on tensor ring decomposition, a low-rank approximation optimization mechanism based on matrix decomposition, and a remaining lifetime prediction method for fractional-order performance degradation differential equation models. The multimodal feature fusion mechanism based on tensor ring decomposition organizes acoustic emission signals, vibration acceleration signals, and thermogram sequences into fourth-order tensors. It extracts compact multilinear feature representations through tensor ring decomposition. Compared to traditional feature concatenation methods, tensor ring decomposition preserves the high-order coupling relationships between time-frequency and spatial modes, avoiding the storage and computational bottlenecks caused by the exponential growth of the core tensor dimension with order. The extracted multimodal fusion feature vectors contain richer fault discrimination information, reducing the number of training samples required by subsequent classifiers and improving diagnostic accuracy and generalization ability under small sample conditions. The low-rank approximation optimization mechanism based on matrix factorization decomposes the high-dimensional weight matrix of the cross-modal attention module in the fault identification model into the product of multiple low-dimensional matrices. It determines the optimal decomposition dimension through singular value decomposition and introduces kernel norm regularization constraints, significantly reducing the number of model parameters and computational complexity. This reduces the number of floating-point operations in a single forward propagation and lowers memory usage. Simultaneously, the alternating minimization optimization strategy ensures that the parameters of each low-rank matrix are co-optimized to achieve a globally approximate optimal solution. While maintaining fault diagnosis accuracy, it achieves lightweight model deployment to meet the computing power constraints of edge computing devices and accelerates inference speed to meet the millisecond-level fault response requirements of gas turbines. The fractional-order performance degradation differential equation model for remaining lifetime prediction utilizes fractional derivatives to describe the long-range dependence of material aging and performance degradation. The non-integer values ​​of the fractional order characterize the nonlinear change of the degradation rate over time, capturing the slow evolution in the early degradation stage and the dynamic process of accelerated failure in the later stage. Compared with integer-order models, it provides more accurate remaining lifetime predictions, supporting predictive maintenance decisions to avoid economic losses caused by premature downtime and the risk of catastrophic failures due to delayed maintenance. The synergistic effect of the three key technical approaches lies in the fact that the multimodal fusion feature vector extracted by tensor ring decomposition provides high-quality input to the fault identification model, the low-rank approximation optimization mechanism based on matrix decomposition ensures that the model can achieve rapid and accurate diagnosis in resource-constrained environments, and the fractional-order performance degradation differential equation model predicts the remaining life based on the diagnosis results to guide the adjustment of operating parameters. The three form a complete technical chain from multi-source data fusion to intelligent diagnosis and then to life prediction, which solves the core problems of traditional methods such as the dimensionality curse in high-dimensional feature space, real-time inference computing resource bottlenecks, and long-term reliance on modeling difficulties, and realizes early warning and accurate diagnosis of gas turbine faults.

[0069] It should be noted that this invention also addresses the following technical problem: when deploying fault identification models on edge computing devices, computational constraints prevent inference speeds from meeting millisecond-level response requirements. This invention utilizes a low-rank approximation optimization mechanism based on matrix factorization to decompose the weight matrix in the cross-modal attention module into a product of the query low-rank matrix, the key low-rank matrix, and the value low-rank matrix. This significantly reduces the number of model parameters and computational complexity, as well as the number of floating-point operations and memory usage per forward propagation. Simultaneously, it introduces a sparse connection learning framework based on dynamic topology reconstruction. During training, it evaluates the information transmission contribution of connections by calculating the mutual information between neuron pairs and adaptively prunes redundant connections. High-importance connections are regrown at pruned locations based on gradient magnitude, gradually evolving the dense network into a sparse network. This reduces memory accesses and multiplication / accumulation operations during inference, enabling lightweight model deployment and accelerating diagnostic speed to meet the real-time fault response requirements of gas turbines.

