Combustion gas turbine rotor composite fault diagnosis method and related device

Abstract

The invention belongs to the technical field of rotor dynamics, and discloses a gas turbine rotor composite fault diagnosis method and a related device. The method comprises the following steps: extracting a fault signal, and performing diagnosis and identification by using a composite fault diagnosis model; wherein the composite fault diagnosis model comprises a feature extraction module used for inputting a fault signal and outputting a reconstructed feature map, a time sequence feature screening module used for outputting a hidden state containing time sequence features according to the reconstructed feature map, an attention weighting module used for weighting the hidden state and outputting key time sequence features, and a fault diagnosis module used for performing fault diagnosis on the key time sequence features. And the fault diagnosis module is used for mapping the output of the attention weighting module into one-hot coding for converting the classification variables into binary vectors to realize rotor compound fault diagnosis. The technical problems that feature extraction depends on manual work, multi-sensor information utilization is insufficient, diagnosis precision is low and the like in composite fault diagnosis are solved.

Description

Technical Field

[0001] This invention belongs to the field of rotor dynamics technology, and specifically relates to a method and related device for diagnosing complex faults in gas turbine rotors. Background Technology

[0002] The rotor system is the core component of rotating machinery, and its operating status is directly related to the safety and efficiency of the entire machine.

[0003] Currently, traditional rotor system fault diagnosis methods mainly include model-based methods and signal processing-based methods. Model-based methods rely on precise mathematical models, resulting in complex modeling and poor generalization ability. Signal processing-based methods (such as wavelet transform and envelope analysis) require manual feature extraction, depend on expert experience, and have limited effectiveness in diagnosing complex faults. Furthermore, existing deep learning methods (such as fully connected networks, convolutional neural networks, and deep belief networks) are mostly used for single fault diagnosis, lacking sufficient diagnostic capability for complex faults and lacking effective multi-sensor fusion mechanisms.

[0004] Furthermore, existing methods typically treat composite faults as an independent fault type, leading to an explosion in the number of categories, high computational complexity, and neglecting the inherent correlation between faults. In addition, bearing faults are often ignored in rotor systems, while in reality, bearing and rotor faults often occur simultaneously.

[0005] In summary, there is an urgent need to develop a new joint diagnostic scheme for complex faults in rotor systems. Summary of the Invention

[0006] The purpose of this invention is to provide a method and related apparatus for diagnosing complex faults in gas turbine rotors, thereby solving one or more of the aforementioned technical problems. The technical solution disclosed in this invention overcomes common problems in complex fault diagnosis, such as reliance on manual feature extraction, insufficient utilization of multi-sensor information, and low diagnostic accuracy, achieving accurate diagnosis of rotor complex faults.

[0007] To achieve the above objectives, the present invention adopts the following technical solution: In a first aspect, the present invention provides a method for diagnosing complex faults in a gas turbine rotor, comprising the following steps: Fault signals are extracted based on the monitored abnormal vibration data; Based on the extracted fault signals, a trained composite fault diagnosis model is used for diagnosis and identification to obtain rotor composite fault diagnosis results. The composite fault diagnosis model includes: a feature extraction module for inputting fault signals and outputting reconstructed feature maps; a temporal feature filtering module for outputting hidden states containing temporal features based on the reconstructed feature maps; an attention weighting module for weighting the hidden states and outputting key temporal features; and a fault diagnosis module for mapping the output of the attention weighting module to one-hot encoding that converts categorical variables into binary vectors to achieve rotor composite fault diagnosis.

[0008] A further improvement to the technical solution of the present invention lies in that, In the feature extraction module, a one-dimensional convolutional neural network (1DCNN) is used to extract local features from the input data. High-dimensional features are extracted step by step through convolutional layers, ReLU activation function and pooling layer, and the reconstructed feature map is output. In the temporal feature filtering module, the reconstructed feature map is used as the input of the Long Short-Term Memory (LSTM) network to capture temporal dependencies and transform the time series of the fault signal into a hidden state containing temporal features. The attention weighting module introduces a self-attention mechanism to generate query vectors. Q and key vector K The attention score is obtained by dot product operation and then compared with all corresponding vectors. V Multiply and sum the results to output the attention mechanism, which is then used to weight the hidden states. Finally, the fully connected layer outputs the key temporal features. In the fault diagnosis module, the output of the attention weighting module is mapped to a one-hot code that converts categorical variables into binary vectors through a fully connected layer and a Sigmoid activation function, thereby realizing rotor composite fault diagnosis.

