Adaptive multi-domain disentanglement manifold network-based fault detection method for multi-working condition of inverter system

Abstract

This invention discloses a multi-condition fault detection method for inverter systems based on adaptive multi-domain unentangled manifold networks. Addressing the issue of multi-condition operation and scarce samples for some conditions in inverter systems, the method first constructs a dual-path encoder comprising a shared encoder and a private encoder to explicitly separate common features that do not change with the operating conditions from unique features that do change with the operating conditions. Then, an adaptive weighting mechanism based on reconstruction differences is used to quantify the reference value of the source domains, selectively increasing the proportion of highly correlated source domains to suppress negative migration. Next, manifold regularization techniques are used to constrain the latent space, constructing a geometrically coherent healthy manifold benchmark. Finally, the distance of the test sample projected onto the local manifold space is calculated as a fault monitoring index to achieve the fault detection objective for the inverter system.

Description

Technical Field

[0001] This invention belongs to the field of power electronic system fault detection technology. More specifically, it relates to a multi-condition fault detection method for inverter systems based on adaptive multi-domain unentangled manifold networks, which is suitable for monitoring inverter faults with data imbalance under multiple operating conditions. Background Technology

[0002] With the rapid development of new energy power generation, electric vehicles, and industrial drives, the importance of power electronic systems in modern society is becoming increasingly significant. As a key component in the DC-AC conversion process of the power grid, the inverter system's operating status directly affects the reliability and safety of the entire system. A failure in the inverter system can not only lead to equipment downtime and huge economic losses, but also potentially trigger serious safety accidents. Therefore, researching timely and effective fault detection technologies is of significant engineering and social value for ensuring the safe and stable operation of power electronic systems.

[0003] However, in real-world applications, inverter systems often operate under multiple conditions. Influenced by load fluctuations, changes in reference commands, and topology adjustments, the statistical distribution of system data undergoes significant dynamic shifts. Existing data-driven fault detection methods typically assume that training and test data follow the same distribution and require a large amount of complete historical data to train the model. However, in practical engineering, due to strict safety constraints and operating costs, it is difficult to obtain complete data for all different operating conditions, especially for certain marginal or new operating conditions, where only a very small number of normal operating samples can often be collected. This complex situation of "multiple operating conditions" and "small sample size" makes it difficult for existing methods to capture the inherent correlations between operating conditions. They are prone to learning biased latent feature representations, resulting in "negative transfer" or the inability to establish reliable normal behavior benchmarks, severely impacting the accuracy and robustness of fault detection.

[0004] Therefore, a new method is urgently needed that can adapt to changes in multiple operating conditions and achieve efficient fault detection even when target operating condition samples are scarce. Transfer learning has been introduced into the field of fault detection because it can solve the problem of large differences in data distribution across operating conditions. It aims to mine and utilize common knowledge between operating conditions with large sample sizes (source domain) and operating conditions with small sample sizes (target domain). Methods for feature transfer focus on finding a common feature space that is invariant to different operating conditions through feature transformation or subspace alignment techniques, attempting to bring the data distribution of the source and target domains closer or unify them, thereby reducing the impact of distribution shift. Methods for parameter transfer mainly assume that the models in the source and target domains share certain parameters or prior distributions. They reuse model parameters trained on rich historical data (such as the weights of neural networks) and fine-tune them using target domain data to adapt to the new operating environment. Single-source domain transfer or simple multi-source merging methods typically only focus on transferring knowledge from a specific historical operating condition, or simply mix all historical operating condition data as a whole source domain, without finely distinguishing the contributions of different historical operating conditions. However, existing transfer learning methods ignore the relationships between different operating conditions, forcibly transfer irrelevant knowledge, causing negative transfer and thus reducing detection performance. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of the prior art and provide a multi-condition fault detection method for inverter systems based on adaptive multi-domain unentangled manifold networks. This method utilizes unentangled networks to separate condition-specific features from common features, and constructs a geometrically coherent latent space based on manifold regularization, thereby achieving efficient and sensitive fault detection even when target condition samples are scarce.