[0070] Specifically, the principle of this invention is as follows: The reason why tensor ring decomposition can solve the problem of high-order coupling feature extraction in multimodal sensor data is that it decomposes the fourth-order tensor into a ring-connected sequence of kernel tensors. Each kernel tensor corresponds to a physical dimension, and adjacent kernel tensors form a closed-loop structure through tensor contraction operations. This decomposition method is mathematically equivalent to performing a multilinear projection transformation on the original tensor. The temporal kernel tensor captures the evolution pattern of the fault on the time axis, the frequency kernel tensor represents the contribution weight of each frequency component, the spatial kernel tensor reflects the spatial correlation of the sensor's placement, and the modal kernel tensor quantifies the coupling strength of the three physical modes: acoustic emission, vibration, and thermography. The alternating least squares iterative optimization process ensures that all kernel tensors converge to the optimal solution by fixing the parameters of other kernel tensors and optimizing only the current kernel tensor. The multimodal fusion feature vector composed of the extracted principal components and coupling coefficients retains the essential structure of the original data. After this feature vector is input into the cross-modal attention module, the multi-head self-attention mechanism further mines the deep dependencies between features, ultimately achieving accurate identification of complex fault modes.

[0071] The following provides a specific embodiment 1 of the present invention. The specific implementation of steps S01 and S06 in this embodiment 1 is the same as that described above, and will not be repeated in detail here. The specific implementation of other steps is described in detail below.

[0072] The specific implementation of step S02 involves performing wavelet packet decomposition on the acoustic emission signal to extract energy features across eight frequency bands, and then using the Daubechies wavelet basis function for a three-level decomposition. The formula for calculating the energy characteristics of each frequency band is expressed as follows:

[0073] ;

[0074] In the formula, For the first Normalized energy characteristics of each frequency band, dimensionless. The value range is from 1 to 8; For the first The first frequency band Each wavelet packet coefficient is represented in millivolts. This is the wavelet packet coefficient index, with values ​​ranging from 1 to... ; For the first The total number of wavelet packet coefficients in each frequency band is dimensionless. The frequency band index variable used in the summation of the denominator has a value range of 1 to 8; The wavelet packet coefficient index variable in the summation of the denominator; For the first The total number of wavelet packet coefficients in each frequency band is dimensionless; the denominator is the sum of the energies of all frequency bands used for normalization, with units of millivolts squared. A fast Fourier transform is performed on the vibration acceleration signal, with 2048 transform points, to extract the amplitude and phase spectra within the frequency range of 0Hz to 25kHz. The formulas for calculating the amplitude and phase spectra are as follows:

[0075] , ;

[0076] In the formula, For frequency The normalized amplitude spectrum at the location is dimensionless; For frequency The Fourier transform of the value at the given point is a complex value, expressed in meters per second squared. This represents the number of Fourier transform points, which is 2048 in this case and is dimensionless. For reference amplitude, it is usually taken as 1 meter per second squared, and the unit is meters per second squared; For frequency The phase spectrum at the location, in radians; and They are respectively The imaginary and real parts are both in meters per second squared; The frequency variable is expressed in Hertz. Temperature gradient distribution features are extracted from thermal image sequences by calculating the temperature difference between each pixel and its eight neighboring pixels. The formula for calculating the temperature gradient distribution features is as follows:

[0077] ;

[0078] In the formula, For position The normalized temperature gradient characteristics at the location are dimensionless. For position The temperature value at that location, in degrees Celsius; For position The set of eight neighboring pixels; and These are the x and y coordinates of the neighboring pixels, respectively, in pixels; For neighboring pixels Temperature value, in degrees Celsius; For reference temperature, the value is usually taken as the average temperature of the thermal image, in degrees Celsius; and These represent the x and y coordinates of a pixel, in pixels. An 8×8 pixel max pooling operation is performed to reduce the dimensionality of the temperature gradient features. The energy features, amplitude spectrum, phase spectrum, and temperature gradient distribution features of the eight frequency bands are organized into a fourth-order tensor. ,in The time step dimension is set to 1000. The frequency dimension is set to 128. The number of sensors is set to 16. The modality quantity dimension is set to 3.

