Methods, devices, computer equipment and media for judging the performance degradation of magnetic levitation systems

Abstract

This application relates to a method, apparatus, computer equipment, and medium for judging the performance degradation of magnetic levitation systems. The method includes: firstly, using the Symmetric Polar Coordinate (SDP) method, converting one-dimensional time-series data (3 sets of suspension gaps, 2 sets of vertical accelerations, and 1 set of electromagnet currents) into two-dimensional images, achieving interpretable fusion of multi-source parameters, reducing the false positive rate of single-data noise, and improving feature representation. Then, using a pre-trained Swin Transformer to extract multi-scale features from the images, mapping them, and then using a dual encoder to mine discriminative concept embeddings. Features are fused using a concept weight matrix, and a model is trained by combining fine-grained classification loss and dual-objective constraint loss. Real-time running data is converted into images using the same method, and the trained model is used for discrimination, solving the problems of limited receptive field and insensitivity to weak features in traditional models, significantly improving the accuracy of degradation recognition.

Description

Technical Field

[0001] This application relates to the field of magnetic levitation system degradation discrimination technology, and in particular to a method, device, computer equipment and medium for judging the performance degradation of magnetic levitation systems. Background Technology

[0002] With the rapid development of modern rail transit, maglev trains have gradually become a research hotspot. The levitation system, as the core system of a maglev train, provides electromagnetic force to achieve contactless levitation between the car body and the track. This overcomes the mechanical friction limitations of traditional wheel-rail systems, demonstrating significant advantages in high-speed transportation. The safety and reliability of the levitation system directly affect the train's operating efficiency. However, during long-term operation, the performance of the levitation system will slowly degrade, seriously threatening train operation safety. Currently, the identification of performance degradation in levitation systems faces two major challenges: First, the difficulty of multi-data fusion analysis. The data obtained from train detection includes time-series data such as gaps, acceleration, and current; single-physical-quantity threshold detection methods are prone to false alarms. Second, the weak manifestation of abnormal features under dynamic operating conditions. Maglev systems are significantly affected by noise during operation, and traditional Fourier transforms are insufficient to extract fault features from strong background noise.

[0003] However, existing research still has two main limitations: First, in terms of multi-data fusion, most methods are still limited to the abstract processing of one-dimensional time series signals and lack the ability to intuitively and interpretably fuse multi-source and heterogeneous data; Second, for weak features, current strategies mostly rely on improving the overall model prediction accuracy, but lack front-end signal processing mechanisms that directly enhance and strengthen weak degenerate features. Summary of the Invention

[0004] Therefore, it is necessary to provide a method, device, computer equipment, and medium for judging the performance degradation of a magnetic levitation system that can accurately judge degradation based on a variety of operating data, in order to address the above-mentioned technical problems.

[0005] A method for judging the performance degradation of a magnetic levitation system, the method comprising:

[0006] Obtain a training dataset, which includes multi-parameter training data collected during the operation of the magnetic levitation system and arranged in chronological order, along with the corresponding ground value degradation levels.

[0007] The symmetric polar coordinate method is used to transform the multi-parameter training data of one-dimensional time series data into two-dimensional image sample data.

[0008] The performance degradation discrimination model of the magnetic levitation system is trained using the two-dimensional image sample data. Specifically, the backbone network in the pre-trained Swin Transformer architecture is used to extract multi-scale feature maps from the two-dimensional image sample data. The multi-scale feature maps are then mapped to multi-scale depth feature maps. Based on the discriminative concept embedding obtained by concept mining using a dual encoder structure based on the deep-scale depth feature maps, a concept weight matrix is ​​generated. The degradation level is then predicted based on the fused features obtained by fusing the multi-scale depth feature maps using the concept weight matrix.

[0009] The fine-grained classification loss is calculated based on the prediction results and the corresponding truth degradation level. The dual-objective constraint loss is obtained by applying dual-objective constraints to the results obtained from concept mining.

[0010] The performance degradation discrimination model of the magnetic levitation system is trained based on the fine-grained classification loss and the dual-objective constraint loss to obtain the trained performance degradation discrimination model of the magnetic levitation system.

[0011] The system acquires real-time operational data from the magnetic levitation system and inputs the corresponding two-dimensional image data into the trained magnetic levitation system performance degradation discrimination model for degradation discrimination.

[0012] In one embodiment, in the performance degradation discrimination model of the magnetic levitation system, the multi-scale feature map is mapped to a multi-scale depth feature map using a feature mapping unit.

[0013] The feature mapping unit consists of two linear layers and a ReLU activation function.

[0014] In one embodiment, when using a dual encoder structure to perform concept mining based on deep-scale depth features in the multi-scale depth feature map:

[0015] The deep-scale features are mapped to concept embeddings by the first encoder, and the concept embeddings are decoupled into discriminative concept embeddings and inherited concept embeddings.

[0016] The inherited concept embedding is optimized with the assistance of a second encoder to obtain the final discriminative concept embedding.

[0017] In one embodiment, generating a concept weight matrix based on the discriminative concept embedding includes:

[0018] The discriminative concept embedding is then subjected to global average pooling, and then activated by the Sigmoid activation function to obtain the concept weight matrix.

[0019] In one embodiment, the multi-scale deep feature maps are fused using the concept weight matrix to obtain fused features including:

[0020] Based on the maximum response matrix and average response corresponding to the depth feature maps at each scale, and the concept weight matrix, a visual concept matrix for each scale depth feature map is generated using the Sigmoid activation function.

[0021] Guided by the corresponding visual concept matrix, the depth feature maps at each scale are used to obtain visual concept feature maps at each scale.

[0022] Visual concept feature maps at various scales are stitched together after average pooling, and the stitched feature maps are then processed by an activation function to obtain the fused features.

[0023] In one embodiment, the dual-objective constraint loss includes a first objective constraint loss, expressed as:

[0024] ;

[0025] In the above formula, This indicates temperature hyperparameters. This represents the inherited concept embedding obtained from the first encoder for the i-th two-dimensional image sample data. This represents the inherited concept embedding obtained from the second encoder for the i-th two-dimensional image sample data. and These are positive sample pairs. This represents the inherited concept embedding obtained from the second encoder for the j-th two-dimensional image sample data. This represents the number of concept embeddings in a training batch.

