A method for spatiotemporal alignment and fusion based on multi-dimensional state feature data of a transformer

Abstract

The application discloses a kind of space-time alignment fusion methods based on transformer multidimensional state feature data, it is characterized in that, including: using LSTM network layer and improved one-dimensional full convolutional neural network 1D-FCN respectively extract the time feature and spatial feature of multidimensional data, using two networks simultaneously process transformer multidimensional state data, can make the single time feature obtained by LSTM and the spatial feature obtained by 1D-FCN more detailed, then utilize self-learning weight information to correct the deviation between multi-scale features, utilize the fusion of transformer multidimensional state features realized by optimized fully connected layer, finally realize feature classification using classification algorithm, reach the effect that the same category feature distribution is strong, different category feature distribution has obvious spatial interval;Solve the current transformer faces model construction and evolution mode practicability is weak, for transformer multi-source, multi-modal feature quantity expression is not consistent, leading to multidimensional data utilization is not high, data management effect is poor and other problems.

Description

Technical Field

[0001] This invention relates to the field of spatiotemporal fusion technology of multi-scale feature data, and in particular to a spatiotemporal fusion method based on multidimensional state feature data of transformers. Background Technology

[0002] With the continuous development of power systems, transformers, as core electrical equipment, are constantly undergoing upgrades and becoming increasingly efficient and intelligent. Accurately depicting the operating status of transformers under actual conditions is crucial for fault early warning and diagnosis. However, current transformer models face a series of problems, including weak practicality in model construction and evolution methods, inconsistent representations of multi-source and multi-modal characteristics, low utilization of multi-dimensional data, poor data governance, and relatively crude traditional early warning strategies, with maintenance decisions largely relying on manual processes. Therefore, researching spatiotemporal alignment and fusion technology based on multi-dimensional transformer state characteristic data to achieve effective governance of multi-dimensional state characteristic data and predictive analysis and early warning of operating status has become a key focus.

[0003] The current status and problems of spatiotemporal alignment and fusion technology for multidimensional state feature data at home and abroad are summarized as follows:

[0004] 1) A spatiotemporal data fusion method based on Kalman filtering uses the real-time Kalman filtering algorithm to fuse active and passive data based on time series data. Building upon this time series fusion, it further fuses active and passive data according to weights based on spatial distribution characteristics. However, the amount of interactive information between the fused data using this method fails to meet standard requirements, resulting in poor fusion performance.

[0005] 2) A data alignment method based on a multi-algorithm combination and a multi-level attention mechanism focusing on information content addresses the problem of insufficient utilization of multimodal data features in social networks. It achieves user data alignment by fusing different types of data within the same modality and features from different modalities. However, this method employs a comparison approach where other algorithms remain unchanged while a single algorithm is replaced. Theoretically, there exist other specific combinations of algorithms that are superior to the current choice.

[0006] 3) Data spatiotemporal alignment and fusion method based on dilated convolution corrects the position of features corresponding to different scales. When using dilated convolution, it is affected by the size of the convolution kernel, and due to the characteristics of dilated convolution, some features will be lost. The feature position of the same target will be shifted by different proportions.

[0007] 4) The spatiotemporal alignment fusion method based on convolution kernel pyramids directly merges different feature layers using different convolution kernels. Essentially, it is a general convolution method with the same convolution kernel and different amplitudes, but it only fuses local features at different scales and is cumbersome to operate. Therefore, the spatial information obtained by deep features through shallow features is biased, and the redundant information of deep features will also interfere with shallow features.

[0008] The main problems with the spatiotemporal alignment and fusion technology of multidimensional state feature data are poor fusion effect, insufficient information content after fusion, fusion achieved by replacing only a single algorithm, insufficient utilization of multimodal feature data, loss of features and spatial information deviation.

[0009] Therefore, it is necessary to provide a spatiotemporal alignment and fusion method based on transformer multidimensional state feature data to solve the above-mentioned technical problems. Summary of the Invention

[0010] The purpose of this section is to outline some aspects of the embodiments of the present invention and to briefly introduce some preferred embodiments. Some simplifications or omissions may be made in this section, as well as in the abstract and title of the present application, to avoid obscuring the purpose of this section, the abstract and title of the invention. Such simplifications or omissions shall not be used to limit the scope of the present invention.

