Gas turbine fault diagnosis method based on feature extraction and adaptive fusion
By adopting the method of feature extraction and adaptive fusion in gas turbine fault diagnosis, global and local features are extracted using attention mechanism and convolutional neural network, and adaptive fusion of features is achieved through feature fusion gating mechanism, the problem of low fault diagnosis accuracy in the existing technology is solved, and the accuracy and robustness of diagnosis are improved.
Patent Information
- Application Number
- CN202411870219.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-12-18
- Publication Date
- 2025-03-18
AI Technical Summary
The existing gas turbine fault diagnosis technology has shortcomings in terms of accuracy and robustness, and it is difficult to meet the needs of accurate fault positioning and reasonable maintenance plans.
A gas turbine fault diagnosis method based on feature extraction and adaptive fusion is proposed. Global and local features are extracted through attention mechanism and convolutional neural network, and a new feature fusion gating mechanism is designed to realize adaptive fusion of features.
It improves the accuracy and robustness of gas turbine fault diagnosis, can identify faults more accurately, reduce false alarms and missed alarms, and improves the safety and economical operation of gas turbines.
Smart Images
Figure CN119646628A_ABST
Abstract
Description
Technical Field
[0001] The invention belongs to the technical field of gas turbines, and in particular relates to a gas turbine fault diagnosis method based on feature extraction and adaptive fusion. Background Art
[0002] Gas turbines have the characteristics of rapid start and stop, and have become a key source of power for power plants, ships and aircraft. Gas turbines have a complex structure, and the various components are coupled and operated in a harsh environment of high temperature and high pressure, which inevitably leads to performance degradation or even failure. On the other hand, increasing the initial temperature of the gas is the main measure to improve the workmanship of the gas turbine, which will further increase the failure probability of various components of the gas turbine. Due to the important position of gas turbines in various fields, maintenance technology must be applied and reported in advance, so traditional preventive and corrective maintenance can no longer meet the needs. Predictive maintenance based on fault diagnosis can accurately locate faults and specify reasonable maintenance plans. Therefore, the development of accurate and reliable fault diagnosis technology is the focus of current gas turbine maintenance technology.
[0003] Gas turbine fault diagnosis technology needs to meet the following two requirements. First, the fault diagnosis technology must be sensitive enough to prevent post-maintenance due to missed reports and downtime accidents due to faults. Second, the fault diagnosis technology must be stable and robust enough to prevent premature maintenance due to false alarms, which further reduces the economic efficiency of gas turbine operation. With the development of artificial intelligence and sensor technology, fault diagnosis technology based on artificial intelligence has unique advantages in meeting the above requirements, and fault diagnosis technology based on deep learning has received more and more attention.
[0004] The main principle of developing gas turbine fault diagnosis technology based on deep learning algorithms is to use neural networks to extract fault features contained in gas turbine operation data. Fault features can be divided into global features and local features. Global features and local features represent the description and analysis methods of data at different scales. How to achieve more comprehensive feature extraction is the key to fault diagnosis. Most of the existing fault diagnosis technologies are also developed for this problem, such as back propagation (BP) neural networks, convolutional neural networks (CNNs), long short-term memory neural networks, and Transformer models. In order to further improve the adaptive fusion capability of features, an attention mechanism can be introduced to apply different weights to features at different positions. However, most existing networks only extract features of a single dimension, such as CNN pays more attention to local features, and Transformer and attention pay more attention to global features. This results in the use of a single network not being able to meet the actual requirements for fault diagnosis accuracy and robustness, and it is necessary to further develop feature fusion methods.