[0026] In one embodiment, the dual-objective constraint loss includes a second objective constraint loss, denoted as;

[0027] ;

[0028] In the above formula, , They represent , of Norm, , These represent the distinctive concept embedding and the inherited concept embedding output by the first encoder, respectively.

[0029] This application also provides a device for determining the performance degradation of a suspension system, the device comprising:

[0030] The training dataset acquisition module is used to acquire the training dataset, which includes multi-parameter training data collected by the magnetic levitation system during operation and arranged in chronological order, as well as the corresponding ground truth degradation levels.

[0031] The training data conversion module is used to convert multi-parameter training data of one-dimensional time series data into two-dimensional image sample data using the symmetric polar coordinate method.

[0032] The prediction result acquisition module is used to train the performance degradation discrimination model of the magnetic levitation system using the two-dimensional image sample data. Specifically, it uses the backbone network in the pre-trained Swin Transformer architecture to extract multi-scale feature maps from the two-dimensional image sample data, maps the multi-scale feature maps to multi-scale depth feature maps, generates a concept weight matrix based on the discriminative concept embedding obtained by concept mining using a dual encoder structure based on the deep-scale depth feature maps, and performs degradation level prediction to obtain the prediction result based on the fused features obtained by fusing the multi-scale depth feature maps using the concept weight matrix.

[0033] The loss function calculation module is used to calculate the fine-grained classification loss based on the prediction results and the corresponding truth degradation level. The dual-objective constraint loss is obtained by applying dual-objective constraints to the results obtained from concept mining.

[0034] The model training module is used to train the performance degradation discrimination model of the magnetic levitation system based on the fine-grained classification loss and the bi-objective constraint loss, so as to obtain the trained performance degradation discrimination model of the magnetic levitation system.

[0035] The magnetic levitation system performance degradation discrimination module is used to acquire real-time operational data of the magnetic levitation system and input the two-dimensional image data corresponding to the operational data into the trained magnetic levitation system performance degradation discrimination model for degradation discrimination.

[0036] A computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the steps in the above-described method for determining the performance degradation of a suspension system.

[0037] A computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps in the above-described method for determining the performance degradation of a levitation system.

[0038] The aforementioned method, apparatus, computer equipment, and medium for judging the performance degradation of magnetic levitation systems utilize a symmetric polar coordinate method to transform multi-parameter training data from one-dimensional time-series data into two-dimensional image sample data. This two-dimensional image sample data is then used to train a performance degradation judgment model for the magnetic levitation system. Within this model, a pre-trained Swin... The backbone network in the Transformer architecture extracts multi-scale feature maps from two-dimensional image sample data, maps these multi-scale feature maps to multi-scale depth feature maps, and uses a dual-encoder structure to perform concept mining based on the deep-scale feature maps in the multi-scale depth feature maps to obtain discriminative concept embeddings. A concept weight matrix is ​​then generated based on these discriminative concept embeddings. The multi-scale depth feature maps are fused using the concept weight matrix to obtain fused features. Degradation levels are predicted based on these fused features, and fine-grained classification loss is calculated based on the prediction results and the corresponding ground truth degradation levels. A dual-objective constraint loss is obtained by applying dual-objective constraints to the concept mining results. The maglev system performance degradation discrimination model is trained based on the fine-grained classification loss and the dual-objective constraint loss, resulting in a trained maglev system performance degradation discrimination model. The running data is then converted into corresponding two-dimensional image data using the symmetric polar coordinate method. The trained maglev system performance degradation discrimination model is then used to perform degradation discrimination based on the two-dimensional image data, yielding degradation discrimination results. By first using the Symmetric Polar Coordinates (SDP) method to transform one-dimensional time-series data into two-dimensional image sample data, interpretable fusion of key parameters from multiple sources is achieved, avoiding misjudgments caused by noise in single data sources, reducing the false detection rate and improving the data feature representation capability. Multi-scale feature maps are extracted using a pre-trained Swin Transformer, and fine-grained weak features for discriminative concept embedding are mined through a dual encoder structure. Multi-scale features are then fused with a concept weight matrix. At the same time, the model is optimized through dual-objective constraint loss to solve the problems of limited receptive field and insensitivity to weak features in traditional models, significantly improving the accuracy of degradation level recognition. Attached Figure Description

[0039] Figure 1 This is a flowchart illustrating a method for determining the performance degradation of a magnetic levitation system in one embodiment.

[0040] Figure 2 This is a schematic diagram illustrating the levitation system structure and levitation data acquisition system structure of a maglev train in one embodiment.

[0041] Figure 3 This is a schematic diagram of Gap1 data conversion to SDP in one embodiment;

[0042] Figure 4 This is a schematic diagram of the conversion of three sets of data—Gap1, Acc1, and current—to SDP in one embodiment;

[0043] Figure 5 This is a schematic diagram of the SDP generated from six sets of data: Gap1, Gap2, Gap3, Acc1, Acc2, and current, in one embodiment.

[0044] Figure 6 This is a schematic diagram of the basic model structure of the Swin Transformer in one embodiment;

[0045] Figure 7 This is a schematic diagram of a concept-guided learning framework in one embodiment;

[0046] Figure 8 This is a schematic diagram of the thermal generation of CGLM in one embodiment;

[0047] Figure 9 This is a radar diagram illustrating the evaluation metrics for the CNN model and the CGLM model in one embodiment, where... Figure 9 (a) shows a diagram illustrating the accuracy. Figure 9 (b) shows a diagram illustrating recall rate. Figure 9 (c) shows a diagram illustrating the false detection rate. Figure 9 (d) shows a diagram illustrating the F1-Score values;

[0048] Figure 10 This is a structural block diagram of a magnetic levitation system performance degradation discrimination device in one embodiment;

[0049] Figure 11 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0050] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0051] In existing technologies, most methods for multi-data fusion are still limited to the abstract processing of one-dimensional time-series signals, lacking the ability to intuitively and interpretably fuse multi-source, heterogeneous data. Secondly, regarding weak features, current strategies mostly rely on improving the overall model's prediction accuracy, lacking front-end signal processing mechanisms that directly enhance and strengthen weak and degenerate features. In this application, as... Figure 1 As shown, a method for judging the performance degradation of a magnetic levitation system is provided, which specifically includes the following steps:

[0052] Step S100: Obtain the training dataset. The training dataset includes multi-parameter training data collected during the operation of the magnetic levitation system, arranged in chronological order, and the corresponding ground truth degradation level. The multi-parameter training data includes 3 sets of suspension gap data, 2 sets of vertical acceleration data, and 1 set of electromagnet current data.