[0079] The specific implementation of step S03 is to perform tensor ring decomposition on the fourth-order tensor, thereby decomposing the fourth-order tensor into a tensor ring. The decomposition into a ring-connected sequence of kernel tensors is expressed mathematically as follows:

[0080] ;

[0081] In the formula, For a fourth-order tensor at the index The element value at that location is dimensionless. These are the time kernel tensor, frequency kernel tensor, spatial kernel tensor, and modal kernel tensor, respectively, all of which are dimensionless; Let be the rank parameters of each kernel tensor, which are dimensionless and have empirical values ​​of 8, 12, 10, and 6, respectively. For the corresponding rank index variable, it is dimensionless and takes values ​​from 1 to... 1 to 1 to 1 to ; This is a time index, with values ​​ranging from 1 to 1000; This is a frequency index, with values ​​ranging from 1 to 128; This is the sensor index, with a value range of 1 to 16; The modality index ranges from 1 to 3. Alternating least squares iterative optimization updates the current kernel tensor while keeping other kernel tensors fixed. The formula for calculating the reconstruction error is as follows:

[0082] ;

[0083] In the formula, The normalized reconstruction error is dimensionless. It is the original fourth-order tensor, dimensionless; The approximate tensor reconstructed from the kernel tensor is dimensionless; The Frobenius norm is calculated using the following formula: ,in This is a time index summation variable, with values ​​ranging from 1 to... , For frequency index summation variables, the value range is 1 to... , The sensor index summation variable has a value range of 1 to 1. , The modal index summation variable takes values ​​from 1 to 1. The Frobenius norm is dimensionless. Iterative optimization continues until... Less than 0.001. Extract time kernel tensor. First principal component, frequency kernel tensor The first three principal components and modal kernel tensors The coupling coefficients are used as multimodal fusion feature vectors. The coupling coefficients of the modal kernel tensor are obtained through singular value decomposition, and the modal kernel tensor is expanded into a matrix along the modal dimension. Perform singular value decomposition ,in It is a left singular matrix with dimension . Dimensionless It is a diagonal singular value matrix with dimension . Dimensionless It is a right singular matrix with dimension . Dimensionless The diagonal elements are the coupling coefficients, which are dimensionless.

[0084] The specific implementation of step S04 involves inputting the multimodal fusion feature vector into the fault identification model to determine the fault type. The probability distribution of each fault type output by the fault identification model is calculated using the Softmax function. The formula for calculating the probability distribution is as follows:

[0085] ;

[0086] In the formula, For the first The output probability of the fault type is dimensionless and ranges from 0 to 1. For the first The output value of the fully connected layer corresponding to the class is dimensionless. The total number of fault types includes five categories: compressor surge fault, turbine blade crack fault, combustion chamber hot spot fault, bearing wear fault, and seal leakage fault. It is an exponential function; This is a fault type index, with values ​​ranging from 1 to 5; The summation index variable takes values ​​ranging from 1 to... When the maximum probability value If the value exceeds 0.85, it is determined to be the corresponding fault type. When the maximum probability value is lower than 0.85 but higher than 0.60, the secondary diagnostic process is initiated. When the maximum probability value is lower than 0.60, it is determined to be in normal operating condition.

[0087] The specific implementation of step S05 involves establishing a fractional-order performance degradation differential equation model for the gas turbine determined to be in a fault state. The mathematical expression of the fractional-order performance degradation differential equation model is as follows:

[0088] ;

[0089] In the formula, For fractional derivative operators, it is defined as follows: ,in greater than The smallest integer, Let the integral variable represent historical moments, in hours. It is the Gamma function; It is a fractional order, dimensionless, and its value ranges from 0 to 2; For a moment The performance parameter values ​​are in units depending on the parameter type, with common units being kilowatts or efficiency percentages. These are reference values ​​for performance parameters, typically taken as rated performance values, with units of [missing information]. Consistent; This is a time variable, with the unit being hours; This is the normalized performance degradation factor, expressed in negative values ​​per hour. Power; The normalized external input influence coefficient is dimensionless. For a moment The operating condition input is determined by the type of operating condition parameter; These are reference values ​​for operating conditions; units are equal to or less than 1000 units. Consistent; The normalization constant term, in units of hours, is negative. power. Fractional order. The fitness function of the particle swarm optimization algorithm is described as follows, which is obtained from historical monitoring data:

[0090] ;

[0091] In the formula, The normalized mean square error fitness function is dimensionless. The number of historical monitoring data points, dimensionless; For a moment The observation performance parameter values, in units of Consistent; The performance parameter values ​​predicted by the model, in units of and . Consistent; This is a normalization factor, typically taken as the nominal value of the performance parameter, with units of 1 and 2. Consistent; This is an index for historical data points, with values ​​ranging from 1 to... ; For the first Historical monitoring times, in hours. The fractional-order performance degradation differential equation model is solved using the Green's function method. The time-domain problem is transformed into a frequency-domain algebraic equation using the Laplace transform. After solving, the inverse Laplace transform is performed to obtain the predicted remaining lifetime. The formula for calculating the remaining useful life prediction is as follows:

[0092] ;

[0093] In the formula, This is the predicted remaining lifespan, in hours. The performance parameter value at the current moment, in units of and . Consistent; This is the failure threshold performance parameter value, typically taken as 60% of the rated performance, in units of... Consistent; The current performance degradation rate, in units of Divide the unit by the hour; This is the current time, expressed in hours.

[0094] To better understand and implement this invention, the following is a specific application scenario of this invention, Example 2:

[0095] To verify the effectiveness of this invention, technicians set up a test environment and conducted real-time monitoring and fault diagnosis tests on a 185MW heavy-duty gas turbine. This gas turbine had been running continuously for 23,600 hours and recently exhibited abnormal phenomena such as fluctuating exhaust temperature and increased vibration levels. Technicians placed three acoustic emission sensors at the turbine blade roots, two acoustic emission sensors on the combustion chamber wall, and four and seven vibration sensors on the two bearing housings and the casing shell, respectively, for a total of 16 sensor locations. The sampling frequency of the acoustic emission sensors was set to 500kHz, the sampling frequency of the vibration sensors was set to 50kHz, and the thermal imager frame rate was set to 60Hz. Figure 2 As shown, the sensor placement locations have been optimized to cover critical components and potential sources of failure in the gas turbine.

[0096] Technicians continuously collected multimodal signal data at 1000 time points, with a collection time window of 20 seconds. The acoustic emission signal was decomposed using a three-level wavelet packet decomposition with Daubechies wavelet basis functions, extracting energy features from eight frequency bands. The normalized energy values ​​for each band were 0.142, 0.186, 0.234, 0.198, 0.127, 0.065, 0.032, and 0.016, respectively. A 2048-point Fast Fourier Transform was performed on the vibration acceleration signal to extract the amplitude and phase spectra within the frequency range of 0Hz to 25kHz, with a frequency resolution of 12.2Hz, forming frequency domain features of 128 discrete frequency components. Temperature gradient distribution features were extracted from the thermal image sequence. The temperature difference between each pixel and its eight neighboring pixels was calculated, and the absolute value was then subjected to max pooling in an 8×8 pixel window to obtain the dimensionality-reduced temperature gradient distribution features. The extracted energy features of 8 frequency bands, 128 amplitude spectrum components, 128 phase spectrum components, and temperature gradient distribution features were organized into... The fourth-order tensor has a time step of 1000, a frequency dimension of 128, a sensor count of 16, and a modality count of 3.