[0005] Most of the existing global feature and local feature fusion methods use different networks in parallel or series, such as using the attention mechanism to connect CNN frameworks in series. However, the attention mechanism will still ignore the local features in the data, resulting in the inability of the series CNN to extract some of the ignored local features. The current mainstream method is to use the attention and CNN parallel frameworks to first extract features of different scales separately and then add the features. However, the simple accumulation in the existing methods cannot achieve the best feature fusion. Therefore, in order to solve the problem of low accuracy of gas turbine fault diagnosis, a new global feature and local feature extraction and adaptive fusion method is proposed, which is a technical problem that needs to be solved urgently. Summary of the invention
[0006] The purpose of the present invention is to solve the problem of low gas turbine fault diagnosis accuracy and propose a gas turbine fault diagnosis method based on feature extraction and adaptive fusion.
[0007] The technical solution adopted by the present invention to solve the above technical problems is: a gas turbine fault diagnosis method based on feature extraction and adaptive fusion, the method specifically comprising the following steps:
[0008] Step 1: Use n sensors to collect the operating data of the gas turbine at wT consecutive moments. Use the data collected by the n sensors from the (t-1)w+1th moment to the twth moment to form the sensor data of the tth time segment. The sensor data of the tth time segment is recorded as X t , X t ∈R w×n , R is a real number, w is the size of the sliding window, t = 1, 2, ..., T;
[0009] Step 2: Extract the global features of sensor data for each time segment based on the attention mechanism;
[0010] Step 3: Extract local features of sensor data for each time segment respectively;
[0011] Step 4: fuse the global features and local features of the sensor data of each time segment respectively to obtain the fused features of the sensor data of each time segment;
[0012] The fused features are linearly mapped to obtain the linear mapping results, which are then passed through the fully connected layer and the classification layer in sequence, and the fault diagnosis results are output through the classification layer.
[0013] Furthermore, the specific process of step 2 is as follows:
[0014] Sensor data X corresponding to the tth time segment t, the sensor data of the t-th time segment is passed through the linear layer to generate the query vector Q, the value vector V and the key vector K;
[0015] Then get X based on the query vector Q, value vector V and key vector K t The global feature g t :
[0016]
[0017] Among them, dk is the normalization parameter, K T represents the transpose of K, softmax(·) represents the activation function, and ⊙ represents the matrix dot product.
[0018] Furthermore, the specific process of step three is:
[0019] For the sensor data X at the tth time segment t :
[0020] h t =conv(X t ) (2)
[0021] Among them, conv(·) represents the convolution operation, h t Represents X t local features.
[0022] Furthermore, the specific process of step 4 is as follows:
[0023] For the sensor data X at the tth time segment t , calculate the global feature g t and local features h t Fusion weight coefficient:
[0024]
[0025] Where W h , W g and b u are weight parameters, u t is the weight coefficient, σ(·) is the sigmoid function, represents matrix multiplication;
[0026] For the global feature g t and local features h t Perform feature fusion:
[0027] z t ′=h t ⊙u t +g t ⊙(1-u t ) (4)
[0028] Among them, z t ′ is X t The corresponding feature fusion results.
[0029] Furthermore, the fusion features are linearly mapped to obtain a linear mapping result, and then the linear mapping result is passed through a fully connected layer, and the fault diagnosis result of the sensor data of each time segment is output through the fully connected layer; specifically:
[0030] For the sensor data X at the tth time segment t The corresponding feature fusion result z t ′, for z t ′ performs linear mapping:
[0031]
[0032] Where W z and b z are weight parameters, z t It is the result of linear mapping;
[0033] Then z t After passing through the fully connected layer and the classification layer in sequence, the classification layer outputs the sensor data X for the t-th time segment t The fault diagnosis results.