[0009] A further improvement to the technical solution of the present invention is that the offline training step of the composite fault diagnosis model includes: Obtain training set data; The training set data is used to train the model, and after the preset conditions are met, a well-trained composite fault diagnosis model is obtained. The training set data is obtained based on a rotor system test device, which includes: a base platform, a spindle, a disk, a motor, a sensor bracket, and a data acquisition device. The base platform serves as a support for the rotor system test apparatus and provides a vibration isolation environment during the test. The main shaft is detachably rotatably mounted on the base platform via a support bearing and bearing housing. The motor is fixedly mounted on the base platform, and the output end of the motor is connected to the main shaft via a coupling. Multiple discs are detachably mounted on the main shaft, and concentric circular grooves for mounting unbalanced blocks are formed on the discs. The base platform is equipped with the sensor bracket, on which multiple displacement sensors are mounted. These sensors are used to collect response signals in three measurement directions: perpendicular to the main shaft, parallel to the main shaft, and at a preset angle to the main shaft, and output to the data acquisition unit.

[0010] A further improvement of the technical solution of the present invention is that, in the process of training the model using the training set data and obtaining the trained composite fault diagnosis model after reaching the preset conditions, a dropout layer is added to prevent overfitting, an early stopping method is used to control the number of training rounds, a binary cross-entropy loss function and an Adam optimizer are used to carry out model training, and the model parameters are optimized through the backpropagation algorithm; in addition, a validation set data is used to validate the model, and the root mean square error is used as the evaluation index.

[0011] A further improvement to the technical solution of the present invention lies in the following specific steps for obtaining the training set data based on the rotor system test device: The motor drives the spindle and disc to rotate at a specified speed. Multiple displacement sensors mounted on the sensor bracket collect response signals in three measurement directions: perpendicular to the main shaft, parallel to the main shaft, and at a preset angle to the main shaft. These signals are then output to the data acquisition unit to obtain multi-channel sensor acquisition signals. Specifically, an unbalanced block is installed in the concentric circular groove of the disc to simulate rotor imbalance faults, a cut is made in the optical axis section of the main shaft to simulate crack faults, a shim is placed under one end of the bearing housing to simulate misalignment faults, and the fastening bolts of the bearing housing are loosened to simulate bearing housing loosening faults. The signals acquired by multiple sensors are standardized, and the time-series data are sliced ​​using the sliding window method. The data is then labeled according to the composite fault combinations of each segment to obtain the training set data. The length of each segment is constrained by both the rotational speed and the sampling rate, and each segment contains several complete rotational cycles of the vibration signal.

[0012] A further improvement of the technical solution of this invention lies in that, during the standardization process of signals acquired from multiple sensors, the expression for the standardization process is: ; In the formula, For the first i The signals acquired by each displacement sensor The first i The mean and standard deviation of the signals from each displacement sensor. Indicates the first i A standardized signal after processing.

[0013] A further improvement of the technical solution of the present invention is that, in the process of slicing time series data using the sliding window method, the window size is 40 and the step size is 1.

[0014] In a second aspect, the present invention provides a gas turbine rotor composite fault diagnosis system, comprising: The data acquisition unit is used to extract fault signals based on the monitored abnormal vibration data; The diagnostic identification unit is used to perform diagnostic identification based on the extracted fault signals using a trained composite fault diagnosis model, and obtain rotor composite fault diagnosis results. The composite fault diagnosis model includes: a feature extraction module for inputting fault signals and outputting reconstructed feature maps; a temporal feature filtering module for outputting hidden states containing temporal features based on the reconstructed feature maps; an attention weighting module for weighting the hidden states and outputting key temporal features; and a fault diagnosis module for mapping the output of the attention weighting module to one-hot encoding that converts categorical variables into binary vectors to achieve rotor composite fault diagnosis.

[0015] In a third aspect, the present invention provides an electronic device including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the program, implements the gas turbine rotor complex fault diagnosis method as described in any one of the first aspects of the present invention.

[0016] In a fourth aspect, the present invention provides a non-transitory computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the gas turbine rotor complex fault diagnosis method as described in any one of the first aspects of the present invention.

[0017] Compared with the prior art, the present invention has the following beneficial effects: The technical solution disclosed in this invention proposes a composite fault diagnosis model based on a triple coupling of convolutional network, long short-term memory network, and attention mechanism. First, local features are extracted from the signals of each sensor, and temporal dependencies are captured. Then, the signals from multiple sensors are weighted and fused. This overcomes common problems in composite fault diagnosis, such as reliance on manual feature extraction, insufficient utilization of multi-sensor information, and low diagnostic accuracy, achieving accurate diagnosis of rotor composite faults. Specifically, traditional signal processing-based methods require manual feature extraction, relying on expert experience. Therefore, the feature extraction module of this invention uses a one-dimensional convolutional neural network to output a reconstructed feature map. This process automatically completes local feature extraction without manual intervention, overcoming the dependence on expert experience and solving the problem of manual feature extraction. Furthermore, existing deep learning methods are mostly used for single fault diagnosis, and their ability to diagnose complex faults is insufficient. In view of this, the complex fault diagnosis model of the present invention consists of a feature extraction module, a temporal feature filtering module, an attention weighting module, and a fault diagnosis module. The temporal feature filtering module uses a long short-term memory network to capture temporal dependencies and transforms the time series of fault signals into hidden states containing temporal features. The attention weighting module introduces a self-attention mechanism to weight the hidden states and output key temporal features. This architecture can integrate information from multiple aspects, effectively identify complex faults, and overcome the problem of insufficient ability to diagnose complex faults.

