A method for fault diagnosis of aviation inverter based on high-order characteristic tensor decomposition
By combining high-order feature tensor decomposition and support tensor machine, the problem of unutilized multi-dimensional structural characteristics in aviation inverter fault diagnosis is solved, enabling more efficient fault feature analysis and diagnosis, and improving the safety of aviation power supply systems.
Patent Information
- Application Number
- CN202211396980.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-11-09
- Publication Date
- 2026-02-27
- Estimated Expiration
- 2042-11-09
AI Technical Summary
Existing technologies fail to effectively utilize the multidimensional structural characteristics in the fault diagnosis of aviation inverters, resulting in insufficient data analysis and an inability to effectively uncover the multidimensional hidden structure and components of aviation inverters, thus affecting the accuracy and efficiency of fault diagnosis.
A high-order feature tensor decomposition method is adopted. By preprocessing the electrical signals of the aviation inverter, a high-order feature tensor decomposition model is constructed. The factor matrix is updated using Tucker decomposition and NEU algorithm to reconstruct a higher-order fault feature sample set. Finally, a support tensor machine is used for fault diagnosis.
Without increasing the time complexity of the algorithm, it can more comprehensively analyze fault characteristic information, alleviate the data sparsity problem, realize fast and accurate fault diagnosis of aviation inverters, and improve the safety of aviation power supply systems.
Smart Images
Figure CN115684805B_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The application relates to an aviation inverter fault diagnosis method based on high-order characteristic tensor decomposition and belongs to the technical field of aviation power supply system fault diagnosis. BACKGROUND
[0002] The aviation inverter is an important component of the aviation secondary power supply, and the performance state directly affects the performance of the aviation power supply system. Therefore, aviation inverter fault diagnosis has important engineering use value for the safety and health management of the aviation power supply system.
[0003] At present, the sample data of the aviation inverter is distributed with useful structures and components in different dimensions, and these structures and components are often local. The implicit components or structures can be found through matrix or tensor decomposition. When data is represented by vectors, the set of data and linear transformation can be represented by matrices, and matrix decomposition is a powerful tool for vector data analysis and processing. Typical methods include singular value decomposition (SVD) and principle component analysis (PCA). For tensor data, useful structures and components are also sparsely distributed in multidimensional space. However, for a long time, the multidimensional structure characteristics of the tensor itself have been ignored, and the original tensor data has been rearranged into a vector in a certain order, and then analyzed and processed by using matrix analysis theory. This vectorization method will destroy the multidimensional structure of the original data and cannot effectively mine the multidimensional implicit structure and components in the tensor data. The tensor has the multidimensional data structure characteristics, and the implicit components and useful information with physical meaning can be found by fully utilizing the structure characteristics of the multidimensional data. Essentially, the tensor and its tensor algebra are a representation and processing method of high-dimensional data, and are a more natural and intuitive representation and processing method. In general, compared with the vector representation method, the tensor analysis method based on tensor algebra has more advantages. The multidimensional structure characteristics of the data can be maintained, the local spatial correlation of the data and the correlation and complementarity between modes can be fully utilized, the dimension can be effectively reduced through tensor decomposition, and more effective data representation can be obtained, and under certain conditions, the implicit components and data representation with certain physical meaning can be uniquely found. On the basis of tensor decomposition, the application provides an aviation inverter fault diagnosis method based on high-order characteristic tensor decomposition, which can mine high-order characteristic information of fault samples without improving the time complexity of the algorithm, and can further expand the fault sample information. SUMMARY
[0004] The application aims to provide a high-order feature tensor decomposition-based aviation inverter fault diagnosis method, which can obtain higher-order similar matrices and further obtain high-order similar sample sets through continuous iteration and updating without improving the time complexity of the original sample set, thereby providing data support for aviation inverter fault diagnosis and improving the safety of aviation transportation.
[0005] The application aims to provide a high-order feature tensor decomposition-based aviation inverter fault diagnosis method, which can obtain higher-order similar matrices and further obtain high-order similar sample sets through continuous iteration and updating without improving the time complexity of the original sample set, thereby providing data support for aviation inverter fault diagnosis and improving the safety of aviation transportation.
[0006] The application aims to provide a high-order feature tensor decomposition-based aviation inverter fault diagnosis method, which can obtain higher-order similar matrices and further obtain high-order similar sample sets through continuous iteration and updating without improving the time complexity of the original sample set, thereby providing data support for aviation inverter fault diagnosis and improving the safety of aviation transportation.