[0053] Step S100: Obtain the training dataset, which includes multi-parameter training data collected in chronological order during the operation of the magnetic levitation system and the corresponding ground truth degradation levels.

[0054] Step S110: Using the symmetric polar coordinate method, the multi-parameter training data of the one-dimensional time series data is transformed into two-dimensional image sample data.

[0055] Step S120: The performance degradation discrimination model of the magnetic levitation system is trained using two-dimensional image sample data. In this step, the backbone network in the pre-trained Swin Transformer architecture is used to extract multi-scale feature maps from the two-dimensional image sample data. The multi-scale feature maps are mapped to multi-scale depth feature maps. Based on the discriminative concept embedding obtained by concept mining using a dual encoder structure based on the deep-scale depth feature maps, a concept weight matrix is ​​generated. Based on the fused features obtained by fusing the multi-scale depth feature maps using the concept weight matrix, the degradation level is predicted to obtain the prediction result.

[0056] Step S130: Calculate the fine-grained classification loss based on the prediction results and the corresponding ground truth degradation level. Obtain the dual-objective constraint loss by applying dual-objective constraints to the results obtained from concept mining.

[0057] Step S140: Train the performance degradation discrimination model of the magnetic levitation system based on fine-grained classification loss and dual-objective constraint loss to obtain the trained performance degradation discrimination model of the magnetic levitation system.

[0058] Step S150: Obtain the real-time operational data of the magnetic levitation system, and input the two-dimensional image data corresponding to the operational data into the trained magnetic levitation system performance degradation discrimination model for degradation discrimination.

[0059] In this application, a concept-guided learning model based on the Swing Transformer is proposed, utilizing the symmetric point pattern for multi-data fusion. The symmetric point pattern addresses the challenge of multi-data fusion by fusing suspension gap, acceleration, and current data to generate an SDP image. Deep learning is introduced to address the problem of weak feature representation; the concept-guided learning framework and Swing Transformer are fused to obtain the CGLM model, i.e., the performance degradation discrimination model for magnetic levitation systems. This improves the weights of discriminative concepts and reduces the weights of inherited concepts, thereby enhancing the model's ability to recognize subtle features and distinguish images.

[0060] The levitation system mainly consists of levitation electromagnets, controllers, and sensors. It achieves stable, contactless operation between the train and the track through electromagnetic force. However, after long-term operation, key performance indicators of the levitation system will gradually deviate from their design optimum values. To address the performance degradation issue, data-driven methods can be used for condition monitoring and fault diagnosis. During operation, a levitation data acquisition system continuously monitors and records key signals such as levitation gap, vertical acceleration, and current.

[0061] Specifically, maglev trains generally consist of multiple carriages; this article uses a three-carriage train as an example. Each carriage is designed with five suspension frames, each frame containing four suspension units, and each suspension unit is a complete suspension system. Therefore, the data acquisition system for each carriage includes 20 suspension sensors and corresponding data acquisition units, enabling simultaneous data collection from 20 suspension points within one carriage. The suspension data acquisition system mainly consists of modules such as a data acquisition unit, suspension gap sensors, current sensors, and vertical acceleration sensors. The suspension sensors utilize eddy current technology to monitor and provide feedback on the train's suspension status in real time. The structure and operating principle of the maglev train's suspension system and suspension data acquisition system are as follows... Figure 2 As shown.

[0062] Furthermore, suspension sensors are installed at suspension sampling points 4, 5, 6, and 12 on car MC1 (i.e., the lead car), with a data sampling rate of 1 kHz. The acquired signals include the output gaps (i.e., suspension gaps) of four suspension sensors: Gap1, Gap2, Gap3, and Gap4; and the vertical acceleration outputs of two suspension sensors, Acc1 and Acc2, and one electromagnet current. The vertical acceleration sensor has a measurement range of ±5g, a frequency response range of 0~1kHz (-3dB), and a nonlinear error not exceeding ±5%FSO; the current sensor has a measurement range of 0~100A, a nonlinear error not exceeding ±1%FSO, and can measure DC current. After the train starts and runs stably, the rated values ​​of the key parameters are shown in Table 1.

[0063] Table 1

[0064]

[0065] Significant fluctuations in these key parameters indicate system performance degradation. The degradation can be categorized into three levels, as shown in Table 2:

[0066] Table 2 Description of Degradation Levels

[0067]

[0068] Furthermore, in order to better identify and differentiate the current level of degradation of the suspension system, it is first necessary to collect and analyze the data of the suspension system.

[0069] In this embodiment, the SDP analysis method, also known as the symmetric polar coordinate method, is used to fuse the data. This method can transform the acquired one-dimensional time-series data into a two-dimensional image. For multiple sets of discrete data, different data fusions can be performed when generating the SDP image. The SDP image can reflect changes in signal amplitude and frequency, such as gaps, making it more conducive to anomaly feature extraction. Therefore, to more intuitively identify the performance degradation of the suspension system, the acquired suspension data is converted into an SDP image using the SDP method.

[0070] Next, SDP was introduced for experimental data fusion processing. To ensure data consistency, only the first 38 sampling points were selected for processing for all degradation levels, with a time interval of l=2 s and a sampling frequency of 12000.

[0071] Since gap fluctuations are the most direct indicator of levitation system performance, SDP transformation was first performed on only the Gap1 dataset, repeated six times. The six dimensions obtained for each degree of degradation were identical, as shown in the following results. Figure 3 As shown, the gap fluctuation under normal operating conditions is minimal, resulting in a relatively dispersed distribution of scatter points in the generated SDP (Special Purpose Map) image. The scatter points from both Level 1 and Level 2 degradation are distributed from the center to the outermost edge. Since all six dimensions involve gap value fluctuations, the SDP images generated by Level 1 and Level 2 degradation, while different, cannot be precisely determined. The scatter points in the image obtained from Level 3 degradation break midway from the center to the outermost edge. Therefore, the gap fluctuation characteristics are similar across different degradation levels. Thus, determining the degradation level solely based on gap data from a single set of magnetic levitation systems is unreliable.