[0034] The beneficial effects of the present invention are:
[0035] The present invention proposes a gas turbine fault diagnosis method based on local feature and global feature extraction and adaptive fusion. Firstly, the global features and local features in the gas turbine operation data are extracted respectively by using the attention mechanism and the convolutional neural network. Secondly, in order to realize the adaptive fusion of local features and global features, the present invention designs a new feature fusion gating mechanism. The feature fusion gating mechanism calculates the fusion coefficient by using features of different dimensions, and normalizes the fusion coefficient by using the Sigmoid function. Then, the normalized fusion coefficient is used to perform weighted superposition on the features output by the attention mechanism and the convolutional neural network. Finally, the fused features are sent to the fully connected layer and the classification layer to realize the fault diagnosis of the gas turbine. The gas turbine fault diagnosis accuracy can be improved by adaptively fusion of global features and local features. BRIEF DESCRIPTION OF THE DRAWINGS
[0036] Figure 1 It is a flow chart of a gas turbine fault diagnosis method based on feature extraction and adaptive fusion of the present invention;
[0037] Figure 2 It is a flow chart of the adaptive fusion process of local features and global features;
[0038] Figure 3 It is the calculation result diagram of feature fusion coefficient;
[0039] Figure 4 is a fault diagnosis result diagram of the method of the present invention;
[0040] Figure 5 This is the fault diagnosis result diagram of the traditional method. DETAILED DESCRIPTION
[0041] Specific implementation method 1: Combination Figure 1 The present embodiment describes a method for gas turbine fault diagnosis based on feature extraction and adaptive fusion, and the method specifically includes the following steps:
[0042] Step 1: Use n sensors to collect the operating data of the gas turbine at wT consecutive moments. Use the data collected by the n sensors from the (t-1)w+1th moment to the twth moment to form the sensor data of the tth time segment. The sensor data of the tth time segment is recorded as X t , X t ∈R w×n , R is a real number, w is the size of the sliding window (i.e., the total number of moments corresponding to the data in each time segment), t = 1, 2, ..., T;
[0043] And the sensor data X at the tth time segment t In the equation, the data collected by the i-th sensor at the (t-1)w+1th moment is the element in the 1st row and the ith column, the data collected by the i-th sensor at the (t-1)w+2th moment is the element in the 2nd row and the ith column, and so on;
[0044] Step 2: Extract the global features of sensor data for each time segment based on the attention mechanism;
[0045] Step 3: Extract local features of sensor data for each time segment respectively;
[0046] Step 4: fuse the global features and local features of the sensor data of each time segment respectively to obtain the fused features of the sensor data of each time segment;
[0047] The fused features are linearly mapped to obtain linear mapping results, which are then passed through the fully connected layer and the classification layer in sequence, and the fault diagnosis results are output through the classification layer (the fault diagnosis results include whether a fault is about to occur and the type of fault).
[0048] The local features and global features in the gas turbine degradation data refer to the description and analysis methods of data at different scales. The various components of the gas turbine are mutually coupled and affect each other during operation, so the degradation of a certain component will be reflected in the sensors at other positions. For example, when the high-pressure turbine blade has a corrosion failure, resulting in a decrease in efficiency, the corresponding speed sensor and temperature sensor will also have obvious deviations. At this time, the deviation is a global prominent feature in all sensor data. At the same time, other positions such as the low-pressure turbine speed will also have slight deviations, which is a local fault. Therefore, the present invention proposes a fault diagnosis method with strong extraction and adaptive fusion capabilities for both local features and global features. The method of the present invention can be run once at each moment, and the data to be processed at the current moment is updated by continuously collecting the operating data of the gas turbine, that is, when the operating data of the new moment arrives, the algorithm will automatically discard the operating data of the historical moment to ensure that the total number of moments corresponding to each processing data is always wT. The method of the present invention can perform real-time fault diagnosis on the gas turbine, and when the fault detection result corresponding to the sensor data of any time segment is a fault, maintenance should be prepared.
[0049] Specific implementation method 2: This implementation method is different from the specific implementation method 1 in that the specific process of step 2 is as follows:
[0050] Sensor data X corresponding to the tth time segment t , the sensor data of the t-th time segment is passed through the linear layer to generate the query vector Q, the value vector V and the key vector K;
[0051] Q=X t W Q
[0052] K=X t W K
[0053] V=X t W V
[0054] Among them, W Q , W K and W V All are learnable weight parameters;
[0055] Then get X based on the query vector Q, value vector V and key vector K t The global feature g t :
[0056]
[0057] Among them, dk is the normalization parameter, K Trepresents the transpose of K, softmax(·) represents the activation function, and ⊙ represents the matrix dot multiplication, that is, the corresponding elements are multiplied.