[0006] To achieve the above-mentioned objectives, this invention provides a multi-condition fault detection method for inverter systems based on adaptive multi-domain unentangled manifold networks, characterized by comprising the following steps:

[0007] (1) Data acquisition and processing;

[0008] (2) Adaptive weighted calculation;

[0009] (3) Construct an adaptive multi-domain unentangled manifold network model;

[0010] (4) Train an adaptive multi-domain unentangled manifold network model;

[0011] (5) Calculate the manifold deviation anomaly score;

[0012] (6) Solving the fault detection threshold based on kernel density estimation;

[0013] (7) Online fault monitoring and status determination;

[0014] The objective of this invention is achieved as follows:

[0015] This invention presents a multi-condition fault detection method for inverter systems based on adaptive multi-domain unentangled manifold networks. Addressing the issue of scarce samples for some operating conditions in inverter systems, the method first constructs a dual-path encoder comprising a shared encoder and a private encoder to explicitly separate common features that remain unchanged with each operating condition from unique features that change with each condition. Then, it quantifies the reference value of the source domains through an adaptive weighting mechanism based on reconstruction differences, selectively increasing the proportion of highly correlated source domains to suppress negative migration. Next, it utilizes manifold regularization techniques to constrain the latent space, constructing a geometrically coherent healthy manifold benchmark. Finally, it calculates the distance of the test sample projected onto the local manifold space as a fault monitoring indicator to achieve the purpose of fault detection for the inverter system.

[0016] Meanwhile, the multi-condition fault detection method for inverter systems based on adaptive multi-domain unentangled manifold networks of the present invention also has the following beneficial effects:

[0017] (1) The present invention proposes an adaptive weighting strategy, which can automatically adjust the source domain migration weight according to the reconstruction differences between working conditions, effectively solve the negative migration problem caused by irrelevant source domain influence under multiple working conditions, and significantly improve the robustness of the model in complex variable working conditions.

[0018] (2) This invention utilizes the feature deentanglement architecture and manifold geometric constraints to successfully establish a reliable health benchmark under extreme conditions where there are only a very small number of normal samples in the target working condition. This overcomes the limitation of traditional methods that rely on complete target domain data and solves the problem of scarce samples in certain working conditions in actual industrial scenarios.

[0019] (3) The present invention constructs a manifold deviation index anomaly score index, which uses the local geometric structure information of the data to capture deep feature differences. Compared with traditional Euclidean distance or statistical measures, this index has higher detection sensitivity and accuracy for early weak faults that cause small changes in manifold structure. Attached Figure Description

[0020] Figure 1 This is a flowchart of the multi-condition fault detection method for inverter systems based on adaptive multi-domain unentangled manifold networks of the present invention;

[0021] Figure 2 This is a schematic diagram of the experimental platform;

[0022] Figure 3 These are voltage diagrams under different operating conditions;

[0023] Figure 4 It is a cluster distribution diagram of sample points;

[0024] Figure 5 This is a schematic diagram of adaptive weighted calculation;

[0025] Figure 6 This is a diagram of the adaptive multi-domain unentangled manifold network model structure; Detailed Implementation

[0026] The specific embodiments of the present invention will now be described with reference to the accompanying drawings to enable those skilled in the art to better understand the invention. It should be particularly noted that in the following description, detailed descriptions of known functions and designs that might obscure the main content of the invention will be omitted here.