[0007] Step 1: pre-process the electrical signals collected by the aviation inverter under different working conditions, and extract the time-domain fault features, frequency-domain fault features and time-frequency-domain fault features of the electrical signals as the fault feature sample set;
[0008] Step 2: divide the fault feature sample set into a training sample set and a test sample set, and normalize the training sample set and the test sample set;
[0009] Step 3: perform tensor decomposition on the normalized training sample set to obtain three factor matrices U (1) , U (2) , U (3) , perform singular value decomposition on the factor matrix U (1) to obtain the network representation vector R and the context representation vector C of the factor matrix U (1) ; then update R and C to R' and C' through the NEU algorithm, and further obtain the reconstructed factor matrix Finally, the same steps are taken for to obtain the updated and further reconstruct the tensor feature , which is a higher-order fault feature sample set;
[0010] Step 4: input the normalized test sample into the support tensor machine to obtain the aviation inverter fault diagnosis result.
[0011] Further, the fault feature sample set in step 1 uses the time-domain fault features, frequency-domain fault features and time-frequency-domain fault features of the electrical signals as feature columns, which include the frequency mean, frequency root mean square, mean, wavelet scale entropy, variance, peak-to-peak value, kurtosis factor, waveform factor, wavelet energy entropy and wavelet singular entropy of the electrical signals.
[0012] Further, the normalization process in step 2 is as follows:
[0013] The fault feature sample set is divided into a training sample set and a test sample set in a ratio of 8:2; and the zero mean normalization method is used to normalize the training set and the test set respectively:
[0014]
[0015] Wherein, x' represents the data after normalization processing; x is input sample data; μ is sample mean; σ is sample standard deviation.
[0016] Further, in step 3, the tensor decomposition selects Tucker decomposition,
[0017] Tensor representation of aviation inverter fault diagnosis The low rank approximation is represented by using Tucker decomposition operation as:
[0018]
[0019] Wherein
[0020] Further, in step 3, the NEU algorithm is updated by matrix conversion to obtain a higher order factor matrix representation form;
[0021] The updated first order similarity is shown in formula (4) and formula (5):
[0022] R'=R+λU1·R (4)
[0023]
[0024] Wherein λ is a hyperparameter, and λ satisfies λ∈(0, 0.5];
[0025] The updated second order similarity is shown in formula (6) and formula (7):
[0026] R'=R+λ1U1·R+λ2U1·(U1·R) (6)
[0027]
[0028] Wherein λ1 and λ2 are hyperparameters, and λ1λ2 satisfies λ1λ2∈(0, 0.5], through continuous iteration and update, third order, fourth order, n order similarity is obtained, and a high-order approximate factor matrix is further obtained.
[0029] Further, in step 4, the normalized test sample is input into the STM aviation inverter diagnosis model, the 10-fold cross-validation method is used to traverse all test set samples, the average value of 10 test results is taken, and more accurate fault diagnosis results are obtained.
[0030] The beneficial effects of the present application are that:
[0031] 1. By analyzing fault feature information from a higher dimension, the limitations of two-dimensional feature information are avoided, resulting in more comprehensive fault feature information.
[0032] 2. By reconstructing tensor feature information, the problem of data sparsity can be effectively alleviated.
[0033] 3. It can enable aircraft inverters to generate fault diagnosis results quickly and accurately, helping to manage the health of aircraft power supply systems and thus improving the safety of air transport. Attached Figure Description
[0034] Figure 1 This is a schematic diagram of the overall process of a fault diagnosis method for aero-inverters based on high-order feature tensor decomposition.
[0035] Figure 2 This is a schematic diagram of tensor decomposition.
[0036] Figure 3 This is a schematic diagram of tensor expansion.
[0037] Figure 4 A diagram illustrating an example of tensor decomposition. Detailed Implementation
[0038] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments.
[0039] This embodiment proposes a fault diagnosis method for aero-inverters based on high-order characteristic tensor decomposition, such as... Figure 1 As shown, firstly, an initial data sample set is constructed based on the electrical signal data collected by the aircraft inverter. Electrical signal features are extracted to form a fault feature set. Based on this, the fault feature set undergoes data preprocessing to obtain training and test sets, which are then normalized. Next, a high-order feature tensor decomposition model is constructed to learn the implicit relationship between the aircraft inverter's electrical signals and fault modes. Finally, the fault diagnosis results of the aircraft inverter are obtained through a support tensor machine. The specific steps are as follows:
[0040] Step 1: Preprocess the electrical signals collected by the aircraft inverter under different operating conditions, and extract the time-domain fault features, frequency-domain fault features, and time-frequency-domain fault features of the electrical signals as fault feature samples, thereby constructing a three-dimensional initial data sample set.