[0072] Furthermore, the three sets of data (Gap1, Acc1, and current) were sampled twice, and three sets of symmetrical dimensions were obtained for each degradation level, such as... Figure 4As shown. The differences in gap under different degradation levels when using three sets of data are the same as when using one set of data, and will not be repeated. Under normal operating conditions, the scatter points obtained from Acc1 and current conversion are evenly distributed from the center to the outermost circle. The scatter points obtained under Level 1 degradation are distributed from the center to the outermost circle, and are densely distributed at a distance from the center. Under Level 2 degradation, the scatter points obtained from Acc1 are distributed from the center to the outermost circle, and the scatter points obtained from current conversion are concentrated near the center with scattered points on the periphery. Under Level 3 degradation, the scatter points obtained from Acc1 are distributed from the center to the periphery, and the scatter points obtained from current conversion show breaks. In summary, the SDP image obtained using three sets of data is more obvious in terms of the differences in different degradation levels. Therefore, comprehensive analysis of gap, vertical acceleration, and current data can better reflect the degradation of the system. The data fusion results of three different types of data are more intuitive and more conducive to identifying the degradation level of the system than a single type of data.

[0073] Finally, to further reduce the similarity of SDP images at different degradation levels, all six sets of data were sampled, including Gap1, Gap2, Gap3, Acc1, Acc2, and current data. Gap1, Gap2, and Gap3 are the suspension gap data from three different suspension sampling points, and Acc1 and Acc2 are the vertical acceleration data from two different sampling points. The following section mainly focuses on the comprehensive analysis of the SDP images generated by fusing these six sets of data. Figure 5 As shown. Under normal operating conditions, the most obvious feature of the image is the scattered points distributed in broken concentric circles, obtained from the conversion of three sets of gap data. Under first-level degradation, the most obvious feature is the scattered points that cluster at the center and spread outwards, obtained from the conversion of three sets of gap data. Under second-level degradation, the most obvious feature is the scattered points concentrated near the center, obtained from the conversion of one set of current data. Under third-level degradation, the most obvious feature is the scattered points concentrated near the center, obtained from the conversion of three sets of gap data and one set of current data. In summary, the SDP images obtained by fusing six sets of data are not entirely the same under different degradation levels. Compared with the fusion of one or three sets of data, the fusion of six sets of data is more conducive to distinguishing the system performance degradation. Therefore, the SDP image obtained by fusing six sets of data was ultimately chosen for subsequent model training.

[0074] In summary, in step S100, the multi-parameter training data includes three sets of suspension gap data, two sets of vertical acceleration data, and one set of electromagnet current data, namely, six sets of data: Gap1, Gap2, Gap3, Acc1, Acc2, and current data, along with corresponding ground truth degradation levels. The degradation levels are shown in Table 2, including Level 1, Level 2, and Level 3 degradation levels. When constructing the training dataset, it also includes corresponding non-degraded data, allowing the subsequent model training to determine whether degradation exists. If degradation exists, the specific degradation level is predicted. In step S110, the data from the various parameters received in step S100 are fused using SDP analysis to obtain the corresponding SDP graph, which serves as the input data for subsequent model training.

[0075] In this embodiment, identifying performance degradation in the magnetic levitation system requires enhancing the ability to identify and extract subtle features, further improving the accuracy of anomaly detection and state recognition. The Swin Transformer and concept-guided learning address the bottlenecks of traditional CNN models from different perspectives. Therefore, this method proposes a concept-guided learning model (CGLM) based on the Swin Transformer, which can capture complex and subtle features while possessing the interpretability and controllability of concept-guided learning.

[0076] First, Swin-T (Swin Transformer) inherits the powerful global modeling capabilities of the Transformer architecture. Compared to CNN models with limited receptive fields, it introduces hierarchical feature extraction and a windowed multi-head self-attention (W-MSA) mechanism, enabling it to extract globally correlated features from a macroscopic perspective of the image. Swin-T possesses strong feature extraction capabilities, making it suitable as a backbone network. The basic model structure of Swin Transformer is as follows: Figure 6 As shown.

[0077] The process of determining whether a levitation system has degraded and identifying the level of degradation is a typical progression from coarse-grained to fine-grained. Coarse-grained and fine-grained are widely used concepts in image recognition. They describe the granularity level of a system, data, or task, i.e., the level of detail or abstraction. Coarse-grained emphasizes the macroscopic and holistic perspective, focusing on rough categories, outlines, and commonalities. Fine-grained emphasizes the microscopic and local perspective, focusing on fine distinctions, details, and characteristics. A coarse-grained perspective can be used to determine whether a levitation system has degraded. Furthermore, a fine-grained perspective can be used to identify the level of degradation in the system.

[0078] Furthermore, the concepts encompassed in fine-grained categories can be divided into two categories: inherited concepts and distinguishing concepts. Inherited concepts are fundamental features shared by all fine-grained categories within the coarse-grained category, providing a basic semantic framework for the fine-grained categories and ensuring they belong to the same category at the macro level. Distinguishing concepts refer to unique features that are not found in other fine-grained categories within the coarse-grained category, distinguishing them from each other.

[0079] Specifically, concepts are expressed using embedding spaces, using... To represent the concept of a fine-grained category. It is composed of Embedding vectors that are unrelated to this category ,in, The width corresponding to the image feature, The height of the corresponding image feature. Therefore, we can conclude that... It is a distinguishing concept embedding Embedding the concept of inheritance The combination of can be expressed as:

[0080] (1)

[0081] In formula (1), This represents a combination function.

[0082] Furthermore, the embedding of distinguishing concepts This is used to distinguish the fine-grained category from other fine-grained categories within the same coarse-grained category. Inheritance concept embedding... This represents a concept shared by all fine-grained categories, which belong to their corresponding coarse-grained categories.

[0083] In this embodiment, the Swin Transformer is selected as the backbone network, and the concept-guided learning framework proposed in this method includes three main steps: concept mining, concept fusion, and concept constraint. Structurally, the Swin Transformer has four feature blocks, and the outputs of the first to fourth blocks are denoted as... , , and Using four encoders Each Normalized to a fixed number of channels Embedded, denoted as .here , and These refer to the number of channels, width, and height, respectively.