[0027] Example

[0028] In this embodiment, as Figure 1 As shown, the present invention discloses a multi-condition fault detection method for inverter systems based on adaptive multi-domain unentangled manifold networks, comprising the following steps:

[0029] (1) Data acquisition and processing;

[0030] (1.1) In this embodiment, using Figure 2 The experimental platform shown collects three-phase voltage data at the load end of the inverter system at equal intervals under different operating conditions, with the number of sampling points being [number missing]. Point; record the first The three-phase voltage data sequence collected under various operating conditions is represented as follows: , among which, the The first working condition Three-phase voltage data collected at each moment Represented as , Indicates the first The first working condition The data collected at the [time]th moment One sampling point, Represents a three-phase circuit. Indicates the first Total number of sampling times under various working conditions;

[0031] In this embodiment, data for three operating conditions were collected, as shown in Table 1. All operating conditions had a frequency of 400Hz and a load resistance of 50Ω. The DC bus voltage for the first operating condition was 50V, for the second operating condition it was 60V, and for the third operating condition it was 80V. The voltage diagrams for the three operating conditions are shown below. Figure 3 As shown in (a), (b), and (c) in the figure;

[0032] Table 1. Operating Condition Table;

[0033] frequency load resistor DC bus voltage Operating Condition 1 400Hz 50Ω 50V Operating Condition 2 400Hz 50Ω 60V Operating Condition 3 400Hz 50Ω 80V

[0034] (1.2) Traverse the three-phase voltage data sequence under each operating condition The three-phase voltage data at each sampling time are extracted sequentially. The features include: frequency, mean, peak-to-peak value, RMS value, margin, and kurtosis. These six features are then used as the first set of features. Feature vector at each sampling time This allows us to obtain feature datasets for various operating conditions. In this embodiment, 4000 sample points were collected for 50V and 60V operating conditions, and 10 sample points were collected for 80V operating conditions. T-SNE clustering was used, and the cluster distribution of the sample points is as follows: Figure 4 As shown;

[0035] (1.3) Traverse the three-phase voltage data sequence under each operating condition Based on the total number of sampling times corresponding to each working condition With preset threshold Perform domain partitioning, if Then the first The operating condition is defined as the source domain operating condition, denoted as . Otherwise, the first The type of working condition is defined as the target domain working condition, denoted as In this embodiment, the 50V and 60V operating conditions with 4000 sample points are set as the source domain, and the 80V operating condition with 10 sample points is set as the target domain.

[0036] (2) Adaptive weighted calculation, the process is as follows Figure 5 As shown;

[0037] (2.1) Training source domain-specific autoencoders: For each source domain Using the corresponding feature dataset within this domain Training an independent source domain autoencoder ,in, Representing the Training parameters for a source domain autoencoder;

[0038] (2.2) Calculate the reconstruction differences of the target domain under each source domain autoencoder: The feature dataset corresponding to the target domain... The data are input into each of the source domain autoencoders trained in step (2.1), and the target domain sample is calculated according to the following formula at the i-th position. Average reconstruction difference under each source domain model :

[0039] ;

[0040] in, The first term of the target domain Each sample, i.e., the target domain The corresponding feature dataset in the first Feature vectors at each sampling time; For temperature coefficient, Represents the L2 norm. Indicates the first The reconstructed output of a source domain autoencoder on a target domain sample; Indicates the number of source domain autoencoders. Represents the target domain The total number of sampling times in the corresponding feature dataset;

[0041] (2.3) Generate adaptive transfer weights: The weights obtained in step (2.2) One reconstruction difference value Perform Softmax normalization to obtain the first... Adaptive transfer weights from the source domain to the target domain ;

[0042] (3) Construct an adaptive multi-domain unentangled manifold network model;

[0043] like Figure 6 As shown, the adaptive multi-domain unentangled manifold network model is composed of a private encoder. Shared encoder Feature fusion module and shared decoder composition;

[0044] The private encoder and the shared encoder are both composed of two identical temporal convolutional network modules connected in series. Each temporal convolutional network module consists of a causal convolutional layer with a kernel size of 3, a dilation factor of 3, and a padding length of 6, a one-dimensional batch normalization layer, and a ReLU activation function connected in series. The private encoder is used to extract private feature information within the working condition, and the shared encoder is used to extract common feature information between working conditions.