[0011] Therefore, the present invention aims to address the problems currently faced by transformers, such as the weak practicality of model construction and evolution methods, inconsistent descriptions of multi-source and multi-modal characteristic quantities of transformers, resulting in low utilization of multi-dimensional data and poor data governance effects.

[0012] To address the aforementioned technical problems, this invention provides the following technical solution: a spatiotemporal alignment and fusion method based on transformer multidimensional state feature data, characterized by comprising: The spatiotemporal alignment and fusion technology for transformer multidimensional state feature data employs an LSTM network layer and an improved one-dimensional fully convolutional neural network (1D-FCN) to extract the temporal and spatial features of the multidimensional data respectively. By using two networks to process the transformer multidimensional state data simultaneously, the single temporal feature acquired by LSTM and the spatial feature acquired by 1D-FCN become more detailed. Then, self-learning weight information is used to correct the deviation between multi-scale features. An optimized fully connected layer is used to achieve the fusion of transformer multidimensional state features. Finally, a classification algorithm is used to achieve feature classification, resulting in strong clustering of features of the same category and different feature distributions. The system exhibits a clear spatial separation effect. By aligning and fusing multi-dimensional state features of transformers across both temporal and spatial scales, it effectively improves the completeness, accuracy, and timeliness of transformer full-element perception data, laying the foundation for subsequent transformer modeling, fault diagnosis, risk warning, and operational decision-making. Two LSTM network layers are used to extract features from the transformer's time-series data, uncovering hidden temporal patterns. These temporal features effectively assist in extracting spatial features, making the management of multi-dimensional transformer state data more efficient. Transformer vibration signals and other data are transformed into feature maps using data reconstruction methods and then input into an improved one-dimensional fully convolutional neural network (1D-FCN). The fully convolutional neural network employs three one-dimensional convolutional layers to fully extract the spatial features of the transformer's multi-dimensional data. Using two networks simultaneously to process the transformer's multi-dimensional state data allows for more detailed single-temporal features acquired by LSTM and spatial features acquired by 1D-FCN. The unique network structure ensures the integrity of the input data.

[0013] As a preferred embodiment of the spatiotemporal alignment and fusion method based on transformer multidimensional state feature data described in this invention, the LSTM network layer used includes: using two LSTM network layers to extract time features from the transformer multidimensional state data, which can uncover hidden temporal patterns in the data, thereby assisting in the governance of transformer multidimensional state data using spatial features.

[0014] As a preferred embodiment of the spatiotemporal alignment and fusion method based on multidimensional state feature data of transformers described in this invention, the improved one-dimensional fully convolutional neural network 1D-FCN uses three one-dimensional convolutional layers to extract spatial features of multidimensional transformer data. By setting different kernel sizes for the three convolutional layers, the integrity and diversity of the extracted features are ensured.

[0015] As a preferred embodiment of the spatiotemporal alignment and fusion method based on transformer multidimensional state feature data described in this invention, the method of correcting the deviation between multi-scale features by using self-learning weight information includes: using dynamic weights to select corresponding feature points in the local field to reduce feature offset and mutual interference between features, which can adaptively correspond to features of different scales, and adjusting features of different scales to the same scale through an interpolation module for splicing operation.

[0016] As a preferred embodiment of the spatiotemporal alignment fusion method based on transformer multidimensional state feature data described in this invention, the fusion operation of transformer multidimensional state feature data is achieved using an optimized fully connected layer. The improved fully connected layer of this invention consists of two fully connected layers, one Flatten layer, and one Dropout layer, with ReLU activation functions to ensure iteration speed and training effectiveness. The added Dropout layer is to address the overfitting problem and reduce the impact of overfitting on the feature fusion result. The fully connected layer fusion algorithm only concatenates feature information without changing other content, thus preserving the integrity of the feature information.

[0017] As a preferred embodiment of the spatiotemporal alignment and fusion method based on transformer multidimensional state feature data described in this invention, the feature classification using a classification algorithm includes: by leveraging the angle between the feature vector and the weight vector of each class of features, and through a loss function and an improved multi-class classification algorithm, the spatiotemporal alignment and fusion of transformer multidimensional state feature data achieves the expected effect of strong cohesion of the same class of feature distributions and significant spacing between different class of feature distributions.