[0084] In this embodiment, the concept mining step aims to extract deep features Extracting Distinctive Concept Embeddings This is the inverse operation of formula (1). Design a two-stage encoder. The first stage of learning begins with image features. To concept embedding The mapping. The second stage embeds the concepts. Decoupling into Distinctive Concept Embedding Embedding the concept of inheritance Both stages are represented as:

[0085] (2)

[0086] However, formula (2) does not effectively distinguish between inherited concept embeddings and discriminative concept embeddings. Therefore, another encoder needs to be designed. The encoder directly learns from... The given concept embedding is represented as:

[0087] (3)

[0088] By encoder The resulting inheritance concept embedding It can help distinguish encoders The distinctive concept embedding.

[0089] Specifically, through unsupervised learning and contrastive methods, the encoder will be... The resulting concept of inheritance is embedded and from the encoder The resulting concept of inheritance is embedded Hide it. Due to the size of the concept embedding. Since it is low-dimensional, channel normalization is performed on the embeddings of the two concepts to make subsequent feature propagation more stable. For example, suppose yes Previous embedded output. Each individual concept in The calculation is expressed as a formula:

[0090] (4)

[0091] (5)

[0092] In formulas (4) and (5), and These represent the learnable scaling and shift parameters, respectively, which follow the default settings of the normalization function. It is the first before normalization The original embedding vector of each concept is the output of the previous network layer. It is the embedding vector of all concepts The mean. It is the embedding vector of all concepts The standard deviation. It is a very small constant (e.g., 10). -5 (), used to avoid the denominator being zero and to ensure numerical stability.

[0093] Therefore, in step S120, in the magnetic levitation system performance degradation discrimination model, the feature mapping unit is used to map the multi-scale feature map to the multi-scale depth feature map. The feature mapping unit consists of two linear layers and a ReLU activation function.

[0094] Specifically, the feature mapping unit is represented as a function Its goal is to combine different scales Image features Mapping to a fixed number of deep features This facilitates unified processing of multi-scale features (Swin-T features have a channel size of 1536). Function It consists of two linear layers, each followed by a ReLU activation function.

[0095] In this embodiment, when using a dual encoder structure to perform concept mining based on deep-scale features in a multi-scale depth feature map: the deep-scale features are mapped to concept embeddings by the first encoder, the concept embeddings are decoupled into discriminative concept embeddings and inherited concept embeddings, and then the inherited concept embeddings are optimized by the second encoder to obtain the final discriminative concept embeddings.

[0096] Specifically, for encoders Its structural design consists of two main stages. In the first stage, the input data is processed through two consecutive linear layers. The first linear layer maps the input data to a feature space with 64 channels, initially extracting low-level features from the input data. The second linear layer adjusts the dimension of the feature space to... ,in This is a predefined parameter representing the number of channels in the layer's output. The goal of this stage is to progressively abstract the input data, preparing it for subsequent feature extraction and encoding.

[0097] Furthermore, the second stage comprises two branches that process the feature data from the first stage in parallel. Each branch contains a linear layer with an output channel size of [missing information]. This branching structure allows the model to process features from different perspectives, thereby extracting richer and more comprehensive feature information. Therefore, The encoder can effectively integrate features from two branches, further improving the expressive power of features and the performance of the model.

[0098] Specifically, for encoders The channel sizes for the first and second layers are set to 64 and 64 respectively. Specifically, the first layer maps the input features to a 64-channel feature space, initially extracting low-level features; the second layer adjusts the dimension of the feature space to... To obtain a more compact and advanced feature representation. Fine-grained concept embedding in encoders Distinctive concept embedding Embedding the concept of inheritance All are designed to have N There are n class-independent embedding vectors. The dimensions of these embedding vectors are 1. ,in and These represent the width and height of the embedding vector, respectively. This design allows each embedding vector to capture more general feature representations without relying on specific category information, thereby enhancing the model's generalization ability.

[0099] Preferably, The value is set to 32. This choice is based on a comprehensive consideration of model performance and computational efficiency. This ensures that the model remains sensitive to fine-grained features while avoiding excessive computational complexity and memory consumption.

[0100] In this embodiment, concept fusion enables the learning of more representative, fine-grained representations. It primarily utilizes discriminative concept embeddings. and image features This is implemented on image features at each scale.

[0101] Specifically, generating a concept weight matrix based on the discriminative concept embedding includes: performing global average pooling on the discriminative concept embedding, and then applying the Sigmoid activation function to obtain the concept weight matrix, i.e. After global average pooling, a concept weight matrix is ​​generated using the Sigmoid activation function. .

[0102] Furthermore, the multi-scale deep feature maps are fused using the concept weight matrix to obtain fused features. This includes generating visual concept matrices for each scale of deep feature maps using the Sigmoid activation function based on the maximum and average response matrices corresponding to each scale of deep feature maps, as well as the concept weight matrix. Guided by the corresponding visual concept matrices, visual concept feature maps for each scale are obtained. The visual concept feature maps for each scale are then concatenated after average pooling. Finally, the concatenated feature maps are processed by an activation function to obtain the fused features.

[0103] Specifically, from the scale Image features Considering the maximum response and average response Specifically Activate the most significant and highest response. It is the concept weight matrix and The balance measures the overall response of fine-grained categories. A scaled-down visual concept matrix. , is represented as:

[0104] (6)

[0105] In formula (6), Sigmoid This represents the Sigmoid activation function. and It is a scale The weights and bias matrices of the linear layer, This indicates a join operation.

[0106] For scale Concept-guided representation The method is expressed as:

[0107] (7)

[0108] In formula (7), This represents the Hadamard product.

[0109] In this embodiment, the concept-guided learning model based on the Swing Transformer is implemented in an end-to-end manner, that is, the structure of the magnetic levitation system performance degradation discrimination model is as follows: Figure 7 As shown.

[0110] In step S140, in order to emphasize the unity of the inheritance concept and the uniqueness of the discriminative concept in the embedding space, a dual-objective constraint mechanism is proposed, including reducing the inheritance embedding of different instances of the same category. and The distance ensures the stability of coarse-grained shared features while increasing inherited embedding. and distinctive embedding The distance between them facilitates the model's learning of fine-grained, unique discriminative features.