[0045] The shared decoder consists of two identical one-dimensional transposed convolutional network modules connected in series. Each layer consists of a transposed convolutional layer with a kernel size of 3 and a padding length of 3, a one-dimensional batch normalization layer, and a ReLU activation function, which is used to realize the dimensionality restoration of temporal features and the reconstruction and extraction of the original signal.

[0046] (4) Train an adaptive multi-domain unentangled manifold network model;

[0047] (4.1) Perform feature deentanglement extraction;

[0048] Feature datasets under various working conditions The shared encoder and the corresponding private encoder are input into the adaptive multi-domain unentangled manifold network model, and the shared feature vector is extracted through the shared encoder. The private feature vector is calculated using the corresponding working condition private encoder. ;

[0049] (4.2) Feature fusion and reconstruction;

[0050] (4.2.1) Perform feature fusion operation: The feature fusion module processes the shared feature vectors. With private feature vectors The features are concatenated along the channel dimension to obtain the fused feature vector. :

[0051] ;

[0052] (4.2.2) Perform decoding and reconstruction operations; fuse the feature vectors. Input Decoder By utilizing the convolutional layer parameters of the decoder, the latent spatial features are mapped back to the original data space, resulting in reconstructed feature datasets for various operating conditions. Among them, the reconstructed first Feature data at time 1 for:

[0053] ;

[0054] (4.3) Calculate the network training loss function;

[0055] (4.3.1) Calculate the basic reconstruction loss function :

[0056] ;

[0057] in, This indicates the calculation of the square of the Frobenius norm;

[0058] (4.3.2) Calculate the characteristic orthogonal constraint loss function :

[0059] ;

[0060] (4.3.3) Calculate the loss function for private feature differences between domains. :

[0061] ;

[0062] in, Indicates the first All feature data under various working conditions Private feature vectors The mean, This indicates the calculation of cosine similarity.

[0063] (4.3.4) Calculate the adaptive weighted manifold loss function ;

[0064] Using the adaptive transfer weights obtained in step (2) Construct a weighted adjacency matrix Its elements satisfy:

[0065] ;

[0066] Among them, if the sample , They come from the source domain respectively and target domain ,and , If they come from different domains, then ,otherwise ; As a hyperparameter, in this embodiment, Set to 0.5;

[0067] Secondly, calculate the graph Laplacian matrix. ,in, It is a degree matrix;

[0068] Finally, calculate the manifold loss:

[0069] ;

[0070] in, For the matrix representation of all samples in the shared feature space, Represents the trace operation of a matrix;

[0071] (4.3.5) Calculate the overall optimization objective; sum the above four loss functions by weight to obtain the total training loss function of the network. :

[0072] ;

[0073] in, These are the preset balance hyperparameters;

[0074] (4.3.6) Based on the loss function value The parameters of the model are updated using stochastic gradient descent, and then steps (4.1) to (4.3) are repeated until the adaptive multi-domain unentangled manifold network model converges.

[0075] (5) Calculate the manifold deviation anomaly score;

[0076] (5.1) Feature space mapping;

[0077] For any given working condition Feature dataset Data at each time point Input the trained cross-domain shared encoder Extracting shared feature vectors ;

[0078] (5.2) Construct anomaly score reference feature set;

[0079] The sample feature sets from all source domain conditions are input into the trained shared encoder to extract the shared feature vectors of the data at each time point under each source domain condition. Then, the shared feature vectors extracted from the sample feature sets of all source domain conditions are used to construct the anomaly score reference feature set. ;

[0080] (5.3) Determine the local geometric reference;

[0081] In the reference feature set of abnormal scores In the search, the shared feature vector is... Euclidean distance nearest Each neighbor sample constitutes a nearest neighbor set. In this embodiment, Set to 10;

[0082] Using principal component analysis to analyze nearest neighbor sets Decompose the matrix to obtain the basis vector matrix of the normal space. ;

[0083] Wherein, the basis vector matrix of the normal space The specific acquisition process is as follows:

[0084] First, calculate the nearest neighbor set. The local arithmetic mean vector of all samples in :