[0018] In the preferred embodiment of this invention, based on the proposed rotor system test device, training set data is acquired through multi-channel sensor acquisition. A feature extraction module extracts local features from the signals of each sensor channel, and an attention weighting module performs weighted fusion of the multi-channel sensor signals. Instead of treating composite faults as independent fault types, the comprehensive analysis of the model avoids an explosion in the number of categories, reduces computational complexity, and considers the inherent correlation between faults. Furthermore, the rotor system test device design fully considers the simulation of bearing faults. Bearing-related faults are simulated by placing shims under one bearing housing and loosening the fastening bolts of the bearing housing, ensuring that the model training data includes bearing fault information and solving the problem of bearing faults often being overlooked.

[0019] In the preferred embodiment of the present invention, training set data is obtained through a rotor system test device, which can obtain multiple signals containing composite fault feature information, thereby improving diagnostic effectiveness. Training strategies such as using dropout layer to prevent overfitting and early stopping method to control the number of training rounds are adopted to further improve the performance and stability of the model, and finally achieve accurate diagnosis of rotor composite faults. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0021] Figure 1 This is a flowchart illustrating a method for diagnosing complex faults in a gas turbine rotor, as described in an embodiment of the present invention. Figure 2 This is a schematic diagram of the rotor system test apparatus in an embodiment of the present invention; Figure 3 This is a schematic diagram of the training and application process of the composite fault diagnosis model in a specific embodiment of the present invention. Figure 4 This is a schematic diagram of data preprocessing in an embodiment of the present invention; Figure 5 This is a schematic diagram of the 1DCNN feature extraction module in an embodiment of the present invention; Figure 6 This is a schematic diagram of the LSTM timing capture module in an embodiment of the present invention; Figure 7 This is a schematic diagram of Self-Attention weighting in an embodiment of the present invention; Figure 8 This is a schematic diagram of the composite fault diagnosis module in an embodiment of the present invention; Figure 9 This is a schematic diagram illustrating offline model training and online fault diagnosis in an embodiment of the present invention; Figure 10 This is a schematic diagram of a gas turbine rotor composite fault diagnosis system in an embodiment of the present invention; The annotations in the figure are explained as follows: 1. Basic platform; 2. Support bearing; 3. Bearing housing; 4. Disc; 5. Concentric circular groove; 6. Sensor bracket; 7. Spindle; 8. Unbalance block; 9. Eddy current displacement sensor; 10. Coupling; 11. Servo motor. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention; obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0023] Based on the technical solutions disclosed in the embodiments of this invention, all other embodiments obtained by those skilled in the art without inventive effort are within the scope of protection of this invention. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or device that includes a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to these processes, methods, products, or devices.

[0024] Please see Figure 1 The present invention provides a method for diagnosing complex faults in a gas turbine rotor, comprising the following steps: Step 1: Extract fault signals based on the monitored abnormal vibration data; Step 2: Based on the extracted fault signals, use the trained composite fault diagnosis model to perform diagnosis and identification, and obtain the rotor composite fault diagnosis results. The composite fault diagnosis model includes: a feature extraction module for inputting fault signals and outputting reconstructed feature maps; a temporal feature filtering module for outputting hidden states containing temporal features based on the reconstructed feature maps; an attention weighting module for weighting the hidden states and outputting key temporal features; and a fault diagnosis module for mapping the output of the attention weighting module to one-hot encoding that converts categorical variables into binary vectors to achieve rotor composite fault diagnosis.

[0025] The technical solution provided in this invention proposes a composite fault diagnosis model based on the triple coupling of convolutional network, long short-term memory network, and attention mechanism. First, local features are extracted from the signals of each sensor and the temporal dependencies are captured. Then, the signals of multiple sensors are weighted and fused. Finally, it overcomes common problems in composite fault diagnosis such as feature extraction relying on manual methods, insufficient utilization of multi-sensor information, and low diagnostic accuracy, and realizes accurate diagnosis of rotor composite faults.