[0111] Specifically, for the first objective, a contrastive learning paradigm is adopted. Assume... In and In These are positive sample pairs, constrained loss. The InfoNCE loss is calculated using a channel-by-channel implementation and is expressed as follows:

[0112] (8)

[0113] In formula (8), This represents the inherited concept embedding obtained from the first encoder for the i-th two-dimensional image sample data. This represents the inherited concept embedding obtained from the second encoder for the i-th two-dimensional image sample data. and These are positive sample pairs. This represents the inherited concept embedding obtained from the second encoder for the j-th two-dimensional image sample data. This represents the number of concept embeddings in a training batch. This represents the temperature hyperparameter, used to control the scaling ratio of the similarity metric; preferably, it is set to 0.1. Through the design of the loss function, the model can maximize the similarity between positive sample pairs during training, while minimizing the similarity between positive samples and other samples.

[0114] Furthermore, regarding the second objective, by minimizing and cosine similarity This increases the distance between them. Cosine similarity is a commonly used similarity measure that measures the similarity between two vectors by calculating the angle between them. The calculation process is as follows:

[0115] (9)

[0116] In formula (9), express of Norm, express of Norm. By minimizing cosine similarity, the distance between two samples is increased, thereby enabling the differentiation of different features. , These represent the distinctive concept embedding and the inherited concept embedding output by the first encoder, respectively.

[0117] In this embodiment, the final fine-grained semantic prediction is the prediction result. The representation is guided by the concept of all scales. The generation process is represented as follows:

[0118] (10)

[0119] Furthermore, fine-grained classification loss function Based on the cross-entropy loss formula, it can be expressed as:

[0120] (11)

[0121] In formula (11), It is the first The true value of each fine-grained category It is the first The predicted probability value for each fine-grained category.

[0122] In this embodiment, the total loss function yes , , The combination of is represented as:

[0123] (12)

[0124] In formula (12), As a balancing quantity, according to Perform calibration.

[0125] In this embodiment, the backbone network uses a model pre-trained on ImageNet as its initial parameter values. The remaining weights and biases of the framework are randomly initialized. During training, a stochastic gradient descent (SGD) optimizer is used with a maximum learning rate of 0.001, and the network is trained for a total of 10 epochs.

[0126] In step S150, six sets of data are acquired at the current time: Gap1, Gap2, Gap3, Acc1, Acc2 and current data. These six sets of data are converted into SDP images. The trained model is used to output the current performance degradation judgment result of the magnetic levitation system based on the SDP images.

[0127] In this method, to verify its effectiveness, the concept-guided learning CGLM model proposed in this method was used to train six sets of multi-data fusion SDP images. The training epoch was 10 and the learning rate was 0.001.

[0128] First, randomly select an image to generate a heatmap, such as... Figure 8 As shown in the image. Red indicates highly activated regions on the heatmap. From... Figure 8 As can be seen, by introducing the concept-guided learning framework, CGLM can effectively separate discriminative and inherited concepts, and focuses on the extracted discriminative concepts. The weight of discriminative concepts increases, so the high-activation regions in the heatmap generated by CGLM essentially cover all effective features in circles. The weight of inherited concepts decreases, resulting in less activation of irrelevant background.

[0129] In summary, CGLM is highly sensitive to discriminative concepts within the circle, exhibiting only a small portion of false activations outside the circle. Compared to CNN models, it suppresses irrelevant background responses outside the circle while increasing the responses of effective features within the circle. Therefore, the model demonstrates strong fine-grained recognition capabilities, validating its effectiveness. It can be proven that CGLM outperforms CNN models in recognizing degradation in suspended systems.

[0130] Further quantitative analysis of evaluation metrics was conducted to verify the effectiveness of CGLM. To comprehensively evaluate the recognition performance of the CGLM model compared to three typical CNN models, four evaluation metrics—mean precision, recall, false positive rate (FPR), and F1-Score—were selected for quantitative analysis. In each of the 10 iterations, corresponding evaluation metric values ​​were generated for each model. A radar chart was plotted for each evaluation metric as shown below. Figure 9 As shown.

[0131] like Figure 9 As shown in (a), the precision is calculated as the average precision across different degradation levels. Since the CNN model fails to distinguish between inherited and discriminative concepts, its recognition performance is poor, resulting in a lower precision. In contrast, CGLM can distinguish between inherited and discriminative concepts and can identify subtle features, thus achieving a higher precision. The precision obtained by CGLM is 31.1%, 19.5%, and 19.9% ​​higher than that of the AlexNet, GoogleNet, and ResNet50 models, respectively.

[0132] like Figure 9 As shown in (b), the recall rate is calculated as the average recall rate across different degradation levels. Because the CNN model makes more errors in classifying different degradation levels, its recall rate is lower. CGLM adds a fine-grained module, which has strong fine-grained recognition capabilities, thus resulting in a higher recall rate. Furthermore, the recall rate obtained by CGLM is 32.3%, 21.2%, and 24.2% higher than that of the AlexNet, GoogleNet, and ResNet50 models, respectively.

[0133] like Figure 9As shown in (c), the false positive rate is calculated as the average of the false positive rates for different degradation levels. CGLM incorporates a concept-guided learning framework to improve its feature selection ability, resulting in an extremely low false negative rate. Furthermore, CGLM's false negative rate is reduced by 5.7%, 3.3%, and 4.5% compared to the AlexNet, GoogleNet, and ResNet50 models, respectively.

[0134] like Figure 9 As shown in (d), the F1-Score is calculated as the average of different degradation levels. The F1-Score is the harmonic mean of precision and recall, comprehensively reflecting the model's balanced performance. Because CNN models have limited discriminative power and limited ability to extract and amplify discriminative concepts, both precision and recall are low. Therefore, their F1-Scores are also low. CGLM can extract discriminative concepts and increase their weights, resulting in better classification ability and a higher F1-Score. Furthermore, CGLM's F1-Score is 32.5%, 22.7%, and 25.3% higher than that of AlexNet, GoogleNet, and ResNet50 models, respectively.

[0135] In summary, the CGLM model achieves better evaluation metrics than the CNN model and performs better in identifying degradation in magnetic levitation systems. CGLM introduces a concept-guided learning framework that includes concept mining, concept fusion, and concept constraints, enabling it to distinguish between inherited and discriminative concepts. By increasing the weights of discriminative concepts, making them more prominent during prediction, the model's ability to recognize similar images is improved. CGLM effectively identifies the performance degradation levels of magnetic levitation systems, thus significantly improving recognition accuracy.