[0085] ;

[0086] Secondly, the samples in the nearest neighbor set are centered, and a local covariance matrix is ​​constructed. :

[0087] ;

[0088] Next, eigenvalue decomposition is performed on the local covariance matrix: Solve its eigenvalues and corresponding orthogonal eigenvectors ;

[0089] Finally, the obtained The eigenvalues ​​are sorted in ascending order, that is: Then select the previous one. The orthogonal eigenvectors corresponding to the smallest eigenvalues And form the basis vector matrix of the normal space. :

[0090] ;

[0091] (5.4) Calculate the adaptive reference center;

[0092] The shared encoder in step (5.1) Extracting feature datasets The shared feature vector is represented as follows: ;

[0093] Traversing the shared feature vector set Each shared feature vector Find its corresponding reference mean center. :

[0094] ;

[0095] in, Nearest neighbor reference features The adaptive transfer weights corresponding to the source domain;

[0096] (5.5) Calculate the scoring indicators;

[0097] Shared feature vectors With Reference Center Project the difference vector onto the local normal space The square of its projection modulus is calculated as the manifold deviation anomaly score. :

[0098] ;

[0099] Among them, the superscript " " indicates transpose;

[0100] Finally, the shared feature vectors are... Calculated manifold deviation anomaly score Composition of abnormal score set ,in, for Abbreviated form;

[0101] (6) Solving the fault detection threshold based on kernel density estimation;

[0102] (6.1) Construct the fitted probability density;

[0103] The kernel density estimation method is used to statistically model the anomaly score set to obtain its probability density function. :

[0104] ;

[0105] in, For function variables, For bandwidth parameters, The Gaussian kernel function;

[0106] (6.2) Determine any working condition Control limits;

[0107] Set confidence level , To set the confidence level, in this embodiment, the confidence level is... The value is set to 0.01 to strictly control the false alarm rate.

[0108] Solving for any working condition by integrating the probability density function. Corresponding fault detection control limits To satisfy:

[0109] ;

[0110] (7) Online fault monitoring and status determination;

[0111] (7.1) Real-time data acquisition;

[0112] During the operation of the inverter system, voltage data samples are collected in real time at the current moment. And determine the current operating conditions;

[0113] (7.2) Calculate the real-time score;

[0114] Calculate the voltage data sample according to step (5). Corresponding manifold deviation anomaly score ;

[0115] (7.3) Fault diagnosis;

[0116] Manifold deviation anomaly score Fault detection and control limits under corresponding operating conditions Comparison: If Determine if a fault has occurred in the inverter system at the current moment; if The inverter system is determined to be operating normally at the current moment.

[0117] Experimental simulation

[0118] The platform used for verification in this invention consists of four parts: a DC power supply, an inverter circuit system, electronic loads, and a data acquisition module. The DC power supply voltage was set to simulate three different operating conditions: 50V, 60V, and 80V. The inverter circuit system uses Semikron SKM100GAL12T4 modules with a 2SC0435T2F1-17 IGBT driver core and a 2BB043 adapter board. The three electronic loads are set to AC constant resistance mode with a resistance value of 50 ohms. The implanted fault is located in the filter inductor at the output of the inverter circuit system, with a resistor connected in parallel to simulate inductor degradation. In this experiment, 4000 normal data points were collected for each source domain condition to train the model, while only 10 normal data points were collected for the target domain condition. The test data for each operating condition consisted of 4000 samples, of which 0-2000 samples were normal data, and the parallel resistance of samples 2000-4000 was 330Ω. The fault detection rate was calculated based on the number of samples exceeding the control limit during the fault duration. Then, the false alarm rate was obtained using the number of falsely identified normal samples outside the fault duration.