[0026] Please see Figure 2 This invention provides a rotor system testing device, which mainly includes: a base platform 1, a main shaft 7, a disk 4, a servo motor 11, a sensor bracket 6, a support bearing 2, a bearing housing 3, a coupling 10, an unbalance block 8, an eddy current displacement sensor 9, and a data acquisition unit, etc. The basic platform 1 is used to provide support for the main body of the rotor system test device and to provide a vibration isolation environment for the test process; The bearing housing 3 and the support bearing 2 are used to support the rotating parts during the test, while bearing the load and reducing friction. The spindle 7 is used to detachably connect multiple disks 4 into one unit and is supported by two support bearings 2; The disk 4 has concentric circular grooves 5 for mounting unbalance blocks 8 to simulate a rotor system with unbalance faults; Servo motor 11 is used to drive spindle 7 and disk 4 to rotate at a specified speed; coupling 10 is used to connect servo motor 11 and spindle 7. The unbalance block 8 is used to simulate the unbalance and can be connected to the concentric groove 5 of the disk 4 by bolts and nuts; Several eddy current displacement sensors 9 are installed on the openings of the sensor bracket 6 respectively, and are used to output response signals in three measurement directions: perpendicular to the main shaft 7, parallel to the main shaft 7, and at a 45° angle to the main shaft 7, and output to the data acquisition unit.

[0027] The rotor system testing apparatus provided in this invention can collect sample data through experimental simulation. Specifically, an unbalanced block is installed in the concentric groove of the disc to simulate rotor imbalance faults; a steel wire is used to cut along the smooth shaft section of the main shaft to simulate crack faults; a shim is placed under one end of the bearing housing to simulate misalignment faults; and the fastening bolts of the bearing housing are loosened to simulate bearing housing loosening faults. A composite fault diagnosis model trained using the collected sample data can diagnose composite faults of the rotor and bearings, including misalignment, imbalance, bearing housing loosening, and main shaft cracks.

[0028] Please see Figure 3 The specific process of training and applying the composite fault diagnosis model provided in the specific embodiments of the present invention is as follows: Step S1, Fault signal acquisition: Drive the rotor to rotate at a set speed by a servo motor, install eddy current displacement sensors in each direction of the sensor bracket, and simultaneously acquire response signals in three measurement directions: perpendicular to the main shaft, parallel to the main shaft, and at a 45° angle to the main shaft, to obtain multi-channel sensor acquisition signals. Step S2, Data Preprocessing: Standardize the signals acquired by the multi-channel sensors, slice the time series data using the sliding window method with a window size of 40 and a step size of 1; label the composite fault combinations of each segmented signal for model training. Step S3, Feature Extraction: Use a one-dimensional convolutional neural network (1DCNN) to extract local features from the input data. High-dimensional features are extracted step by step through convolutional layers, ReLU activation function and pooling layers. The reconstructed feature map is output as the input of the long short-term memory network (LSTM). Step S4, Temporal Feature Filtering: Input the reconstructed feature map into the LSTM network to capture temporal dependencies and transform the time series of the input rotor composite fault signal into a hidden state containing temporal features. Step S5, Attention Weighting: Introduce a self-attention mechanism to generate query vectors. Q and key vectorK The attention score is obtained by dot product operation and then compared with all corresponding vectors. V Multiply and sum the results to output the attention mechanism, which is then used to weight the hidden states of the LSTM. Finally, the fully connected layer outputs the key temporal features. Step S6, Fault Diagnosis: The fault diagnosis module consists of a fully connected layer and a Sigmoid activation function. It maps the output of attention to a one-hot encoding that converts the categorical variable into a binary vector, thereby achieving composite fault diagnosis.

[0029] Furthermore, it also includes: Step S7, offline model training and composite fault diagnosis: Model training is carried out using training set data, a dropout layer is added to prevent overfitting, and an early stopping method is used to control the number of training rounds. The binary cross-entropy loss function and Adam optimizer are used for model training, and the model parameters are optimized through the backpropagation algorithm; Model validation is carried out using validation set data, with root mean square error as the evaluation index, to obtain an offline model for rotor composite fault diagnosis; When new abnormal vibration data is detected, the fault signal is extracted in real time as a test set and input into the model for identification, thereby realizing online diagnosis of rotor composite faults.

[0030] Please see Figure 4 The diagram illustrates the specific process of data preprocessing in step S2. In this embodiment of the invention, by installing multiple sensors in the rotor system test apparatus, multiple signals containing composite fault characteristic information can be obtained, thereby improving diagnostic effectiveness; then, the multi-channel acquired data is used as input, wherein the signal segment from each sensor occupies one channel respectively, for diagnosing composite faults in the rotor system.