[0136] In the aforementioned method for identifying the performance degradation of magnetic levitation systems, firstly, a symmetric point model is used to process multiple sets of operational data from the magnetic levitation system. Data on gap, vertical acceleration, and current are collected from the operation of the magnetic levitation system under different degradation levels. SDP images obtained by fusing one, three, and six sets of data are then generated using the symmetric point model. A basic CNN model is selected to train different sets of SDP images. The results show that fusing six sets of data is most effective in improving recognition performance. Secondly, a concept-guided learning framework (CGLM) model is proposed by fusing it with a Swin Transformer network structure. To enhance feature representation, the weights of discriminative concepts are increased while the weights of inherited concepts are decreased. Finally, the various indicators of CGLM are compared with those of the CNN model, successfully verifying that CGLM is more effective than the CNN model in identifying the performance degradation of magnetic levitation systems, with an average accuracy improvement of 8.9%.

[0137] It should be understood that, although Figure 1The steps in the flowchart are shown sequentially as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified herein, there is no strict order in which these steps are executed, and they can be performed in other orders. Figure 1 At least some of the steps in the process may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be executed in turn or alternately with other steps or at least some of the sub-steps or stages of other steps.

[0138] In one embodiment, such as Figure 10 As shown, a magnetic levitation system performance degradation discrimination device is provided, comprising: a training dataset acquisition module 200, a training data conversion module 210, a prediction result acquisition module 220, a loss function calculation module 230, a model training module 240, and a magnetic levitation system performance degradation discrimination module 250, wherein:

[0139] The training dataset acquisition module 200 is used to acquire a training dataset, which includes multi-parameter training data collected by the magnetic levitation system during operation and arranged in chronological order, as well as the corresponding ground value degradation level.

[0140] The training data conversion module 210 is used to convert multi-parameter training data of one-dimensional time series data into two-dimensional image sample data using the symmetric polar coordinate method.

[0141] The prediction result acquisition module 220 is used to train the performance degradation discrimination model of the magnetic levitation system using the two-dimensional image sample data. Specifically, it uses the backbone network in the pre-trained Swin Transformer architecture to extract multi-scale feature maps of the two-dimensional image sample data, maps the multi-scale feature maps to multi-scale depth feature maps, generates a concept weight matrix based on the discriminative concept embedding obtained by concept mining using a dual encoder structure based on the deep-scale depth feature maps, and performs degradation level prediction based on the fused features obtained by fusing the multi-scale depth feature maps using the concept weight matrix to obtain the prediction result.

[0142] The loss function calculation module 230 is used to calculate the fine-grained classification loss based on the prediction results and the corresponding truth degradation level, and to obtain the dual-objective constraint loss by applying dual-objective constraints to the results obtained from concept mining.

[0143] The model training module 240 is used to train the performance degradation discrimination model of the magnetic levitation system based on the fine-grained classification loss and the dual-objective constraint loss, so as to obtain the trained performance degradation discrimination model of the magnetic levitation system.

[0144] The magnetic levitation system performance degradation discrimination module 250 is used to acquire real-time operational data of the magnetic levitation system and input the two-dimensional image data corresponding to the operational data into the trained magnetic levitation system performance degradation discrimination model for degradation discrimination.

[0145] Specific limitations regarding the performance degradation detection device for magnetic levitation systems can be found in the limitations of the performance degradation detection method for magnetic levitation systems described above, and will not be repeated here. Each module in the aforementioned performance degradation detection device for magnetic levitation systems can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in hardware or independently of the processor in a computer device, or stored in software in the memory of a computer device, so that the processor can call and execute the corresponding operations of each module.

[0146] In one embodiment, a computer device is provided, which may be a terminal, and its internal structure diagram may be as follows: Figure 11 As shown, the computer device includes a processor, memory, network interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The network interface is used to communicate with external terminals via a network connection. When executed by the processor, the computer program implements a method for judging the performance degradation of a magnetic levitation system. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad mounted on the computer device casing, or an external keyboard, touchpad, or mouse.

[0147] Those skilled in the art will understand that Figure 11 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0148] In one embodiment, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to perform the following steps:

[0149] Obtain a training dataset, which includes multi-parameter training data collected during the operation of the magnetic levitation system and arranged in chronological order, along with the corresponding ground value degradation levels.

[0150] The symmetric polar coordinate method is used to transform the multi-parameter training data of one-dimensional time series data into two-dimensional image sample data.

[0151] The performance degradation discrimination model of the magnetic levitation system is trained using the two-dimensional image sample data. Specifically, the backbone network in the pre-trained Swin Transformer architecture is used to extract multi-scale feature maps from the two-dimensional image sample data. The multi-scale feature maps are then mapped to multi-scale depth feature maps. Based on the discriminative concept embedding obtained by concept mining using a dual encoder structure based on the deep-scale depth feature maps, a concept weight matrix is ​​generated. The degradation level is then predicted based on the fused features obtained by fusing the multi-scale depth feature maps using the concept weight matrix.

[0152] The fine-grained classification loss is calculated based on the prediction results and the corresponding truth degradation level. The dual-objective constraint loss is obtained by applying dual-objective constraints to the results obtained from concept mining.

[0153] The performance degradation discrimination model of the magnetic levitation system is trained based on the fine-grained classification loss and the dual-objective constraint loss to obtain the trained performance degradation discrimination model of the magnetic levitation system.

[0154] The system acquires real-time operational data from the magnetic levitation system and inputs the corresponding two-dimensional image data into the trained magnetic levitation system performance degradation discrimination model for degradation discrimination.

[0155] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, the computer program performing the following steps when executed by a processor:

[0156] Obtain a training dataset, which includes multi-parameter training data collected during the operation of the magnetic levitation system and arranged in chronological order, along with the corresponding ground value degradation levels.

[0157] The symmetric polar coordinate method is used to transform the multi-parameter training data of one-dimensional time series data into two-dimensional image sample data.

[0158] The performance degradation discrimination model of the magnetic levitation system is trained using the two-dimensional image sample data. Specifically, the backbone network in the pre-trained Swin Transformer architecture is used to extract multi-scale feature maps from the two-dimensional image sample data. The multi-scale feature maps are then mapped to multi-scale depth feature maps. Based on the discriminative concept embedding obtained by concept mining using a dual encoder structure based on the deep-scale depth feature maps, a concept weight matrix is ​​generated. The degradation level is then predicted based on the fused features obtained by fusing the multi-scale depth feature maps using the concept weight matrix.