[0119] In this experiment, Temporal Convolutional Network (TCN), Transformer, Long Short-Term Memory-Convolutional Neural Network (LSTM-CNN), Self-tuning Transfer Dynamic Autoencoder (STDAE), One-class Transfer Autoencoder (OCTAE), and the AMDMN proposed in this invention were used to verify their performance in the inverter system. TCN has a hidden layer dimension of 16. The Transformer model is configured with 6 attention heads and 16 hidden layers. LSTM-CNN has 16 LSTM hidden layers and 8 convolutional latent space dimensions. OCTAE uses a hidden layer dimension of 16. STDAE has both shared and private latent space dimensions set to 16, and uses maximum mean difference loss for domain alignment, with loss weight coefficients... Set to 0.01. This invention's AMDMN also employs a 1D-CNN encoder architecture, with its core manifold deviation exponential configuration having 10 nearest neighbors and an intrinsic manifold dimension of 3. All models use the Adam optimizer, with an initial learning rate set to... The weight decay coefficient is The model training run was uniformly set to 30 rounds. During the anomaly detection phase, the significance level of each detector was set to 1%, and its statistical control limits were automatically calculated by the kernel density estimation algorithm based on the score distribution of the normal training data.

[0120] Table 2 lists the performance metrics of various verification methods in multi-condition fault detection of inverter systems. Among them, standard deep learning methods such as TCN, Transformer, and LSTM-CNN show unsatisfactory detection results and significant instability. Taking TCN as an example, although it can achieve a high fault detection rate (FDR) under certain conditions, its false alarm rate (FAR) is as high as 45.20% under the 50V condition, failing to achieve reliable fault detection. Due to the introduction of transfer learning mechanisms, the detection performance of OCTAE and STDAE is significantly improved compared to non-transfer methods, effectively suppressing the false alarm rate to a low level (e.g., OCTAE is only 0.75% under the first condition), while its detection rate exceeds 60% in most tasks, even reaching a peak detection performance of 99.90% under specific conditions. The AMDMN proposed in this invention achieves a high fault detection rate of 97.65% under the 50V / 400Hz condition. Especially in the target domain with only 10 training samples, AMDMN still maintains a detection rate of 94.70%, which proves the superiority and effectiveness of the proposed method in dealing with multi-condition data drift and sample scarcity problems.

[0121] Table 2 shows the failure detection rate (%) of the verification method for intermittent inductor degradation faults in inverter systems;

[0122] method 50V / 400Hz FAR (%) 50V / 400Hz FDR (%) 60V / 400Hz FAR(%) 60V / 400Hz FDR (%) 80V / 400Hz FAR (%) 80V / 400Hz FDR (%) TCN 45.20 99.95 53.50 74.50 17.55 82.55 Transformer 0.05 33.80 1.60 30.00 7.25 13.75 LSTM-CNN 0.80 50.30 17.10 78.65 1.00 46.50 OCTAE 0.75 81.10 16.55 64.75 1.25 65.30 STDAE 18.55 50.80 10.15 99.90 16.55 57.25 AMDMN 0.95 97.65 0.00 92.70 4.65 94.70

[0123] Although the illustrative specific embodiments of the present invention have been described above to enable those skilled in the art to understand the invention, it should be understood that the invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the invention as defined and determined by the appended claims, and all inventions utilizing the concept of the present invention are protected.