[0041] The electrical signals of an aviation inverter mainly include voltage signals and current signals. As an example, this embodiment uses electrical signal samples for six operating conditions as shown in Table 1.
[0042] Table 1 Operating Conditions of Aircraft Inverters
[0043] Operating condition number Change 1 Input +28V DC voltage, output power 200VA, load resistance 66Ω 2 Input +22V DC voltage, output power 200VA, load resistance 66Ω 3 Input +30V DC voltage, output power 200VA, load resistance 66Ω 4 Input +28V DC voltage, output power 300VA, load resistance 44Ω 5 Input +22V DC voltage, output power 300VA, load resistance 44Ω 6 Input +30V DC voltage, output power 300VA, load resistance 44Ω
[0044] As an example, the electrical signal data of the aviation inverter used in this embodiment includes the fault modes of different states as shown in Table 2, each fault mode corresponds to a set of continuously collected electrical signals, and each set of signal samples corresponds to an electrical circuit fault mode label.
[0045] Table 2 Circuit fault mode
[0046]
[0047]
[0048] The time domain fault features, frequency domain fault features and time-frequency domain fault features of the electrical signals are extracted as feature columns, such as the frequency mean, frequency root mean square, mean, wavelet scale entropy, variance, peak-to-peak value, kurtosis factor, waveform factor, wavelet energy entropy and wavelet singular entropy of the voltage / current signal, thereby forming a three-dimensional fault feature sample set.
[0049] Step 2: Divide the fault feature sample set into a training sample set and a test sample set, and normalize the training sample set and the test sample set.
[0050] In this embodiment, the three-dimensional fault feature sample set is divided into a training sample set and a test sample set in a ratio of 8:2, and the data in the training sample set and the test set are all labeled data samples.
[0051] The zero-mean normalization method is used to normalize the training sample set and the test sample set respectively, thereby generating a dimensionless data set, and the normalization processing formula is as follows:
[0052]
[0053] Where x' represents the normalized data, x is the input sample data, μ is the sample mean, and σ is the sample standard deviation.
[0054] Step 3: Use the normalized training sample set to construct a high-order feature tensor decomposition model. The construction process of the high-order feature tensor decomposition model is as follows:
[0055] Step 3-1, use the training sample set to construct a three-dimensional fault feature matrix A, three dimensions are respectively working condition, fault feature and sample number.
[0056] Step 3-2, tensor decomposition is performed on A to obtain three factor matrices U (1) , U (2) , U (3); Since the numerical value of each factor matrix is the same, changing one factor matrix will change the other two factor matrices accordingly. Therefore, the factor matrix U1 is singular value decomposed (SVD) to obtain the factor matrix U (1) , the network representation vector R and the context representation vector C.
[0057] The tensor decomposition selects Tucker decomposition, also known as higher-order SVD (HOSVD), which is to decompose the tensor into a core tensor and a mode product of multiple orthogonal factor matrices, as shown in Figure 2 The expansion process is shown in Figure 3 The actual example expansion process is shown in Figure 4 Tucker decomposition can be seen as a form of multiplication of a core tensor and N factor matrices on mode-n. An N-order tensor, its low-rank approximation can be represented as a truncated Tucker decomposition, as follows:
[0058] X≈G×1U (1) ×2U (2) …× N U (N) (2)
[0059] The factor matrices on mode-1, mode-2, …, mode-n subspaces represent the main components of the tensor in each mode, and any two columns of them are orthogonal to each other. The truncated core tensor in the low-rank approximation process represents the degree of correlation between each subspace component, and the core tensor can be calculated as follows: G = X ×1U (1)T ×2U (2)T …× N U (N)T
[0060] The embodiment relates to tensor representation of aviation inverter fault diagnosis The low-rank approximation using Tucker decomposition operation is represented as:
[0061]
[0062] Among them
[0063] Step 3-3, update R and C using the NUE algorithm to obtain R' and C', and further obtain the reconstructed factor matrix Through The same steps of Further reconstruct the tensor feature That is, a higher-order fault feature sample set.
[0064] The core idea of the Network Embedding Updata (NEU) updating method is to update the factor matrix through simple matrix conversion, thereby obtaining a higher-order factor matrix representation form.