[0159] The fine-grained classification loss is calculated based on the prediction results and the corresponding truth degradation level. The dual-objective constraint loss is obtained by applying dual-objective constraints to the results obtained from concept mining.

[0160] The performance degradation discrimination model of the magnetic levitation system is trained based on the fine-grained classification loss and the dual-objective constraint loss to obtain the trained performance degradation discrimination model of the magnetic levitation system.

[0161] The system acquires real-time operational data from the magnetic levitation system and inputs the corresponding two-dimensional image data into the trained magnetic levitation system performance degradation discrimination model for degradation discrimination.

[0162] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0163] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0164] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. A method for determining performance degradation of a magnetic levitation system, characterized by, The method comprises: obtaining a training data set, the training data set comprising multi-parameter training data arranged in time sequence and corresponding true value degradation levels collected by a magnetic suspension system during operation; using a symmetric polar coordinate method to convert the one-dimensional time series data of the multi-parameter training data into two-dimensional image sample data; training a magnetic suspension system performance degradation discrimination model using the two-dimensional image sample data, wherein a backbone network in a pre-trained Swin Transformer architecture is used to extract multi-scale feature maps of the two-dimensional image sample data, the multi-scale feature maps are mapped to multi-scale deep feature maps, a distinctive concept embedding is obtained based on concept mining using a double-encoder structure according to deep-scale deep feature maps, a concept weight matrix is generated, and a fusion feature is obtained by fusing the multi-scale deep feature maps using the concept weight matrix, and a prediction result is obtained by predicting the degradation level; calculating a fine-grained classification loss according to the prediction result and the corresponding true value degradation level, and obtaining a double-target constraint loss by double-target constraining the result obtained by concept mining; training the magnetic suspension system performance degradation discrimination model according to the fine-grained classification loss and the double-target constraint loss to obtain a trained magnetic suspension system performance degradation discrimination model; obtaining real-time collected operation data of the magnetic suspension system, and inputting the corresponding two-dimensional image data of the operation data into the trained magnetic suspension system performance degradation discrimination model for degradation discrimination.

2. The magnetic levitation system performance degradation determination method according to claim 1, characterized by, In the magnetic suspension system performance degradation discrimination model, the multi-scale feature maps are mapped to multi-scale deep feature maps using a feature mapping unit. The feature mapping unit is composed of two linear layers and a ReLU activation function.

3. The magnetic levitation system performance degradation determination method according to claim 1, characterized by, When mining concepts from deep-scale deep features in the multi-scale deep feature maps using a double-encoder structure: the deep-scale deep features are mapped to concept embeddings by a first encoder, and the concept embeddings are decoupled into distinctive concept embeddings and inherited concept embeddings; the inherited concept embeddings are optimized by a second encoder to obtain the final distinctive concept embeddings.

4. The magnetic levitation system performance degradation determination method according to claim 1, characterized by, Generating a concept weight matrix based on the distinctive concept embedding comprises: after the distinctive concept embedding is globally averaged and pooled, a Sigmoid activation function is used to obtain the concept weight matrix.

5. The method of claim 1, wherein the magnetic levitation system performance degradation determination method is characterized by, Fusing the multi-scale deep feature maps using the concept weight matrix to obtain a fusion feature comprises: a visual concept matrix of each scale deep feature map is generated by a Sigmoid activation function according to the maximum response matrix and the average response of each scale deep feature map and the concept weight matrix; each scale deep feature map obtains a visual concept feature map of each scale under the guidance of the corresponding visual concept matrix; after the visual concept feature maps of each scale are averaged and pooled, the fused feature is obtained by processing the feature maps after activation function.

6. The method of claim 3, wherein the magnetic levitation system performance degradation determination method is characterized by, The double-target constraint loss comprises a first target constraint loss, which is represented as: In the above formulae, denotes the temperature hyperparameter, denotes the inherited conceptual embedding of the i-th 2D image sample data by the first encoder, denotes the inherited conceptual embedding of the i-th 2D image sample data by the second encoder, and is a positive sample pair, denotes the inherited conceptual embedding of the j-th 2D image sample data by the second encoder, denotes the number of conceptual embeddings in a training batch.

7. The method of claim 4, wherein the magnetic levitation system performance degradation determination method is characterized by, The double-target constraint loss comprises a second target constraint loss, which is represented as: In the above formulae, , respectively represent , the norms, , respectively represent the discriminative concept embeddings and the inherited concept embeddings output by the first encoder.

8. A device for determining degradation of performance of a suspension system, characterized in that The device comprises: The training data set acquisition module is configured to acquire a training data set, wherein the training data set comprises time-sequentially arranged multi-parameter training data collected during operation of the magnetic suspension system and corresponding true value degradation levels; The training data conversion module is configured to convert the one-dimensional time-series data multi-parameter training data into two-dimensional image sample data by using a symmetric polar coordinate method; The prediction result obtaining module is configured to train the magnetic suspension system performance degradation discrimination model by using the two-dimensional image sample data, wherein a backbone network in a pre-trained Swin Transformer architecture is used to extract multi-scale feature maps of the two-dimensional image sample data, the multi-scale feature maps are mapped to multi-scale deep feature maps, a discriminative concept embedding obtained by using a double-encoder structure to perform concept mining on deep-scale deep feature maps is used to generate a concept weight matrix, and a fusion feature obtained by fusing the multi-scale deep feature maps by using the concept weight matrix is used to perform degradation level prediction to obtain a prediction result; The loss function calculation module is configured to calculate a fine-grained classification loss according to the prediction result and the corresponding true value degradation level, and obtain a double-target constraint loss by performing double-target constraint on a result obtained by concept mining; The model training module is configured to train the magnetic suspension system performance degradation discrimination model according to the fine-grained classification loss and the double-target constraint loss, and obtain a trained magnetic suspension system performance degradation discrimination model. The magnetic suspension system performance degradation discrimination module is configured to acquire real-time collected operation data of the magnetic suspension system, input two-dimensional image data corresponding to the operation data into the trained magnetic suspension system performance degradation discrimination model, and perform degradation discrimination. 9.A computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the computer device is configured to perform the method according to any one of claims 1-8 when the computer program is executed by the processor. The processor executes the computer program to implement the steps of the method of any one of claims 1 to 7.

10. A computer-readable storage medium having stored thereon a computer program, characterized in that, The computer program is executed by the processor to implement the steps of the method of any one of claims 1 to 7.

Images