[0031] Because the vibration signal of the rotor system exhibits periodicity when rotating at a specific speed, and the length of the data segment is constrained by both the rotor speed and the signal sampling rate, the complete signal sequence acquired within a specific time period should be evenly divided into multiple data segments. Through data preprocessing methods, the signals acquired by multiple sensors are standardized. A sliding window method is used to slice the time-series data, dynamically reading the data until the entire data sequence has been traversed. The window size is 40, and the step size is 1, generating multiple segments of standardized input data, ensuring that each data segment contains several complete rotational cycles of the vibration signal.

[0032] In a specific exemplary technical solution, the standardization process is as follows:

[0033] In the formula, For the first i The signals collected by each sensor The first i The mean and standard deviation of the signals from each sensor. Indicates the first i A standardized signal after processing.

[0034] A multi-label classification method is used to label some data segments with their corresponding real fault labels for model training.

[0035] The fault labels include misalignment, imbalance, loose bearing housing, and spindle crack. A "1" indicates the presence of a fault, and a "0" indicates its absence. There are 16 different combinations of composite faults, as shown in Table 1. These labels are used to mark the composite fault combinations for each segmented signal. These labels indicate the operating status of the monitored rotor system, including information on whether the rotor system is operating healthily and what composite faults are present.

[0036] Table 1. Composite Fault Combination Table

[0037] Please see Figure 5 This diagram illustrates the feature extraction module of a Convolutional Neural Network (CNN). CNNs capture local features through convolution operations and progressively abstract higher-level feature representations through a hierarchical structure, exhibiting outstanding feature extraction and learning capabilities, and demonstrating unique advantages when processing data with spatial structure. Explained, a CNN network typically consists of multiple convolutional layers, activation layers, pooling layers, and fully connected layers, each with convolutional kernels, strides, and padding. The convolutional layers perform convolution operations on specific regions of the input data using convolutional kernels. Each convolutional kernel shares the same weights to extract local features from the input information, reducing the number of parameters in the network and avoiding overfitting, thereby reducing training complexity.

[0038] To illustrate, the operations between one-dimensional convolutional layers can be described as follows: ; In the formula, They represent the first l The input and output of each convolutional layer These represent the number of channels and the size of the input signal, respectively. The weight tensor of the one-dimensional convolution kernel. For the first l The kernel size of each convolutional layer These represent the number of channels and the size of the output signal, respectively. For the ( l +1) bias terms for convolutional layers, where * indicates a one-dimensional convolution operation, calculated as follows: ;in, k The kernel size is [size]. s The step size, p This is the size of the fill element.

[0039] Pooling layers reduce the spatial size of the data by downsampling intermediate feature vectors, thereby reducing the number of parameters and computational cost, while simultaneously extracting key features and achieving translation invariance of the model. In classification tasks, pooling layers help extract the most representative key features. Similar to fully connected networks, CNN networks also employ activation functions. f (﹒) is used to introduce nonlinear changes, thereby improving the expressive power of the network: ; In the formula, This is the network output after the activation function. f (﹒) represents the activation function; commonly used activation functions include ReLU. At the end of the network, a flattening operation is performed to transform the tensor of multiple channels into a one-dimensional vector, which is then passed to the fully connected layer. ; In the formula, The input vector of the fully connected layer. This is the output vector of the fully connected layer. This is the weight matrix. This is the bias vector.

[0040] In the technical solution provided by this invention, a one-dimensional convolutional neural network (1DCNN) is applied to extract features from multi-sensor data. By utilizing spatial topology to create incompletely linked sparse spatial connections between adjacent layers, the number of parameters required for model training is reduced, and overfitting is avoided using a weight sharing algorithm. Furthermore, pooling operations further enhance network performance by leveraging the local characteristics of the data. After multiple convolutional and pooling layers and ReLU activation functions, the length of the input to the convolutional module is shortened while its dimension remains unchanged, and the extracted input features are passed to subsequent modules.

[0041] Please see Figure 6 It shows the LSTM timing capture module, which consists of three parts: a forget gate, an input gate, and an output gate. The forget gate... f t Control adaptive adjustment of the unit state at the previous moment c t-1 The degree of retention, with an output value ranging from 0 to 1 (0 indicating complete forgetting of prior information, and 1 indicating complete retention); input gate i t Control the current moment t Input information x t The update weights for the cell state are output with values ​​ranging from 0 to 1 (1 indicates that the cell state is updated entirely using the current input); output gate ot Adjust the hidden output of the LSTM cell based on the updated cell state. h t This ensures the validity of the output information. LSTM updates the information through the synergistic effect of filtering pre-order memories via the forget gate and incorporating current information into the input gate, directly determining the transmission of long-term dependent information. Furthermore, the output gate is obtained by element-wise multiplying the tanh-activated cell state, serving as the valid output at the current moment and propagating it to the next layer or subsequent modules. The calculation formulas for the three gate units are as follows: ; in, x t For input variables, c t For the current storage unit, h t The current hidden state, c t-1 For the storage unit of the previous moment, h t-1 This is the hidden state from the previous moment; w f , w i , w o , w c These are the input weight vectors for the forget gate, input gate, output gate, and candidate unit, respectively. b f , b i , b o , b c It is the bias vector of each door control unit; σ (﹒) represents the change in the gating unit; tanh(﹒) is the activation function.