[0097] like Figure 3As shown, the technicians performed tensor ring decomposition on the fourth-order tensor, decomposing it into a ring product structure of time kernel tensor, frequency kernel tensor, spatial kernel tensor, and modal kernel tensor. Alternating least squares iterative optimization of the parameters of each kernel tensor was used, with the ranks of the kernel tensors initialized to 32, 48, 24, and 8, respectively. After 127 iterations, the reconstruction error converged to 0.00087. The first principal component was extracted from the time kernel tensor as a time-domain feature, and the first three principal components were extracted from the frequency kernel tensor as frequency-domain features. The coupling coefficient was calculated from the modal kernel tensor, and three singular values ​​were obtained by singular value decomposition along the modal dimension: 4.73, 1.85, and 0.62. The 32-dimensional vector of the first principal component of the time kernel tensor, the 144-dimensional vector of the first three principal components of the frequency kernel tensor, and the 80-dimensional vector of the coupling coefficient of the modal kernel tensor were concatenated to form a 256-dimensional multimodal fusion feature vector.

[0098] Technicians input a 256-dimensional multimodal fusion feature vector into the fault identification model for fault type discrimination. The ratio of the maximum to minimum value of the modal kernel tensor coupling coefficient is 7.63; therefore, the number of heads in the multi-head self-attention mechanism of the cross-modal attention module is set to 16. After receiving the 256-dimensional feature vector at the model input layer, it is mapped to a 512-dimensional latent space through a feature embedding layer. Then, feature interaction occurs through four layers of cross-modal attention modules, each containing a 16-head self-attention mechanism and a feedforward neural network with a hidden layer dimension of 2048. The output of the global average pooling layer is mapped to the dimensions of five fault types through two fully connected layers, and the probability distribution of each fault type is output through the Softmax activation function, as shown in Table 1.

[0099] Table 1. Fault Type Probability Distribution Table

[0100]

[0101] The maximum probability value output by the fault identification model was 0.783, falling between 0.60 and 0.85, thus initiating a secondary diagnostic process. Technicians increased the sampling time window to 60 seconds and reacquired acoustic emission signals, vibration acceleration signals, and thermal image sequences. The same feature extraction process was performed on the reacquired signals, resulting in a resampled fourth-order tensor. This resampled multimodal fusion feature vector was extracted through tensor loop decomposition. The resampled multimodal fusion feature vector was concatenated with the original multimodal fusion feature vector to form a 512-dimensional extended feature vector, which was then input into the fault identification model for further analysis. The probability values ​​for each fault type identified in this second analysis were: compressor surge fault 0.031, turbine blade crack fault 0.847, combustion chamber hot spot fault 0.083, bearing wear fault 0.028, and seal leakage fault 0.011. The maximum probability value of 0.847 exceeded 0.75, therefore, a turbine blade crack fault was determined to exist in the gas turbine.

[0102] likeFigure 4 As shown, technicians analyzed the multimodal signal characteristics of turbine blade crack failure. Acoustic emission signals generated continuous high-frequency bursts during crack propagation, with a frequency range concentrated between 180kHz and 320kHz and an average amplitude of 0.47V. The vibration acceleration signal showed a significant increase in energy at a rotational frequency of 50Hz and its higher harmonics, with a peak amplitude spectrum of 23.6V at 150Hz. Simultaneously, sideband modulation was observed, with a modulation frequency of 8.3 Hz. The thermal imaging sequence revealed an abnormal temperature gradient at the leading edge of the third-stage turbine blade, with a temperature of 892°C, 37°C higher than the surrounding area, and a heat flux density reaching [missing value]. .

[0103] Technicians established a fractional-order performance degradation differential equation model to predict the remaining life of the gas turbine. Historical monitoring data from the past 6 months were collected, including acoustic emission signals, vibration acceleration signals, thermal image sequences, and corresponding power output, exhaust temperature, compressor efficiency, and turbine inlet temperature as observed values ​​for performance parameter degradation, as shown in Table 2.