[0018] As a preferred embodiment of the spatiotemporal alignment and fusion method based on transformer multidimensional state feature data described in this invention, the interpolation module includes: the module for realizing the alignment of transformer multidimensional state features is composed of a series of interpolation modules (IFM), each IFM consisting of an amplified feature alignment module (MFA) and a scaled feature alignment module (DFA), which adjusts the correspondence between feature points by learning a set of weights, and can adaptively align features at different scales.

[0019] As a preferred embodiment of the spatiotemporal alignment fusion method based on transformer multidimensional state feature data described in this invention, the fusion algorithm includes: obtaining temporal and spatial features, aligning the features, obtaining a feature map with the same size as the input information through upsampling, using a one-dimensional pooling layer to reduce the feature dimensionality to obtain a new feature map, and obtaining the final fused features through a feature concatenation operation; based on the improved softmax multi-class classification algorithm, the spatiotemporal alignment fusion of transformer multidimensional state feature data achieves the expected effect of strong cohesion of the same type of feature distribution and obvious spacing between different types of feature distributions.

[0020] The beneficial effects of this invention are: it can provide digital decision support for improving the state perception and control capabilities of transformers, and facilitate the development of solutions such as multimodal data governance, data-driven model construction, proactive risk warning, and predictive operation and maintenance decision-making based on the power grid resource business platform, thereby supporting the construction of new power systems and ensuring the safe and stable operation of the power grid.

[0021] (1) The improved 1D-FCN network can process sensor data and leverage the advantages of convolutional neural networks in extracting spatial features, while LSTM is better at processing temporal features. Combining the strengths of both, the accuracy and effectiveness of the two types of features obtained are higher.

[0022] (2) The 1D-FCN spatial model proposed in this invention uses deconvolutional layers to upsample the features of the last convolutional layer in order to restore its size to be the same as the size of the input data, thus preserving the spatial information of the original data. In addition, the 1D-FCN model also adopts a skip structure, combining the earlier convolutional layers and the upsampled convolutional layers, which makes the data partitioning more refined while preserving global features. The 1D-FCN network can extract fault feature details while maintaining network stability and accelerating training.

[0023] (3) Use dynamic weights to select corresponding feature points. If multi-scale feature fusion is performed directly, it will cause target feature shift and interference of target features to surrounding features, and generate redundant information that will disturb the accuracy of the network model. By first aligning the transformer multi-dimensional features through self-learning weight information, and then using the information filtering module to filter redundant information, these problems can be reduced, thereby ensuring the stability of the system.

[0024] (4) The fully connected layer fusion algorithm only splices feature information without changing other content, thus ensuring the integrity of feature information. The improved multi-classification algorithm has stronger cohesion in the feature distribution of the same class and more obvious intervals in the feature distribution of different classes. Attached Figure Description

[0025] To more clearly illustrate the technical solutions in the embodiments of the present 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 only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0026] Figure 1 A model diagram of the spatiotemporal alignment and fusion method based on multidimensional state feature data of transformers, as provided in an embodiment of the present invention;

[0027] Figure 2A flowchart of 1D-FCN extraction of spatial features for a spatiotemporal alignment and fusion method based on multidimensional state feature data of transformers, provided as an embodiment of the present invention. Detailed Implementation

[0028] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0029] The terms "first" and "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish different objects, not to describe a specific order. 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 apparatus that includes a series of steps or units may include steps or units not listed, but rather steps or units not listed.