[0042] LSTM uses a "gating" mechanism to change the memory state during information transmission, allowing the network to selectively forget and retain information, thus enabling information from the previous time step to be passed to units in subsequent time steps. Compared to traditional deep learning techniques, it alleviates the gradient vanishing problem to some extent and improves prediction accuracy.

[0043] Please see Figure 7This paper illustrates the Attention-based weighted approach. Because input data from different dimensions exhibits varying sensitivities to complex faults, and it's difficult to ensure all dimensions effectively reflect the characteristics of complex faults, traditional multidimensional data processing methods often perform poorly in complex fault diagnosis. Each dimension's input not only contains unique information related to the complex fault but also includes a significant amount of irrelevant noise. Simply stacking these dimensions introduces significant information redundancy, thus reducing diagnostic accuracy. Therefore, this invention introduces a self-attention mechanism that focuses more intently on the fault-related components of certain sensors, thereby more effectively extracting complex fault features from multi-sensor data.

[0044] The method in this invention treats the input set as a list of feature vectors, trains the samples through dot product operations, and generates three different representation vectors for each sample to enhance the discriminative power of feature representation. The attention mechanism can alleviate the problem of LSTM networks losing important information due to excessively long data durations by replacing random weight allocation with probabilistic methods.

[0045] The formula for calculating the self-attention mechanism is: ; in, Q , K , V These represent the query vector, key vector, and value vector, respectively, obtained from the input features through a linear transformation. , I For input data, These are the learnable weight parameters for the model; This is an attention matrix, representing the degree of correlation between the elements in each column; d k The dimension of the input data. (Through...) Q , K Attention scores are calculated using vector dot products. These scores are then standardized using the Softmax function, transforming them into a probability distribution where the sum of all weights is 1. This distribution is then compared with the corresponding vectors. V The results of the multiplication and summation are output as the attention value, which is then used to perform a weighted summation of the LSTM hidden states.

[0046] This invention calculates which sensor data (via key vectors) are obtained through a self-attention mechanism. K ) and the current fault query vector Q The most relevant value is selected, and a weighted score is obtained. Finally, this weight is compared with the value vector. VBy using weighted summation, the output generated by the model is no longer a simple stack of all sensor data, but a new, highly condensed expression that focuses on the features most relevant to the fault. This effectively avoids interference from noise and redundant information, thereby improving the diagnostic effect.

[0047] Please see Figure 8 The diagram illustrates the fault diagnosis module. The output of the self-attention function is connected to the fault diagnosis module, which consists of a fully connected layer and an activation function. It maps the attention output to a one-hot code that converts the categorical variable into a binary vector, indicating the type of composite fault in the rotor system. "0" represents a healthy state, and "1" represents a fault, corresponding to the labels in the data preprocessing. When the predicted one-hot code is completely consistent with the actual composite fault label, it indicates that the model diagnosis is correct, thus realizing the diagnosis of composite faults in the rotor system.

[0048] Please see Figure 9 This invention employs training set data for model training, utilizing a binary cross-entropy loss function and the Adam optimizer. Model parameters are optimized through backpropagation. A dropout layer is added after the fully connected layer with a dropout rate of 0.2 to prevent overfitting, achieving an average ensemble effect and increasing sparsity to prevent co-adaptation between feature detectors. Validation set data is used for model validation, with the root mean square error (RMSE) as the evaluation metric. Model training is completed when the RMSE meets the requirements. Early stopping is used to prevent overfitting; training stops when the validation set error stops decreasing or even begins to increase after 30 consecutive iterations. The model parameters corresponding to the lowest validation set error are used as the final model parameters, completing the training of the offline model for rotor composite fault diagnosis. When new abnormal vibration data is detected, fault signals are extracted in real-time as a test set and input into the model for identification, achieving online diagnosis of rotor composite faults.