Claims

1. A multi-condition fault detection method for inverter systems based on adaptive multi-domain unentangled manifold networks, characterized in that, Includes the following steps: (1) Data acquisition and processing; (1.1) Collect three-phase voltage data of the inverter system at the load end under different operating conditions at equal intervals. The number of sampling points is: Point; record the first The three-phase voltage data sequence collected under various operating conditions is represented as follows: , among which, the The first working condition Three-phase voltage data collected at each moment Represented as , Indicates the first The first working condition The data collected at the [time]th moment One sampling point, Represents a three-phase circuit. Indicates the first Total number of sampling times under various working conditions; (1.2) Traverse the three-phase voltage data sequence under each operating condition The three-phase voltage data at each sampling time are extracted sequentially. The features include: frequency, mean, peak-to-peak value, RMS value, margin, and kurtosis. These six features are then used as the first set of features. Feature vector at each sampling time This allows us to obtain feature datasets for various operating conditions. ; (1.3) Traverse the three-phase voltage data sequence under each operating condition Based on the total number of sampling times corresponding to each working condition With preset threshold Perform domain partitioning, if Then the first The operating condition is defined as the source domain operating condition, denoted as . Otherwise, the first The type of working condition is defined as the target domain working condition, denoted as ; (2) Adaptive weighted calculation; (2.1) Training source domain-specific autoencoders: For each source domain Using the corresponding feature dataset within this domain Training an independent source domain autoencoder ,in, Representing the Training parameters for a source domain autoencoder; (2.2) Calculate the reconstruction differences of the target domain under each source domain autoencoder: The feature dataset corresponding to the target domain... The data are input into each of the source domain autoencoders trained in step (2.1), and the target domain sample is calculated according to the following formula at the i-th position. Average reconstruction difference under each source domain model : ； in, Represents the first of the target domains Each sample, i.e., the target domain The corresponding feature dataset in the first Feature vectors at each sampling time; For temperature coefficient, Represents the L2 norm. Indicates the first The reconstructed output of a source domain autoencoder on a target domain sample; Indicates the number of source domain autoencoders. Represents the target domain The total number of sampling times in the corresponding feature dataset; (2.3) Generate adaptive transfer weights: The weights obtained in step (2.2) One reconstruction difference value Perform Softmax normalization to obtain the first... Adaptive transfer weights from the source domain to the target domain ; (3) Construct an adaptive multi-domain unentangled manifold network model; The adaptive multi-domain unentangled manifold network model is based on a proprietary encoder. Shared encoder Feature fusion module and shared decoder composition; The private encoder and the shared encoder are both composed of two identical temporal convolutional network modules connected in series. Each temporal convolutional network module consists of a causal convolutional layer with a kernel size of 3, a dilation factor of 3, and a padding length of 6, a one-dimensional batch normalization layer, and a ReLU activation function connected in series. The private encoder is used to extract private feature information within the working condition, and the shared encoder is used to extract common feature information between working conditions. The shared decoder consists of two identical one-dimensional transposed convolutional network modules connected in series. Each layer consists of a transposed convolutional layer with a kernel size of 3 and a padding length of 3, a one-dimensional batch normalization layer, and a ReLU activation function, which is used to realize the dimensionality restoration of temporal features and the reconstruction and extraction of the original signal. (4) Train an adaptive multi-domain unentangled manifold network model; (4.1) Perform feature deentanglement extraction; Feature datasets under various working conditions The shared encoder and the corresponding private encoder are input into the adaptive multi-domain unentangled manifold network model, and the shared feature vector is extracted through the shared encoder. The private feature vector is calculated using the corresponding working condition private encoder. ; (4.2) Feature fusion and reconstruction; (4.2.1) Perform feature fusion operation: The feature fusion module processes the shared feature vectors. With private feature vectors The features are concatenated along the channel dimension to obtain the fused feature vector. : ； (4.2.2) Perform decoding and reconstruction operations; fuse the feature vectors. Input Decoder By utilizing the convolutional layer parameters of the decoder, the latent spatial features are mapped back to the original data space, resulting in reconstructed feature datasets for various operating conditions. Among them, the reconstructed first Feature data at time 1 for: ； (4.3) Calculate the network training loss function; (4.3.1) Calculate the basic reconstruction loss function : ； in, This indicates the calculation of the square of the Frobenius norm; (4.3.2) Calculate the characteristic orthogonal constraint loss function : ； (4.3.3) Calculate the loss function for private feature differences between domains. : ； in, Indicates the first All feature data under various working conditions Private feature vectors The mean, This indicates the calculation of cosine similarity. (4.3.4) Calculate the adaptive weighted manifold loss function ; Using the adaptive transfer weights obtained in step (2) Construct a weighted adjacency matrix Its elements satisfy: ； Among them, if the sample , They come from the source domain respectively and target domain ,and , If they come from different domains, then ,otherwise ; For hyperparameters; Secondly, calculate the graph Laplacian matrix. ,in, It is a degree matrix; Finally, calculate the manifold loss: ； in, For the matrix representation of all samples in the shared feature space, Represents the trace operation of a matrix; (4.3.5) Calculate the overall optimization objective; sum the above four loss functions by weight to obtain the total training loss function of the network. : ； in, These are the preset balance hyperparameters; (4.3.6) Based on the loss function value The parameters of the model are updated using stochastic gradient descent, and then steps (4.1) to (4.3) are repeated until the adaptive multi-domain unentangled manifold network model converges. (5) Calculate the manifold deviation anomaly score; (5.1) Feature space mapping; For any given working condition Feature dataset Data at each time point Input the trained cross-domain shared encoder Extracting shared feature vectors ; (5.2) Construct anomaly score reference feature set; The sample feature sets from all source domain conditions are input into the trained shared encoder to extract the shared feature vectors of the data at each time point under each source domain condition. Then, the shared feature vectors extracted from the sample feature sets of all source domain conditions are used to construct the anomaly score reference feature set. ; (5.3) Determine the local geometric reference; In the reference feature set of abnormal scores In the search, the shared feature vector is... Euclidean distance nearest Each neighbor sample constitutes a nearest neighbor set. ; Using principal component analysis to analyze nearest neighbor sets Decompose the matrix to obtain the basis vector matrix of the normal space. ; (5.4) Calculate the adaptive reference center; The shared encoder in step (5.1) Extracting feature datasets The shared feature vector is represented as follows: ; Traversing the shared feature vector set Each shared feature vector Find its corresponding reference mean center. : ； in, Nearest neighbor reference features Adaptive transfer weights corresponding to the source domain; (5.5) Calculate the scoring indicators; Shared feature vectors With Reference Center Project the difference vector onto the local normal space The square of its projection modulus is calculated as the manifold deviation anomaly score. : ； Among them, the superscript " " indicates transpose; Finally, the shared feature vectors are... Calculated manifold deviation anomaly score Composition of abnormal score set ,in, for Abbreviated form; (6) Solving the fault detection threshold based on kernel density estimation; (6.1) Construct the fitted probability density; The kernel density estimation method is used to statistically model the anomaly score set to obtain its probability density function. : ； in, For function variables, For bandwidth parameters, The Gaussian kernel function; (6.2) Determine any working condition Control limits; Set confidence level , The set confidence level; Solving for any working condition by integrating the probability density function. Corresponding fault detection control limits To satisfy: ； (7) Online fault monitoring and status determination; (7.1) Real-time data acquisition; During the operation of the inverter system, voltage data samples are collected in real time at the current moment. And determine the current operating conditions; (7.2) Calculate the real-time score; Calculate the voltage data sample according to step (5). Corresponding manifold deviation anomaly score ; (7.3) Fault diagnosis; Manifold deviation anomaly score Fault detection and control limits under corresponding operating conditions Comparison: If Determine if a fault has occurred in the inverter system at the current moment; if The inverter system is determined to be operating normally at the current moment.

2. The method for multi-condition fault detection of inverter systems based on adaptive multi-domain unentangled manifold networks according to claim 1, characterized in that, The normal space basis vector matrix The specific acquisition process is as follows: First, calculate the nearest neighbor set. The local arithmetic mean vector of all samples in : ； Secondly, the samples in the nearest neighbor set are centered, and a local covariance matrix is ​​constructed. : ； Next, eigenvalue decomposition is performed on the local covariance matrix: Solve its eigenvalues and corresponding orthogonal eigenvectors ; Finally, the obtained The eigenvalues ​​are sorted in ascending order, that is: Then select the previous one. The orthogonal eigenvectors corresponding to the smallest eigenvalues And form the basis vector matrix of the normal space. : 。

Images