[0104] Table 2 Performance Parameter Degradation History Data Table

[0105]

[0106] like Figure 5 As shown, the technicians used a particle swarm optimization algorithm to identify the parameters of a fractional-order performance degradation differential equation model. Fifty particles were initialized and randomly distributed in the model parameter space with fractional orders ranging from 0.3 to 0.9. After 186 iterations, the fitness function converged to a minimum of 0.0023, obtaining an optimal fractional order of 0.67. The model parameters included the degradation coefficient. The fractional-order performance degradation differential equation model was solved using the Green's function method. The time-domain problem was transformed into a frequency-domain problem through the Laplace transform. After solving the algebraic equation, the remaining lifetime prediction value was obtained as 287 hours through the inverse Laplace transform.

[0107] Based on the predicted remaining service life of 287 hours, which falls between 200 and 500 hours, technicians reduced the gas turbine power output to 80% of its rated power, or 148 MW. Simultaneously, they lowered the exhaust temperature monitoring threshold to 580°C, set the compressor efficiency monitoring threshold to 84%, and the turbine inlet temperature monitoring threshold to 1420°C. Under reduced power operation mode, technicians continuously monitored the gas turbine's multimodal signal changes; the high-frequency burst intensity of the acoustic emission signal decreased to 0.31V, and the peak amplitude of the vibration acceleration signal at 150Hz decreased to 17.2V. The thermal imaging sequence showed that the temperature in the abnormal temperature gradient region at the leading edge of the blade decreased to 865℃. Reduced power operation effectively slowed the propagation rate of the turbine blade cracks, providing a window of opportunity for subsequent planned maintenance.

[0108] This invention represents a significant technological advancement compared to traditional gas turbine fault diagnosis methods. Traditional methods typically employ single-modal signal analysis or simple multimodal signal concatenation, failing to effectively capture the high-order coupling relationships between different physical modes, resulting in incomplete fault feature representation. This invention organizes multimodal signals into high-order tensors and decomposes them into a ring-connected kernel tensor sequence through tensor ring decomposition, preserving the multilinear structure between time, frequency, space, and modes. The extracted compact feature representation contains richer fault discrimination information. Traditional fault identification models often use shallow classifiers or fully connected neural networks, lacking the ability to model intermodal interactions. This invention employs a cross-modal attention module to construct a deep feature interaction network, capturing fine-grained interactions between strongly coupled modes through a multi-head self-attention mechanism, dynamically adjusting the number of attention heads to adapt to the feature distribution of different fault modes. Traditional methods use linear regression or integer-order differential equations in performance degradation modeling, failing to describe the long-range dependence of material aging and performance degradation. This invention introduces a fractional-order performance degradation differential equation model, using non-integer derivatives to characterize the nonlinear change in degradation rate over time. This captures the slow evolution in the early degradation stage and the dynamic process of accelerated failure in the later stage, providing more accurate remaining lifetime prediction. Traditional methods suffer from large model parameters, resulting in high computational overhead and difficulty in deployment on edge computing devices. This invention achieves model lightweighting through a low-rank approximation optimization mechanism based on matrix factorization and a sparse connection learning framework based on dynamic topology reconstruction. This reduces computational complexity and memory usage while maintaining diagnostic accuracy, meeting the millisecond-level fault response requirements of gas turbines.

[0109] It should be noted that the variables involved in this invention are explained in detail in Tables 3 and 4.

[0110] Table 3. Variable Explanation Table (Part 1)

[0111]

[0112] Table 4. Variable Explanation Table (Part Two)

[0113]

[0114] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A gas turbine fault diagnosis method based on a multimodal large model, characterized in that, Acoustic emission signals, vibration acceleration signals, and thermogram sequences were collected during the operation of the gas turbine. Wavelet packet decomposition was performed on the acoustic emission signals to extract the energy features of the frequency band. Fourier transform was performed on the vibration acceleration signals to extract the amplitude and phase spectra. Temperature gradient distribution features were extracted from the thermogram sequences. The energy features, amplitude and phase spectra, and temperature gradient distribution features of the frequency bands were organized into fourth-order tensors. Tensor ring decomposition was performed on the fourth-order tensors to decompose them into a ring-shaped product structure of time kernel tensor, frequency kernel tensor, spatial kernel tensor, and modal kernel tensor. The principal components of the time kernel tensor, the principal components of the frequency kernel tensor, and the coupling coefficients of the modal kernel tensor were extracted as multimodal fusion feature vectors. The multimodal fusion feature vectors were input into a fault identification model to determine the fault type. For gas turbines determined to be in a fault state, a fractional-order performance degradation differential equation model was established to predict the remaining life, and the fault type determination result and the remaining life were output.