[0058] The other steps and parameters are the same as those in the first embodiment.
[0059] Similarly, the method of this embodiment is used to obtain the global features of the sensor data of each time segment.
[0060] Specific implementation method three: This implementation method is different from specific implementation methods one or two in that the specific process of step three is as follows:
[0061] For the sensor data X at the tth time segment t :
[0062] h t =conv(X t ) (2)
[0063] Among them, conv(·) represents the convolution operation, h t Represents X t local features.
[0064] The other steps and parameters are the same as those in the first or second embodiment.
[0065] Similarly, the method of this embodiment is used to obtain the local features of the sensor data of each time segment.
[0066] Specific implementation method four: Combination Figure 2 This embodiment is different from the first to third embodiments in that the specific process of step 4 is as follows:
[0067] For the sensor data X at the tth time segment t , calculate the global feature g t and local features h t Fusion weight coefficient:
[0068]
[0069] Where W h , W g and b u are all learnable weight parameters, u t is the weight coefficient, u t ∈(0,1), σ(·) is the sigmoid function, Represents matrix multiplication (it should be noted that the matrix multiplication here is a standard matrix multiplication, with the matrix as For example, the number of columns of matrix A should be equal to the number of rows of matrix B, that is, the element-by-element product of rows and columns and the sum of them);
[0070] For the global feature g t and local features h t Perform feature fusion:
[0071] z t ′=h t ⊙u t +g t ⊙(1-u t ) (4)
[0072] Among them, z t ′ is X t The corresponding feature fusion results.
[0073] The other steps and parameters are the same as those in Specific Embodiments 1 to 3.
[0074] Similarly, the method of this embodiment is used to obtain the feature fusion results of the sensor data of each time segment. During the training process, the present invention uses error back propagation to automatically update all parameters and fusion coefficients in the network, so that the fusion ratio of the two features can be adaptively optimized, and then the features of different dimensions are adaptively fused. In the traditional fault diagnosis method, only the features used by the parallel network are simply added. In comparison, the adaptive feature fusion method of the present invention can effectively improve the diagnosis effect and ensure the safety of gas turbine operation.
[0075] Specific implementation method five: This implementation method is different from any one of specific implementation methods one to four in that the linear mapping is performed on the fusion features to obtain a linear mapping result, and then the linear mapping result is passed through a fully connected layer, and the fault diagnosis result of the sensor data of each time segment is output through the fully connected layer; specifically:
[0076] For the sensor data X at the tth time segment t The corresponding feature fusion result z t ′, for z t ′ performs linear mapping:
[0077]
[0078] Where W z and b z are all learnable weight parameters, z t It is the result of linear mapping;
[0079] Then z t After passing through the fully connected layer and the classification layer in sequence, the classification layer outputs the sensor data X for the t-th time segment t The fault diagnosis results.
[0080] The other steps and parameters are the same as those in Specific Embodiments 1 to 4.
[0081] Similarly, the method of this embodiment is used to obtain the fault detection results of the sensor data of each time segment. The main process of the algorithm of the present invention is shown in Table 1:
[0082] Table 1
[0083]
[0084] The method of the present invention can realize the extraction and adaptive fusion of local features and global features of gas turbine fault data, and realize high-precision fault diagnosis.
[0085] Experimental Section
[0086] In order to verify the effectiveness of the method of the present invention, a verification experiment was designed on the gas turbine simulation data. First, five typical fault types were simulated: normal gas turbine, fan damage, high-pressure compressor corrosion, flame tube crack, and high-pressure turbine blade corrosion. Each type of state training data has 7200 samples, and the test set has 1800 samples. The calculation results of the feature fusion coefficient are shown in Figure 2. Figure 3 shown.