[0065] The updated first-order similarity is shown in equations (4) and (5):
[0066] R' = R + lU1 R (4)
[0067]
[0068] wherein l is a hyperparameter, satisfying l e (0, 0.5], U1 is a tensor decomposition factor matrix, and R and C are network representation vectors and context representation vectors of the factor matrix.
[0069] The updated second-order similarity is shown in equations (6) and (7):
[0070] R' = R + l1U1 R + l2U1 (U1 R) (6)
[0071]
[0072] wherein l1 and l2 are hyperparameters, satisfying l1l2 e (0, 0.5], and through continuous iteration and updating, a third-order, fourth-order, or n-order similarity can be obtained, and a high-order approximate factor matrix can be obtained.
[0073] Step 4: input the normalized test sample into a support tensor machine (STM) to obtain an aviation inverter fault diagnosis result. A 10-fold cross-validation method is used to traverse all test set samples, and the average value of 10 test results is taken to obtain a more accurate fault diagnosis result, which is used to verify the classification effect of the support tensor machine.
[0074] It can be understood that, for those skilled in the art, equivalent replacements or changes can be made according to the technical solutions and inventive concepts of the present application, and all such changes or replacements shall belong to the protection scope of the appended claims of the present application.
Claims
1. A fault diagnosis method for aero-inverters based on high-order characteristic tensor decomposition, characterized in that... Includes the following steps: Step 1: Preprocess the electrical signals collected by the aircraft inverter under different operating conditions, and extract the time-domain fault features, frequency-domain fault features, and time-frequency-domain fault features of the electrical signals as a fault feature sample set; Step 2: Divide the fault feature sample set into a training sample set and a test sample set, and normalize the training sample set and the test sample set. Step 3: Construct a three-dimensional fault feature matrix A using normalized training samples, and perform tensor decomposition on the three-dimensional fault feature matrix A to obtain three factor matrices U. (1) U (2) U (3) The factor matrix U is subjected to singular value decomposition. (1) Perform singular value decomposition to obtain the factor matrix U. (1) The network representation vector R and the context representation vector C are obtained; then R and C are updated using the NEU algorithm to obtain R' and C', and thus the reconstruction factor matrix is obtained. Finally passed The same steps were used to update the result. Then reconstruct tensor features This is a higher-order fault feature sample set; the NEU algorithm updates the factor matrix through matrix transformation to obtain a higher-order factor matrix representation. The updated first-order similarity is shown in equations (4) and (5): R'=R+λU1·R (4) Where λ is a hyperparameter satisfying λ∈(0,0.5]; The updated second-order similarity is shown in equations (6) and (7): R'=R+λ1U1·R+λ2U1·(U1·R) (6) Where λ1 and λ2 are hyperparameters, satisfying λ1, λ2∈(0, 0.5]. Through continuous iteration and updating, third-order, fourth-order, and n-order similarities are obtained, and then higher-order approximate factor matrices are obtained. Step 4: Input the normalized test samples into the support tensor machine to obtain the fault diagnosis results of the aviation inverter.
2. The method for fault diagnosis of aero-inverters based on high-order feature tensor decomposition according to claim 1, characterized in that... The fault feature sample set mentioned in step 1 uses the time-domain fault features, frequency-domain fault features, and time-frequency-domain fault features of the electrical signal as feature columns, including the frequency mean, root mean square frequency, mean, wavelet scale entropy, variance, peak-to-peak value, kurtosis factor, waveform factor, wavelet energy entropy, and wavelet singular entropy of the electrical signal.
3. The method for fault diagnosis of aero-inverters based on high-order feature tensor decomposition according to claim 1, characterized in that... The normalization process in step 2 is implemented as follows: The fault feature sample set was divided into a training sample set and a test sample set in an 8:2 ratio; zero-mean normalization was used to normalize the training set and the test set respectively. Where x' represents the normalized data; x is the input sample data; μ is the sample mean; and σ is the sample standard deviation.
4. The method for fault diagnosis of aero-inverters based on high-order feature tensor decomposition according to claim 1, characterized in that... In step 3, the tensor decomposition uses Tucker decomposition. Tensor representation for fault diagnosis of aviation inverters The low-rank approximation using Tucker decomposition is as follows: A≈S×1U (1) ×2U (2) ×3U (3) =[S;U (1) ,IN (2) ,IN (3) ] (3) in 5. The method for fault diagnosis of aero-inverters based on high-order feature tensor decomposition according to claim 1, characterized in that... In step 4, the normalized test samples are input into the STM aviation inverter diagnostic model. The 10-fold cross-validation method is used to traverse all test set samples and take the average of 10 test results to obtain a more accurate fault diagnosis result.