[0049] In this embodiment of the invention, Sigmoid is used as the activation function of the output layer, and binary cross-entropy loss function is used as the loss function: ; The probability that the one-hot encoding predicted by the model is "1" is: The true one-hot encoding is , , For activation function, k The number of samples. When, the loss function is ,when When the value is closer to 1, the loss function is closer to 0; when When, the loss function is ,when As the value approaches 0, the loss function also approaches 0. Therefore, by iteratively optimizing the model parameters using the Adam optimizer, the aforementioned loss function can be improved. L Minimum.

[0050] In summary, this invention discloses a multi-sensor fusion-based complex fault diagnosis system and method for gas turbine rotor systems. Specifically, a one-dimensional convolutional neural network is used to extract local features from the input multi-sensor data. A long short-term memory network is used to capture temporal dependencies, and a self-attention mechanism is introduced to calculate attention weights. The signals from multiple sensors are then weighted and fused. Finally, a fully connected layer outputs key features. A binary cross-entropy loss function and an optimizer are used for model training to obtain an offline model for rotor complex fault diagnosis. When new abnormal vibration data is detected, fault signals are extracted in real time and input into the model for identification. This overcomes common problems in complex fault diagnosis, such as reliance on manual feature extraction, insufficient utilization of multi-sensor information, and low diagnostic accuracy, achieving accurate diagnosis of rotor complex faults.

[0051] The following are embodiments of the apparatus of the present invention, which can be used to execute embodiments of the method of the present invention. For details not disclosed in the apparatus embodiments, please refer to the embodiments of the method of the present invention.

[0052] Please see Figure 10 In this embodiment of the invention, a gas turbine rotor composite fault diagnosis system is provided, comprising: The data acquisition unit is used to extract fault signals based on the monitored abnormal vibration data; The diagnostic identification unit is used to perform diagnostic identification based on the extracted fault signals using a trained composite fault diagnosis model, and obtain rotor composite fault diagnosis results. The composite fault diagnosis model includes: a feature extraction module for inputting fault signals and outputting reconstructed feature maps; a temporal feature filtering module for outputting hidden states containing temporal features based on the reconstructed feature maps; an attention weighting module for weighting the hidden states and outputting key temporal features; and a fault diagnosis module for mapping the output of the attention weighting module to one-hot encoding that converts categorical variables into binary vectors to achieve rotor composite fault diagnosis.

[0053] In one embodiment of the present invention, a computer device is provided, comprising a processor and a memory. The memory stores a computer program, which includes program instructions. The processor executes the program instructions stored in the computer storage medium. The processor may be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. It is the computing and control core of the terminal, suitable for implementing one or more instructions, specifically suitable for loading and executing one or more instructions in the computer storage medium to achieve a corresponding method flow or corresponding function. The processor described in this embodiment of the present invention can be used to perform the operation of a gas turbine rotor composite fault diagnosis method.

[0054] In one embodiment of the present invention, a storage medium is provided, specifically a computer-readable storage medium (Memory), which is a memory device in a computer device used to store programs and data. It is understood that the computer-readable storage medium here can include both the built-in storage medium in the computer device and extended storage media supported by the computer device. The computer-readable storage medium provides storage space that stores the operating system of the terminal. Furthermore, the storage space also stores one or more instructions suitable for loading and execution by a processor. These instructions can be one or more computer programs (including program code). It should be noted that the computer-readable storage medium here can be high-speed RAM (Random Access Memory) or non-volatile memory, such as at least one disk storage device. The processor can load and execute one or more instructions stored in the computer-readable storage medium to implement the corresponding steps of the gas turbine rotor composite fault diagnosis method in the above embodiments.

[0055] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, optical storage, etc.) containing computer-usable program code.

[0056] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0057] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0058] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0059] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A method for diagnosing complex faults in a gas turbine rotor, characterized in that, Includes the following steps: Fault signals are extracted based on the monitored abnormal vibration data; Based on the extracted fault signals, a trained composite fault diagnosis model is used for diagnosis and identification to obtain rotor composite fault diagnosis results. The composite fault diagnosis model includes: a feature extraction module for inputting fault signals and outputting reconstructed feature maps; a temporal feature filtering module for outputting hidden states containing temporal features based on the reconstructed feature maps; an attention weighting module for weighting the hidden states and outputting key temporal features; and a fault diagnosis module for mapping the output of the attention weighting module to one-hot encoding that converts categorical variables into binary vectors to achieve rotor composite fault diagnosis.