[0087] Depend on Figure 3 It can be seen that the value of the feature fusion coefficient is between (0, 1), which can adaptively fuse global features and local features. The fault diagnosis results of the method of the present invention on the test set are as follows: Figure 4 As shown in Figure 2, the fault diagnosis results of the traditional method on the test set are as follows: Figure 5 As shown, after calculation, the average diagnostic accuracy of the method of the present invention on the test set is 99.93%, and the average diagnostic accuracy of the traditional method on the test set is only 97.78%. Therefore, the method of the present invention can effectively improve the fault diagnosis accuracy and improve the safety of gas turbine operation.
[0088] The above calculation examples of the present invention are only used to explain the calculation model and calculation process of the present invention in detail, and are not intended to limit the implementation methods of the present invention. For ordinary technicians in the relevant field, other different forms of changes or modifications can be made based on the above description. It is impossible to list all the implementation methods here. All obvious changes or modifications derived from the technical solution of the present invention are still within the scope of protection of the present invention.
Claims
1. A gas turbine fault diagnosis method based on feature extraction and adaptive fusion, characterized in that: The method specifically comprises the following steps: Step 1: Use n sensors to collect the operating data of the gas turbine at wT consecutive moments. Use the data collected by the n sensors from the (t-1)w+1th moment to the twth moment to form the sensor data of the tth time segment. The sensor data of the tth time segment is recorded as X t , X t ∈R w×n , R is a real number, w is the size of the sliding window, t = 1, 2, ..., T; Step 2: Extract the global features of sensor data for each time segment based on the attention mechanism; Step 3: Extract local features of sensor data for each time segment respectively; Step 4: fuse the global features and local features of the sensor data of each time segment respectively to obtain the fused features of the sensor data of each time segment; The fused features are linearly mapped to obtain the linear mapping results, which are then passed through the fully connected layer and the classification layer in sequence, and the fault diagnosis results are output through the classification layer.
2. A gas turbine fault diagnosis method based on feature extraction and adaptive fusion according to claim 1, characterized in that: The specific process of step 2 is as follows: Sensor data X corresponding to the tth time segment t , the sensor data of the t-th time segment is passed through the linear layer to generate the query vector Q, the value vector V and the key vector K; Then get X based on the query vector Q, value vector V and key vector K t The global feature g t : Among them, dk is the normalization parameter, K T represents the transpose of K, softmax(·) represents the activation function, and ⊙ represents the matrix dot product.
3. A gas turbine fault diagnosis method based on feature extraction and adaptive fusion according to claim 2, characterized in that: The specific process of step three is: For the sensor data X at the tth time segment t : h t =conv(X t ) (2) Among them, conv(·) represents the convolution operation, h t Represents X t local features.
4. A gas turbine fault diagnosis method based on feature extraction and adaptive fusion according to claim 3, characterized in that: The specific process of step 4 is as follows: For the sensor data X at the tth time segment t , calculate the global feature g t and local features h t Fusion weight coefficient: Where W h , W g and b u are weight parameters, u t is the weight coefficient, σ(·) is the sigmoid function, represents matrix multiplication; For the global feature g t and local features h t Perform feature fusion: z t ′=h t ⊙u t +g t ⊙(1-u t ) (4) Among them, z t ′ is X t The corresponding feature fusion results.
5. A gas turbine fault diagnosis method based on feature extraction and adaptive fusion according to claim 4, characterized in that: The linear mapping is performed on the fusion features to obtain a linear mapping result, and then the linear mapping result is passed through a fully connected layer, and the fault diagnosis result of the sensor data of each time segment is output through the fully connected layer; specifically: For the sensor data X at the tth time segment t The corresponding feature fusion result z t ′, for z t ′ performs linear mapping: Where W z and b z are weight parameters, z t It is the result of linear mapping; Then z t After passing through the fully connected layer and the classification layer in sequence, the classification layer outputs the sensor data X for the t-th time segment t The fault diagnosis results.