2. The gas turbine fault diagnosis method based on a multimodal large model according to claim 1, characterized in that, Wavelet packet decomposition uses the Daubechies wavelet basis function, with a decomposition level of 3. After each level of decomposition, the energy of each sub-band signal is calculated and normalized to obtain the energy characteristics of the frequency band.

3. The gas turbine fault diagnosis method based on a multimodal large model according to claim 2, characterized in that, The Fourier transform uses the Fast Fourier Transform algorithm with 2048 transform points, extracting the amplitude spectrum and phase spectrum in the frequency range of 0Hz to 25kHz as frequency domain features.

4. The gas turbine fault diagnosis method based on a multimodal large model according to claim 3, characterized in that, The temperature gradient distribution feature is obtained by calculating the temperature difference between each pixel and its eight neighboring pixels in the thermal image sequence. The absolute value of the temperature difference is taken and max pooling is performed in the spatial dimension to obtain the dimensionality-reduced temperature gradient distribution feature.

5. The gas turbine fault diagnosis method based on a multimodal large model according to claim 4, characterized in that, The dimensions of the fourth-order tensor are time step multiplied by frequency dimension multiplied by number of sensors multiplied by number of modes. The time step is 1000 continuously acquired time points, the frequency dimension has 128 discrete frequency components, the number of sensors is 16 sensor positions, and the number of modes is three physical modes: acoustic emission vibration and thermography.

6. The gas turbine fault diagnosis method based on a multimodal large model according to claim 5, characterized in that, Tensor ring decomposition decomposes a fourth-order tensor into a ring-connected sequence of kernel tensors. Each kernel tensor is a third-order tensor. Adjacent kernel tensors are connected through tensor contraction operations to form a closed-loop structure. The parameters of each kernel tensor are optimized through alternating least squares iterations until the reconstruction error is less than 0.

001.

7. The gas turbine fault diagnosis method based on a multimodal large model according to claim 6, characterized in that, The coupling coefficient of the modal kernel tensor characterizes the interaction strength between different physical modes. The coupling coefficient is obtained by performing singular value decomposition on the modal kernel tensor along the modal dimension. The magnitude of the singular values ​​in the singular value decomposition reflects the degree of contribution of the corresponding mode combination.

8. The gas turbine fault diagnosis method based on a multimodal large model according to claim 7, characterized in that, The input layer of the fault identification model receives a 256-dimensional multimodal fusion feature vector, which is then mapped to a 512-dimensional latent space through a feature embedding layer. After the feature embedding layer, four cross-modal attention modules are connected. The cross-modal attention modules include a multi-head self-attention mechanism and a feedforward neural network.

9. The gas turbine fault diagnosis method based on a multimodal large model according to claim 8, characterized in that, The fault identification model utilizes a low-rank approximation optimization mechanism based on matrix factorization to decompose the weight matrix in the cross-modal attention module into a product of the query low-rank matrix, the key low-rank matrix, and the value low-rank matrix. It introduces a nuclear norm regularization term to constrain the sum of the singular values ​​of the query low-rank matrix, the key low-rank matrix, and the value low-rank matrix.

10. The gas turbine fault diagnosis method based on a multimodal large model according to claim 9, characterized in that, The fault identification model utilizes a sparse connection learning framework based on dynamic topology reconstruction. During training, it calculates the mutual information between neuron pairs to evaluate the information transmission contribution of the connection. When the mutual information of a connection is lower than a dynamic threshold, the connection is pruned, and highly important connections are regrown at the pruned connection locations based on the gradient magnitude.

Images