[0030] Example 1

[0031] The spatiotemporal alignment and fusion model structure based on transformer multidimensional state feature data is as follows: Figure 1As shown, this invention designs two LSTM layers to extract the temporal features from the multidimensional state data of transformers, uncovering hidden temporal patterns. It leverages an improved 1D-FCN to fully utilize the spatial feature extraction advantages of convolutional neural networks. The features processed by the 1D-FCN network and LSTM layers are input into an alignment and fusion algorithm, then fed into a fully connected layer to update and optimize the output. A Dropout layer addresses overfitting, and finally, a multi-classification algorithm is created for training and classification features to obtain the feature classification results. The invention employs an LSTM network to mine the temporal pattern features of transformers, obtaining temporal context data information and rationally extracting the temporal features of multidimensional data. An improved one-dimensional fully convolutional neural network (1D-FCN) is used to extract the spatial features of the multidimensional transformer data, and self-learning weight information is used to further correct the deviations between multi-scale features, achieving feature alignment on both temporal and spatial scales. An improved fully connected layer algorithm fuses feature information, and finally, a multi-classification algorithm is used to cluster similar features and disperse dissimilar features. Two LSTM network layers are used to extract features from the time-series data of transformers, uncovering hidden temporal patterns. These temporal features effectively assist spatial features in transformer anomaly detection, making the management of multi-dimensional transformer state data more efficient. Transformer vibration signals and other data are transformed into feature maps using a data reconstruction method and then input into an improved one-dimensional fully convolutional neural network (1D-FCN). The fully convolutional neural network uses three one-dimensional convolutional layers to fully extract the spatial features of the multi-dimensional transformer data. Using two networks simultaneously to process the multi-dimensional transformer state data allows for more detailed single-time features acquired by LSTM and spatial features acquired by 1D-FCN. The special network structure ensures the integrity of the input data. Self-learning weight information is used to correct biases between multi-scale features by learning a set of weights to adjust... The system establishes the correspondence between feature points and can adaptively correspond to features at different scales. An interpolation module adjusts features at different scales to the same scale for splicing. An optimized fully connected layer is used to fuse multi-dimensional state feature data of the transformer. This improved fully connected layer consists of two fully connected layers, one Flatten layer, and one Dropout layer, all using ReLU activation functions to ensure iteration speed and training effectiveness, reducing the impact of overfitting on the feature fusion results. The fully connected layer fusion algorithm only splices feature information without changing other content, preserving the integrity of the feature information. An improved multi-classification algorithm is employed to achieve the expected effect of strong cohesion among feature distributions of the same class and significant spacing between feature distributions of different classes after spatiotemporal alignment and fusion of the transformer multi-dimensional state feature data.

[0032] Example 2

[0033] Corresponding to the spatiotemporal alignment and fusion model of transformer multidimensional state feature data disclosed in Embodiment 1 of this invention, Embodiment 2 of this invention also provides a specific implementation method for extracting spatial features using 1D-FCN, see [link to implementation]. Figure 2 The method includes:

[0034] FCN (Flexible Convolutional Network) transforms fully connected layers into convolutional layers on top of the CNN (Convolutional Network) architecture. This structure can adapt to any input, which is beneficial for enhancing the input performance of the transformer text dataset. The spatial feature extraction process of the proposed 1D-FCN is shown in the figure. A deconvolutional layer is used to upsample the features of the last convolutional layer to keep its size consistent with the size of the input data, thereby preserving the spatial information of the original data. The original data signal can be transformed into a feature map form through data reconstruction. Then, the feature map of the signal is input into the 1D-FCN for processing. To ensure the completeness and diversity of the extracted features, the 1D-FCN uses three one-dimensional convolutional layers to extract spatial information features, and uses a max pooling layer for parsing. The first convolutional layer uses 128 convolutional kernels, the second convolutional layer uses 64 convolutional kernels, and the third convolutional layer uses 32 convolutional kernels. The three convolutional kernels are the same size. By changing the size of the convolutional kernels, the important parts in the transformer data are fully extracted. All convolutional layers use the ReLU function as the activation function to avoid the gradient vanishing problem and ensure the iteration speed and training effect of the experiment.

[0035] Example 3

[0036] Corresponding to the spatiotemporal alignment and fusion model of transformer multidimensional state feature data disclosed in Embodiment 1 of the present invention, Embodiment 3 of the present invention also provides a specific implementation method for aligning transformer multidimensional state features, the method comprising:

[0037] If multi-dimensional target features are directly fused, the resulting fused features will not cluster together, making it difficult to determine the location of the target features. This invention utilizes self-learned weight information to correct the deviation between multi-scale features. The essence of multi-dimensional feature alignment is feature point matching. Multi-dimensional features are the result of convolution operations, possessing local similarity, and the offset of feature points is within the range of convolution operations. Therefore, this invention uses dynamic weights to select corresponding feature points in the local context to reduce feature offset and mutual interference between features. The alignment module consists of a series of interpolation modules (IFM). Each IFM consists of a magnified feature alignment module (MFA) and a scaled feature alignment module (DFA). By learning a set of weights, it adjusts the correspondence between feature points and can adaptively align features of different scales. Finally, the IFM adjusts features of different scales to the same scale for splicing.