2. The method for diagnosing complex faults in a gas turbine rotor according to claim 1, characterized in that, In the feature extraction module, a one-dimensional convolutional neural network (1DCNN) is used to extract local features from the input data. High-dimensional features are extracted step by step through convolutional layers, ReLU activation function and pooling layer, and the reconstructed feature map is output. In the temporal feature filtering module, the reconstructed feature map is used as the input of the Long Short-Term Memory (LSTM) network to capture temporal dependencies and transform the time series of the fault signal into a hidden state containing temporal features. The attention weighting module introduces a self-attention mechanism to generate query vectors. Q and key vector K The attention score is obtained by dot product operation and then compared with all corresponding vectors. V The results of the attention mechanism are multiplied and summed to output the value, which is then used to weight the hidden states. Finally, the fully connected layer outputs the key temporal features. In the fault diagnosis module, the output of the attention weighting module is mapped to a one-hot code that converts categorical variables into binary vectors through a fully connected layer and a Sigmoid activation function, thereby realizing rotor composite fault diagnosis.

3. The method for diagnosing complex faults in a gas turbine rotor according to claim 1, characterized in that, The offline training steps of the composite fault diagnosis model include: Obtain training set data; The training set data is used to train the model, and after the preset conditions are met, a well-trained composite fault diagnosis model is obtained. The training set data is obtained based on a rotor system test device, which includes: a base platform, a spindle, a disk, a motor, a sensor bracket, and a data acquisition device. The base platform serves as a support for the rotor system test apparatus and provides a vibration isolation environment during the test. The main shaft is detachably rotatably mounted on the base platform via a support bearing and bearing housing. The motor is fixedly mounted on the base platform, and the output end of the motor is connected to the main shaft via a coupling. Multiple discs are detachably mounted on the main shaft, and concentric circular grooves for mounting unbalanced blocks are formed on the discs. The base platform is equipped with the sensor bracket, on which multiple displacement sensors are mounted. These sensors are used to collect response signals in three measurement directions: perpendicular to the main shaft, parallel to the main shaft, and at a preset angle to the main shaft, and output to the data acquisition unit.

4. The method for diagnosing complex faults in a gas turbine rotor according to claim 3, characterized in that, In the process of training the model using the training set data and obtaining the trained composite fault diagnosis model after reaching the preset conditions, a dropout layer is added to prevent overfitting, an early stopping method is used to control the number of training rounds, a binary cross-entropy loss function and an Adam optimizer are used for model training, and the model parameters are optimized through the backpropagation algorithm. In addition, the model is validated using the validation set data, with the root mean square error as the evaluation index.

5. The method for diagnosing complex faults in a gas turbine rotor according to claim 3, characterized in that, The specific steps for obtaining the training set data based on the rotor system test device are as follows: The motor drives the spindle and disc to rotate at a specified speed. Multiple displacement sensors mounted on the sensor bracket collect response signals in three measurement directions: perpendicular to the main shaft, parallel to the main shaft, and at a preset angle to the main shaft. These signals are then output to the data acquisition unit to obtain multi-channel sensor acquisition signals. Specifically, an unbalanced block is installed in the concentric circular groove of the disc to simulate rotor imbalance faults, a cut is made in the optical axis section of the main shaft to simulate crack faults, a shim is placed under one end of the bearing housing to simulate misalignment faults, and the fastening bolts of the bearing housing are loosened to simulate bearing housing loosening faults. The signals acquired by multiple sensors are standardized, and the time-series data are sliced ​​using the sliding window method. The data is then labeled according to the composite fault combinations of each segment to obtain the training set data. The length of each segment is constrained by both the rotational speed and the sampling rate, and each segment contains several complete rotational cycles of the vibration signal.

6. The method for diagnosing complex faults in a gas turbine rotor according to claim 5, characterized in that, In the process of standardizing signals acquired from multiple sensors, the expression for standardization is: ； In the formula, For the first i The signals acquired by each displacement sensor The first i The mean and standard deviation of the signals from each displacement sensor. Indicates the first i A standardized signal after processing.

7. The method for diagnosing complex faults in a gas turbine rotor according to claim 5, characterized in that, When using the sliding window method to slice time series data, the window size is 40 and the step size is 1.

8. A composite fault diagnosis system for a gas turbine rotor, characterized in that, include: The data acquisition unit is used to extract fault signals based on the monitored abnormal vibration data; The diagnostic identification unit is used to perform diagnostic identification based on the extracted fault signals using a trained composite fault diagnosis model, and obtain rotor composite fault diagnosis results. The composite fault diagnosis model includes: a feature extraction module for inputting fault signals and outputting reconstructed feature maps; a temporal feature filtering module for outputting hidden states containing temporal features based on the reconstructed feature maps; an attention weighting module for weighting the hidden states and outputting key temporal features; and a fault diagnosis module for mapping the output of the attention weighting module to one-hot encoding that converts categorical variables into binary vectors to achieve rotor composite fault diagnosis.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the gas turbine rotor composite fault diagnosis method as described in any one of claims 1 to 7.

10. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the gas turbine rotor composite fault diagnosis method as described in any one of claims 1 to 7.

Images