[0038] Example 4

[0039] Corresponding to the spatiotemporal alignment fusion model of transformer multidimensional state feature data disclosed in Embodiment 1 of the present invention, Embodiment 4 of the present invention also provides a specific implementation method for fusing transformer multidimensional state features, the method comprising:

[0040] The improved fully connected layer in this invention consists of two fully connected layers. The second fully connected layer is a Dropout layer, designed to address overfitting and reduce its impact on subsequent feature classification. The activation function is ReLU. The fusion algorithm is shown below:

[0041] Input: Time series data, sensor signals associated with the transformer.

[0042] Initialization parameters: number of training epochs, learning rate, number of neural network nodes, batch size

[0043] 1. Input time series data into LSTM, input sensor signals into 1D-FCN, extract raw data features, and perform normalization.

[0044] 2. Maintain the temporal order of the entire sequence, use 1D-FCN to extract local spatial features, obtain local information in the spatial domain, and obtain a feature map of the same size as the input information through upsampling.

[0045] 3. Use a one-dimensional pooling layer to perform feature dimensionality reduction to obtain a new feature map.

[0046] 4. By using multiple LSTMs to learn the characteristics of data sequences changing over time, the learned features can assist spatial features in enhancing the ability to identify transformer anomalies.

[0047] 5. The final merged feature is obtained through the feature concatenation operation.

[0048] Output: Fused feature data

[0049] Example 5

[0050] Corresponding to the spatiotemporal alignment fusion model of transformer multidimensional state feature data disclosed in Embodiment 1 of this invention, Embodiment 5 of this invention also provides a specific implementation method for feature classification, which includes:

[0051] To achieve stronger cohesion among feature distributions of the same class and more significant spacing between feature distributions of different classes, the improved feature classification algorithm is as follows:

[0052] W c x = ||W c ||2||x||2cos(θ) c )

[0053] Where c is the feature category, Wc This represents the weight parameters for feature class c in the last fully connected layer. This formula shows that the cosine distance is related to the magnitude of the vector itself, the magnitude of the feature, and the angle between them. For a feature to be classified, the specific class it is assigned to depends on the angle between the feature vector and the weight vector of each class. Using the loss function...

[0054]

[0055] This ensures good angular separability between vectors, facilitating feature classification. The y-th weight of the W weight matrix represents... i column, x i y i These are features and feature labels. express and x i The included angle, j is the number of feature categories, W j It represents the weight of the j-th feature, and the output value determines the category of the feature.

[0056] In summary, this invention realizes a spatiotemporal alignment and fusion technology based on multidimensional transformer state feature data. It utilizes a Long Short-Term Memory (LSTM) network and an improved one-dimensional fully convolutional neural network (1D-FCN) to extract the temporal and spatial features of the transformer's multidimensional data, respectively. Then, self-learning weight information is used to correct the bias between multiple scales, achieving spatiotemporal alignment of the transformer's multidimensional features. An innovative fully connected layer algorithm is then used to fuse the spatiotemporal features. By achieving the alignment and fusion of multidimensional transformer state features at both temporal and spatial scales, the completeness, accuracy, and timeliness of the transformer's full-element perception data can be effectively improved, laying the foundation for subsequent transformer modeling, fault diagnosis, risk warning, and operational decision-making.

[0057] It is important to note that the constructions and arrangements of this application shown in several different exemplary embodiments are merely illustrative. Although only a few embodiments are described in detail in this disclosure, those who consult this disclosure will readily understand that many modifications are possible (e.g., changes in the size, dimensions, structure, shape, and proportions of various elements, as well as parameter values ​​(e.g., temperature, pressure, etc.), mounting arrangements, use of materials, color, orientation, etc.) without substantially departing from the novel teachings and advantages of the subject matter described in this application). For example, an element shown as integrally formed may be composed of multiple parts or elements, the position of elements may be inverted or otherwise altered, and the nature or number or position of discrete elements may be changed or altered. Therefore, all such modifications are intended to be included within the scope of the invention. The order or sequence of any process or method steps may be changed or rearranged according to alternative embodiments. In the claims, any "device plus function" clause is intended to cover the structure described herein that performs the function, and not only structurally equivalent but also equivalent in structure. Other substitutions, modifications, alterations, and omissions may be made in the design, operation, and arrangement of the exemplary embodiments without departing from the scope of the invention. Therefore, the present invention is not limited to the specific embodiments, but extends to various modifications that still fall within the scope of the appended claims.

[0058] Furthermore, in order to provide a concise description of exemplary embodiments, not all features of actual embodiments (i.e., those features that are not relevant to the best mode of carrying out the invention as currently considered, or those features that are not relevant to implementing the invention) may be omitted.

[0059] It should be understood that numerous specific implementation decisions can be made during the development of any practical implementation, such as in any engineering or design project. Such development efforts may be complex and time-consuming, but for those skilled in the art who benefit from this disclosure, the development effort will be a routine work of design, manufacturing, and production without requiring much experimentation.

[0060] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A spatiotemporal alignment and fusion method based on multidimensional state feature data of transformers, characterized in that, include: The spatiotemporal alignment and fusion technology for transformer multidimensional state feature data employs an LSTM network layer and an improved one-dimensional fully convolutional neural network (1D-FCN) to extract the temporal and spatial features of the multidimensional data, respectively. Using two networks simultaneously to process the transformer multidimensional state data allows for more detailed single temporal features acquired by LSTM and spatial features acquired by 1D-FCN. Then, self-learning weight information is used to correct the bias between multi-scale features. An optimized fully connected layer is used to fuse the transformer multidimensional state features. Finally, a classification algorithm is used to classify the features, achieving strong clustering of features of the same category and significant spatial spacing of features of different categories. By aligning and fusing the transformer multidimensional state features at both temporal and spatial scales, the completeness, accuracy, and timeliness of the transformer's full-element perception data can be effectively improved, laying the foundation for subsequent transformer modeling, fault diagnosis, risk warning, and operational decision-making. Correcting the bias between multi-scale features using self-learning weight information includes: using dynamic weights to select corresponding feature points in the local scene to reduce feature offset and mutual interference between features; adaptively corresponding to features of different scales; and adjusting features of different scales to the same scale through the interpolation module for stitching operation. The method of using the optimized fully connected layer to achieve the fusion of multi-dimensional state features of the transformer includes: the improved fully connected layer consists of two fully connected layers, one Flatten layer and one Dropout layer, which ensures the iteration speed and training effect, and reduces the impact of overfitting on the feature fusion result. Then, the fusion algorithm is used to achieve the fusion of multi-dimensional features of the transformer. The fusion algorithm includes: obtaining temporal and spatial features, aligning the features, upsampling to obtain a feature map with the same size as the input information, using a one-dimensional pooling layer to reduce the feature dimensionality to obtain a new feature map, and finally obtaining the fused features through a feature concatenation operation.

2. The spatiotemporal alignment and fusion method based on transformer multidimensional state feature data according to claim 1, characterized in that, The LSTM network layers used include: using two LSTM network layers to extract time features from the multidimensional state data of transformers, which can uncover hidden temporal patterns in the data, thereby assisting in the governance of multidimensional state data of transformers using spatial features.

3. The spatiotemporal alignment and fusion method based on transformer multidimensional state feature data according to claim 1, characterized in that, The improved one-dimensional fully convolutional neural network 1D-FCN uses three one-dimensional convolutional layers to extract spatial features of transformer multidimensional data. By setting different kernel sizes for the three convolutional layers, the integrity and diversity of the extracted features are ensured.

4. The spatiotemporal alignment and fusion method based on transformer multidimensional state feature data according to claim 1, characterized in that, The feature classification using the classification algorithm includes: by using the angle between the feature vector and the weight vector of each class of features, and through the loss function and the improved multi-class classification algorithm, the multi-dimensional state feature data of the transformer is spatiotemporally aligned and fused to achieve the expected effect of strong cohesion of the same class of feature distributions and obvious intervals between different class of feature distributions.

5. The spatiotemporal alignment and fusion method based on transformer multidimensional state feature data according to claim 1, characterized in that, The interpolation module includes: the module for realizing the alignment of multi-dimensional state features of transformer is composed of a series of interpolation modules (IFM), each IFM consisting of an amplified feature alignment module (MFA) and a scaled feature alignment module (DFA), which adjusts the correspondence between feature points by learning a set of weights and can adaptively align features at different scales.

Images